xref: /netbsd-src/sys/kern/kern_descrip.c (revision 1b9578b8c2c1f848eeb16dabbfd7d1f0d9fdefbd)
1 /*	$NetBSD: kern_descrip.c,v 1.215 2011/06/26 16:42:41 christos Exp $	*/
2 
3 /*-
4  * Copyright (c) 2008, 2009 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Andrew Doran.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 /*
33  * Copyright (c) 1982, 1986, 1989, 1991, 1993
34  *	The Regents of the University of California.  All rights reserved.
35  * (c) UNIX System Laboratories, Inc.
36  * All or some portions of this file are derived from material licensed
37  * to the University of California by American Telephone and Telegraph
38  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
39  * the permission of UNIX System Laboratories, Inc.
40  *
41  * Redistribution and use in source and binary forms, with or without
42  * modification, are permitted provided that the following conditions
43  * are met:
44  * 1. Redistributions of source code must retain the above copyright
45  *    notice, this list of conditions and the following disclaimer.
46  * 2. Redistributions in binary form must reproduce the above copyright
47  *    notice, this list of conditions and the following disclaimer in the
48  *    documentation and/or other materials provided with the distribution.
49  * 3. Neither the name of the University nor the names of its contributors
50  *    may be used to endorse or promote products derived from this software
51  *    without specific prior written permission.
52  *
53  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
54  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
55  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
56  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
57  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
58  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
59  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
60  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
61  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
62  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
63  * SUCH DAMAGE.
64  *
65  *	@(#)kern_descrip.c	8.8 (Berkeley) 2/14/95
66  */
67 
68 /*
69  * File descriptor management.
70  */
71 
72 #include <sys/cdefs.h>
73 __KERNEL_RCSID(0, "$NetBSD: kern_descrip.c,v 1.215 2011/06/26 16:42:41 christos Exp $");
74 
75 #include <sys/param.h>
76 #include <sys/systm.h>
77 #include <sys/filedesc.h>
78 #include <sys/kernel.h>
79 #include <sys/proc.h>
80 #include <sys/file.h>
81 #include <sys/socket.h>
82 #include <sys/socketvar.h>
83 #include <sys/stat.h>
84 #include <sys/ioctl.h>
85 #include <sys/fcntl.h>
86 #include <sys/pool.h>
87 #include <sys/unistd.h>
88 #include <sys/resourcevar.h>
89 #include <sys/conf.h>
90 #include <sys/event.h>
91 #include <sys/kauth.h>
92 #include <sys/atomic.h>
93 #include <sys/syscallargs.h>
94 #include <sys/cpu.h>
95 #include <sys/kmem.h>
96 #include <sys/vnode.h>
97 #include <sys/sysctl.h>
98 #include <sys/ktrace.h>
99 
100 /*
101  * A list (head) of open files, counter, and lock protecting them.
102  */
103 struct filelist		filehead	__cacheline_aligned;
104 static u_int		nfiles		__cacheline_aligned;
105 kmutex_t		filelist_lock	__cacheline_aligned;
106 
107 static pool_cache_t	filedesc_cache	__read_mostly;
108 static pool_cache_t	file_cache	__read_mostly;
109 static pool_cache_t	fdfile_cache	__read_mostly;
110 
111 static int	file_ctor(void *, void *, int);
112 static void	file_dtor(void *, void *);
113 static int	fdfile_ctor(void *, void *, int);
114 static void	fdfile_dtor(void *, void *);
115 static int	filedesc_ctor(void *, void *, int);
116 static void	filedesc_dtor(void *, void *);
117 static int	filedescopen(dev_t, int, int, lwp_t *);
118 
119 static int sysctl_kern_file(SYSCTLFN_PROTO);
120 static int sysctl_kern_file2(SYSCTLFN_PROTO);
121 static void fill_file(struct kinfo_file *, const file_t *, const fdfile_t *,
122 		      int, pid_t);
123 
124 const struct cdevsw filedesc_cdevsw = {
125 	filedescopen, noclose, noread, nowrite, noioctl,
126 	nostop, notty, nopoll, nommap, nokqfilter, D_OTHER | D_MPSAFE,
127 };
128 
129 /* For ease of reading. */
130 __strong_alias(fd_putvnode,fd_putfile)
131 __strong_alias(fd_putsock,fd_putfile)
132 
133 /*
134  * Initialize the descriptor system.
135  */
136 void
137 fd_sys_init(void)
138 {
139 	static struct sysctllog *clog;
140 
141 	mutex_init(&filelist_lock, MUTEX_DEFAULT, IPL_NONE);
142 
143 	file_cache = pool_cache_init(sizeof(file_t), coherency_unit, 0,
144 	    0, "file", NULL, IPL_NONE, file_ctor, file_dtor, NULL);
145 	KASSERT(file_cache != NULL);
146 
147 	fdfile_cache = pool_cache_init(sizeof(fdfile_t), coherency_unit, 0,
148 	    PR_LARGECACHE, "fdfile", NULL, IPL_NONE, fdfile_ctor, fdfile_dtor,
149 	    NULL);
150 	KASSERT(fdfile_cache != NULL);
151 
152 	filedesc_cache = pool_cache_init(sizeof(filedesc_t), coherency_unit,
153 	    0, 0, "filedesc", NULL, IPL_NONE, filedesc_ctor, filedesc_dtor,
154 	    NULL);
155 	KASSERT(filedesc_cache != NULL);
156 
157 	sysctl_createv(&clog, 0, NULL, NULL,
158 		       CTLFLAG_PERMANENT, CTLTYPE_NODE, "kern", NULL,
159 		       NULL, 0, NULL, 0, CTL_KERN, CTL_EOL);
160 	sysctl_createv(&clog, 0, NULL, NULL,
161 		       CTLFLAG_PERMANENT,
162 		       CTLTYPE_STRUCT, "file",
163 		       SYSCTL_DESCR("System open file table"),
164 		       sysctl_kern_file, 0, NULL, 0,
165 		       CTL_KERN, KERN_FILE, CTL_EOL);
166 	sysctl_createv(&clog, 0, NULL, NULL,
167 		       CTLFLAG_PERMANENT,
168 		       CTLTYPE_STRUCT, "file2",
169 		       SYSCTL_DESCR("System open file table"),
170 		       sysctl_kern_file2, 0, NULL, 0,
171 		       CTL_KERN, KERN_FILE2, CTL_EOL);
172 }
173 
174 static bool
175 fd_isused(filedesc_t *fdp, unsigned fd)
176 {
177 	u_int off = fd >> NDENTRYSHIFT;
178 
179 	KASSERT(fd < fdp->fd_dt->dt_nfiles);
180 
181 	return (fdp->fd_lomap[off] & (1 << (fd & NDENTRYMASK))) != 0;
182 }
183 
184 /*
185  * Verify that the bitmaps match the descriptor table.
186  */
187 static inline void
188 fd_checkmaps(filedesc_t *fdp)
189 {
190 #ifdef DEBUG
191 	fdtab_t *dt;
192 	u_int fd;
193 
194 	dt = fdp->fd_dt;
195 	if (fdp->fd_refcnt == -1) {
196 		/*
197 		 * fd_free tears down the table without maintaining its bitmap.
198 		 */
199 		return;
200 	}
201 	for (fd = 0; fd < dt->dt_nfiles; fd++) {
202 		if (fd < NDFDFILE) {
203 			KASSERT(dt->dt_ff[fd] ==
204 			    (fdfile_t *)fdp->fd_dfdfile[fd]);
205 		}
206 		if (dt->dt_ff[fd] == NULL) {
207 			KASSERT(!fd_isused(fdp, fd));
208 		} else if (dt->dt_ff[fd]->ff_file != NULL) {
209 			KASSERT(fd_isused(fdp, fd));
210 		}
211 	}
212 #endif
213 }
214 
215 static int
216 fd_next_zero(filedesc_t *fdp, uint32_t *bitmap, int want, u_int bits)
217 {
218 	int i, off, maxoff;
219 	uint32_t sub;
220 
221 	KASSERT(mutex_owned(&fdp->fd_lock));
222 
223 	fd_checkmaps(fdp);
224 
225 	if (want > bits)
226 		return -1;
227 
228 	off = want >> NDENTRYSHIFT;
229 	i = want & NDENTRYMASK;
230 	if (i) {
231 		sub = bitmap[off] | ((u_int)~0 >> (NDENTRIES - i));
232 		if (sub != ~0)
233 			goto found;
234 		off++;
235 	}
236 
237 	maxoff = NDLOSLOTS(bits);
238 	while (off < maxoff) {
239 		if ((sub = bitmap[off]) != ~0)
240 			goto found;
241 		off++;
242 	}
243 
244 	return -1;
245 
246  found:
247 	return (off << NDENTRYSHIFT) + ffs(~sub) - 1;
248 }
249 
250 static int
251 fd_last_set(filedesc_t *fd, int last)
252 {
253 	int off, i;
254 	fdfile_t **ff = fd->fd_dt->dt_ff;
255 	uint32_t *bitmap = fd->fd_lomap;
256 
257 	KASSERT(mutex_owned(&fd->fd_lock));
258 
259 	fd_checkmaps(fd);
260 
261 	off = (last - 1) >> NDENTRYSHIFT;
262 
263 	while (off >= 0 && !bitmap[off])
264 		off--;
265 
266 	if (off < 0)
267 		return -1;
268 
269 	i = ((off + 1) << NDENTRYSHIFT) - 1;
270 	if (i >= last)
271 		i = last - 1;
272 
273 	/* XXX should use bitmap */
274 	while (i > 0 && (ff[i] == NULL || !ff[i]->ff_allocated))
275 		i--;
276 
277 	return i;
278 }
279 
280 static inline void
281 fd_used(filedesc_t *fdp, unsigned fd)
282 {
283 	u_int off = fd >> NDENTRYSHIFT;
284 	fdfile_t *ff;
285 
286 	ff = fdp->fd_dt->dt_ff[fd];
287 
288 	KASSERT(mutex_owned(&fdp->fd_lock));
289 	KASSERT((fdp->fd_lomap[off] & (1 << (fd & NDENTRYMASK))) == 0);
290 	KASSERT(ff != NULL);
291 	KASSERT(ff->ff_file == NULL);
292 	KASSERT(!ff->ff_allocated);
293 
294 	ff->ff_allocated = 1;
295 	fdp->fd_lomap[off] |= 1 << (fd & NDENTRYMASK);
296 	if (__predict_false(fdp->fd_lomap[off] == ~0)) {
297 		KASSERT((fdp->fd_himap[off >> NDENTRYSHIFT] &
298 		    (1 << (off & NDENTRYMASK))) == 0);
299 		fdp->fd_himap[off >> NDENTRYSHIFT] |= 1 << (off & NDENTRYMASK);
300 	}
301 
302 	if ((int)fd > fdp->fd_lastfile) {
303 		fdp->fd_lastfile = fd;
304 	}
305 
306 	fd_checkmaps(fdp);
307 }
308 
309 static inline void
310 fd_unused(filedesc_t *fdp, unsigned fd)
311 {
312 	u_int off = fd >> NDENTRYSHIFT;
313 	fdfile_t *ff;
314 
315 	ff = fdp->fd_dt->dt_ff[fd];
316 
317 	/*
318 	 * Don't assert the lock is held here, as we may be copying
319 	 * the table during exec() and it is not needed there.
320 	 * procfs and sysctl are locked out by proc::p_reflock.
321 	 *
322 	 * KASSERT(mutex_owned(&fdp->fd_lock));
323 	 */
324 	KASSERT(ff != NULL);
325 	KASSERT(ff->ff_file == NULL);
326 	KASSERT(ff->ff_allocated);
327 
328 	if (fd < fdp->fd_freefile) {
329 		fdp->fd_freefile = fd;
330 	}
331 
332 	if (fdp->fd_lomap[off] == ~0) {
333 		KASSERT((fdp->fd_himap[off >> NDENTRYSHIFT] &
334 		    (1 << (off & NDENTRYMASK))) != 0);
335 		fdp->fd_himap[off >> NDENTRYSHIFT] &=
336 		    ~(1 << (off & NDENTRYMASK));
337 	}
338 	KASSERT((fdp->fd_lomap[off] & (1 << (fd & NDENTRYMASK))) != 0);
339 	fdp->fd_lomap[off] &= ~(1 << (fd & NDENTRYMASK));
340 	ff->ff_allocated = 0;
341 
342 	KASSERT(fd <= fdp->fd_lastfile);
343 	if (fd == fdp->fd_lastfile) {
344 		fdp->fd_lastfile = fd_last_set(fdp, fd);
345 	}
346 	fd_checkmaps(fdp);
347 }
348 
349 /*
350  * Look up the file structure corresponding to a file descriptor
351  * and return the file, holding a reference on the descriptor.
352  */
353 file_t *
354 fd_getfile(unsigned fd)
355 {
356 	filedesc_t *fdp;
357 	fdfile_t *ff;
358 	file_t *fp;
359 	fdtab_t *dt;
360 
361 	/*
362 	 * Look up the fdfile structure representing this descriptor.
363 	 * We are doing this unlocked.  See fd_tryexpand().
364 	 */
365 	fdp = curlwp->l_fd;
366 	dt = fdp->fd_dt;
367 	if (__predict_false(fd >= dt->dt_nfiles)) {
368 		return NULL;
369 	}
370 	ff = dt->dt_ff[fd];
371 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
372 	if (__predict_false(ff == NULL)) {
373 		return NULL;
374 	}
375 
376 	/* Now get a reference to the descriptor. */
377 	if (fdp->fd_refcnt == 1) {
378 		/*
379 		 * Single threaded: don't need to worry about concurrent
380 		 * access (other than earlier calls to kqueue, which may
381 		 * hold a reference to the descriptor).
382 		 */
383 		ff->ff_refcnt++;
384 	} else {
385 		/*
386 		 * Multi threaded: issue a memory barrier to ensure that we
387 		 * acquire the file pointer _after_ adding a reference.  If
388 		 * no memory barrier, we could fetch a stale pointer.
389 		 */
390 		atomic_inc_uint(&ff->ff_refcnt);
391 #ifndef __HAVE_ATOMIC_AS_MEMBAR
392 		membar_enter();
393 #endif
394 	}
395 
396 	/*
397 	 * If the file is not open or is being closed then put the
398 	 * reference back.
399 	 */
400 	fp = ff->ff_file;
401 	if (__predict_true(fp != NULL)) {
402 		return fp;
403 	}
404 	fd_putfile(fd);
405 	return NULL;
406 }
407 
408 /*
409  * Release a reference to a file descriptor acquired with fd_getfile().
410  */
411 void
412 fd_putfile(unsigned fd)
413 {
414 	filedesc_t *fdp;
415 	fdfile_t *ff;
416 	u_int u, v;
417 
418 	fdp = curlwp->l_fd;
419 	ff = fdp->fd_dt->dt_ff[fd];
420 
421 	KASSERT(fd < fdp->fd_dt->dt_nfiles);
422 	KASSERT(ff != NULL);
423 	KASSERT((ff->ff_refcnt & FR_MASK) > 0);
424 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
425 
426 	if (fdp->fd_refcnt == 1) {
427 		/*
428 		 * Single threaded: don't need to worry about concurrent
429 		 * access (other than earlier calls to kqueue, which may
430 		 * hold a reference to the descriptor).
431 		 */
432 		if (__predict_false((ff->ff_refcnt & FR_CLOSING) != 0)) {
433 			fd_close(fd);
434 			return;
435 		}
436 		ff->ff_refcnt--;
437 		return;
438 	}
439 
440 	/*
441 	 * Ensure that any use of the file is complete and globally
442 	 * visible before dropping the final reference.  If no membar,
443 	 * the current CPU could still access memory associated with
444 	 * the file after it has been freed or recycled by another
445 	 * CPU.
446 	 */
447 #ifndef __HAVE_ATOMIC_AS_MEMBAR
448 	membar_exit();
449 #endif
450 
451 	/*
452 	 * Be optimistic and start out with the assumption that no other
453 	 * threads are trying to close the descriptor.  If the CAS fails,
454 	 * we lost a race and/or it's being closed.
455 	 */
456 	for (u = ff->ff_refcnt & FR_MASK;; u = v) {
457 		v = atomic_cas_uint(&ff->ff_refcnt, u, u - 1);
458 		if (__predict_true(u == v)) {
459 			return;
460 		}
461 		if (__predict_false((v & FR_CLOSING) != 0)) {
462 			break;
463 		}
464 	}
465 
466 	/* Another thread is waiting to close the file: join it. */
467 	(void)fd_close(fd);
468 }
469 
470 /*
471  * Convenience wrapper around fd_getfile() that returns reference
472  * to a vnode.
473  */
474 int
475 fd_getvnode(unsigned fd, file_t **fpp)
476 {
477 	vnode_t *vp;
478 	file_t *fp;
479 
480 	fp = fd_getfile(fd);
481 	if (__predict_false(fp == NULL)) {
482 		return EBADF;
483 	}
484 	if (__predict_false(fp->f_type != DTYPE_VNODE)) {
485 		fd_putfile(fd);
486 		return EINVAL;
487 	}
488 	vp = fp->f_data;
489 	if (__predict_false(vp->v_type == VBAD)) {
490 		/* XXX Is this case really necessary? */
491 		fd_putfile(fd);
492 		return EBADF;
493 	}
494 	*fpp = fp;
495 	return 0;
496 }
497 
498 /*
499  * Convenience wrapper around fd_getfile() that returns reference
500  * to a socket.
501  */
502 int
503 fd_getsock(unsigned fd, struct socket **sop)
504 {
505 	file_t *fp;
506 
507 	fp = fd_getfile(fd);
508 	if (__predict_false(fp == NULL)) {
509 		return EBADF;
510 	}
511 	if (__predict_false(fp->f_type != DTYPE_SOCKET)) {
512 		fd_putfile(fd);
513 		return ENOTSOCK;
514 	}
515 	*sop = fp->f_data;
516 	return 0;
517 }
518 
519 /*
520  * Look up the file structure corresponding to a file descriptor
521  * and return it with a reference held on the file, not the
522  * descriptor.
523  *
524  * This is heavyweight and only used when accessing descriptors
525  * from a foreign process.  The caller must ensure that `p' does
526  * not exit or fork across this call.
527  *
528  * To release the file (not descriptor) reference, use closef().
529  */
530 file_t *
531 fd_getfile2(proc_t *p, unsigned fd)
532 {
533 	filedesc_t *fdp;
534 	fdfile_t *ff;
535 	file_t *fp;
536 	fdtab_t *dt;
537 
538 	fdp = p->p_fd;
539 	mutex_enter(&fdp->fd_lock);
540 	dt = fdp->fd_dt;
541 	if (fd >= dt->dt_nfiles) {
542 		mutex_exit(&fdp->fd_lock);
543 		return NULL;
544 	}
545 	if ((ff = dt->dt_ff[fd]) == NULL) {
546 		mutex_exit(&fdp->fd_lock);
547 		return NULL;
548 	}
549 	if ((fp = ff->ff_file) == NULL) {
550 		mutex_exit(&fdp->fd_lock);
551 		return NULL;
552 	}
553 	mutex_enter(&fp->f_lock);
554 	fp->f_count++;
555 	mutex_exit(&fp->f_lock);
556 	mutex_exit(&fdp->fd_lock);
557 
558 	return fp;
559 }
560 
561 /*
562  * Internal form of close.  Must be called with a reference to the
563  * descriptor, and will drop the reference.  When all descriptor
564  * references are dropped, releases the descriptor slot and a single
565  * reference to the file structure.
566  */
567 int
568 fd_close(unsigned fd)
569 {
570 	struct flock lf;
571 	filedesc_t *fdp;
572 	fdfile_t *ff;
573 	file_t *fp;
574 	proc_t *p;
575 	lwp_t *l;
576 	u_int refcnt;
577 
578 	l = curlwp;
579 	p = l->l_proc;
580 	fdp = l->l_fd;
581 	ff = fdp->fd_dt->dt_ff[fd];
582 
583 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
584 
585 	mutex_enter(&fdp->fd_lock);
586 	KASSERT((ff->ff_refcnt & FR_MASK) > 0);
587 	if (__predict_false(ff->ff_file == NULL)) {
588 		/*
589 		 * Another user of the file is already closing, and is
590 		 * waiting for other users of the file to drain.  Release
591 		 * our reference, and wake up the closer.
592 		 */
593 		atomic_dec_uint(&ff->ff_refcnt);
594 		cv_broadcast(&ff->ff_closing);
595 		mutex_exit(&fdp->fd_lock);
596 
597 		/*
598 		 * An application error, so pretend that the descriptor
599 		 * was already closed.  We can't safely wait for it to
600 		 * be closed without potentially deadlocking.
601 		 */
602 		return (EBADF);
603 	}
604 	KASSERT((ff->ff_refcnt & FR_CLOSING) == 0);
605 
606 	/*
607 	 * There may be multiple users of this file within the process.
608 	 * Notify existing and new users that the file is closing.  This
609 	 * will prevent them from adding additional uses to this file
610 	 * while we are closing it.
611 	 */
612 	fp = ff->ff_file;
613 	ff->ff_file = NULL;
614 	ff->ff_exclose = false;
615 
616 	/*
617 	 * We expect the caller to hold a descriptor reference - drop it.
618 	 * The reference count may increase beyond zero at this point due
619 	 * to an erroneous descriptor reference by an application, but
620 	 * fd_getfile() will notice that the file is being closed and drop
621 	 * the reference again.
622 	 */
623 	if (fdp->fd_refcnt == 1) {
624 		/* Single threaded. */
625 		refcnt = --(ff->ff_refcnt);
626 	} else {
627 		/* Multi threaded. */
628 #ifndef __HAVE_ATOMIC_AS_MEMBAR
629 		membar_producer();
630 #endif
631 		refcnt = atomic_dec_uint_nv(&ff->ff_refcnt);
632 	}
633 	if (__predict_false(refcnt != 0)) {
634 		/*
635 		 * Wait for other references to drain.  This is typically
636 		 * an application error - the descriptor is being closed
637 		 * while still in use.
638 		 * (Or just a threaded application trying to unblock its
639 		 * thread that sleeps in (say) accept()).
640 		 */
641 		atomic_or_uint(&ff->ff_refcnt, FR_CLOSING);
642 
643 		/*
644 		 * Remove any knotes attached to the file.  A knote
645 		 * attached to the descriptor can hold references on it.
646 		 */
647 		mutex_exit(&fdp->fd_lock);
648 		if (!SLIST_EMPTY(&ff->ff_knlist)) {
649 			knote_fdclose(fd);
650 		}
651 
652 		/*
653 		 * Since the file system code doesn't know which fd
654 		 * each request came from (think dup()), we have to
655 		 * ask it to return ERESTART for any long-term blocks.
656 		 * The re-entry through read/write/etc will detect the
657 		 * closed fd and return EBAFD.
658 		 * Blocked partial writes may return a short length.
659 		 */
660 		(*fp->f_ops->fo_restart)(fp);
661 		mutex_enter(&fdp->fd_lock);
662 
663 		/*
664 		 * We need to see the count drop to zero at least once,
665 		 * in order to ensure that all pre-existing references
666 		 * have been drained.  New references past this point are
667 		 * of no interest.
668 		 * XXX (dsl) this may need to call fo_restart() after a
669 		 * timeout to guarantee that all the system calls exit.
670 		 */
671 		while ((ff->ff_refcnt & FR_MASK) != 0) {
672 			cv_wait(&ff->ff_closing, &fdp->fd_lock);
673 		}
674 		atomic_and_uint(&ff->ff_refcnt, ~FR_CLOSING);
675 	} else {
676 		/* If no references, there must be no knotes. */
677 		KASSERT(SLIST_EMPTY(&ff->ff_knlist));
678 	}
679 
680 	/*
681 	 * POSIX record locking dictates that any close releases ALL
682 	 * locks owned by this process.  This is handled by setting
683 	 * a flag in the unlock to free ONLY locks obeying POSIX
684 	 * semantics, and not to free BSD-style file locks.
685 	 * If the descriptor was in a message, POSIX-style locks
686 	 * aren't passed with the descriptor.
687 	 */
688 	if (__predict_false((p->p_flag & PK_ADVLOCK) != 0 &&
689 	    fp->f_type == DTYPE_VNODE)) {
690 		lf.l_whence = SEEK_SET;
691 		lf.l_start = 0;
692 		lf.l_len = 0;
693 		lf.l_type = F_UNLCK;
694 		mutex_exit(&fdp->fd_lock);
695 		(void)VOP_ADVLOCK(fp->f_data, p, F_UNLCK, &lf, F_POSIX);
696 		mutex_enter(&fdp->fd_lock);
697 	}
698 
699 	/* Free descriptor slot. */
700 	fd_unused(fdp, fd);
701 	mutex_exit(&fdp->fd_lock);
702 
703 	/* Now drop reference to the file itself. */
704 	return closef(fp);
705 }
706 
707 /*
708  * Duplicate a file descriptor.
709  */
710 int
711 fd_dup(file_t *fp, int minfd, int *newp, bool exclose)
712 {
713 	proc_t *p = curproc;
714 	int error;
715 
716 	while ((error = fd_alloc(p, minfd, newp)) != 0) {
717 		if (error != ENOSPC) {
718 			return error;
719 		}
720 		fd_tryexpand(p);
721 	}
722 
723 	curlwp->l_fd->fd_dt->dt_ff[*newp]->ff_exclose = exclose;
724 	fd_affix(p, fp, *newp);
725 	return 0;
726 }
727 
728 /*
729  * dup2 operation.
730  */
731 int
732 fd_dup2(file_t *fp, unsigned new, int flags)
733 {
734 	filedesc_t *fdp = curlwp->l_fd;
735 	fdfile_t *ff;
736 	fdtab_t *dt;
737 
738 	/*
739 	 * Ensure there are enough slots in the descriptor table,
740 	 * and allocate an fdfile_t up front in case we need it.
741 	 */
742 	while (new >= fdp->fd_dt->dt_nfiles) {
743 		fd_tryexpand(curproc);
744 	}
745 	ff = pool_cache_get(fdfile_cache, PR_WAITOK);
746 
747 	/*
748 	 * If there is already a file open, close it.  If the file is
749 	 * half open, wait for it to be constructed before closing it.
750 	 * XXX Potential for deadlock here?
751 	 */
752 	mutex_enter(&fdp->fd_lock);
753 	while (fd_isused(fdp, new)) {
754 		mutex_exit(&fdp->fd_lock);
755 		if (fd_getfile(new) != NULL) {
756 			(void)fd_close(new);
757 		} else {
758 			/*
759 			 * Crummy, but unlikely to happen.
760 			 * Can occur if we interrupt another
761 			 * thread while it is opening a file.
762 			 */
763 			kpause("dup2", false, 1, NULL);
764 		}
765 		mutex_enter(&fdp->fd_lock);
766 	}
767 	dt = fdp->fd_dt;
768 	if (dt->dt_ff[new] == NULL) {
769 		KASSERT(new >= NDFDFILE);
770 		dt->dt_ff[new] = ff;
771 		ff = NULL;
772 	}
773 	fd_used(fdp, new);
774 	mutex_exit(&fdp->fd_lock);
775 
776 	dt->dt_ff[new]->ff_exclose = (flags & O_CLOEXEC) != 0;
777 	fp->f_flag |= flags & FNONBLOCK;
778 	/* Slot is now allocated.  Insert copy of the file. */
779 	fd_affix(curproc, fp, new);
780 	if (ff != NULL) {
781 		pool_cache_put(fdfile_cache, ff);
782 	}
783 	return 0;
784 }
785 
786 /*
787  * Drop reference to a file structure.
788  */
789 int
790 closef(file_t *fp)
791 {
792 	struct flock lf;
793 	int error;
794 
795 	/*
796 	 * Drop reference.  If referenced elsewhere it's still open
797 	 * and we have nothing more to do.
798 	 */
799 	mutex_enter(&fp->f_lock);
800 	KASSERT(fp->f_count > 0);
801 	if (--fp->f_count > 0) {
802 		mutex_exit(&fp->f_lock);
803 		return 0;
804 	}
805 	KASSERT(fp->f_count == 0);
806 	mutex_exit(&fp->f_lock);
807 
808 	/* We held the last reference - release locks, close and free. */
809 	if ((fp->f_flag & FHASLOCK) && fp->f_type == DTYPE_VNODE) {
810 		lf.l_whence = SEEK_SET;
811 		lf.l_start = 0;
812 		lf.l_len = 0;
813 		lf.l_type = F_UNLCK;
814 		(void)VOP_ADVLOCK(fp->f_data, fp, F_UNLCK, &lf, F_FLOCK);
815 	}
816 	if (fp->f_ops != NULL) {
817 		error = (*fp->f_ops->fo_close)(fp);
818 	} else {
819 		error = 0;
820 	}
821 	KASSERT(fp->f_count == 0);
822 	KASSERT(fp->f_cred != NULL);
823 	pool_cache_put(file_cache, fp);
824 
825 	return error;
826 }
827 
828 /*
829  * Allocate a file descriptor for the process.
830  */
831 int
832 fd_alloc(proc_t *p, int want, int *result)
833 {
834 	filedesc_t *fdp = p->p_fd;
835 	int i, lim, last, error;
836 	u_int off, new;
837 	fdtab_t *dt;
838 
839 	KASSERT(p == curproc || p == &proc0);
840 
841 	/*
842 	 * Search for a free descriptor starting at the higher
843 	 * of want or fd_freefile.
844 	 */
845 	mutex_enter(&fdp->fd_lock);
846 	fd_checkmaps(fdp);
847 	dt = fdp->fd_dt;
848 	KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
849 	lim = min((int)p->p_rlimit[RLIMIT_NOFILE].rlim_cur, maxfiles);
850 	last = min(dt->dt_nfiles, lim);
851 	for (;;) {
852 		if ((i = want) < fdp->fd_freefile)
853 			i = fdp->fd_freefile;
854 		off = i >> NDENTRYSHIFT;
855 		new = fd_next_zero(fdp, fdp->fd_himap, off,
856 		    (last + NDENTRIES - 1) >> NDENTRYSHIFT);
857 		if (new == -1)
858 			break;
859 		i = fd_next_zero(fdp, &fdp->fd_lomap[new],
860 		    new > off ? 0 : i & NDENTRYMASK, NDENTRIES);
861 		if (i == -1) {
862 			/*
863 			 * Free file descriptor in this block was
864 			 * below want, try again with higher want.
865 			 */
866 			want = (new + 1) << NDENTRYSHIFT;
867 			continue;
868 		}
869 		i += (new << NDENTRYSHIFT);
870 		if (i >= last) {
871 			break;
872 		}
873 		if (dt->dt_ff[i] == NULL) {
874 			KASSERT(i >= NDFDFILE);
875 			dt->dt_ff[i] = pool_cache_get(fdfile_cache, PR_WAITOK);
876 		}
877 		KASSERT(dt->dt_ff[i]->ff_file == NULL);
878 		fd_used(fdp, i);
879 		if (want <= fdp->fd_freefile) {
880 			fdp->fd_freefile = i;
881 		}
882 		*result = i;
883 		KASSERT(i >= NDFDFILE ||
884 		    dt->dt_ff[i] == (fdfile_t *)fdp->fd_dfdfile[i]);
885 		fd_checkmaps(fdp);
886 		mutex_exit(&fdp->fd_lock);
887 		return 0;
888 	}
889 
890 	/* No space in current array.  Let the caller expand and retry. */
891 	error = (dt->dt_nfiles >= lim) ? EMFILE : ENOSPC;
892 	mutex_exit(&fdp->fd_lock);
893 	return error;
894 }
895 
896 /*
897  * Allocate memory for a descriptor table.
898  */
899 static fdtab_t *
900 fd_dtab_alloc(int n)
901 {
902 	fdtab_t *dt;
903 	size_t sz;
904 
905 	KASSERT(n > NDFILE);
906 
907 	sz = sizeof(*dt) + (n - NDFILE) * sizeof(dt->dt_ff[0]);
908 	dt = kmem_alloc(sz, KM_SLEEP);
909 #ifdef DIAGNOSTIC
910 	memset(dt, 0xff, sz);
911 #endif
912 	dt->dt_nfiles = n;
913 	dt->dt_link = NULL;
914 	return dt;
915 }
916 
917 /*
918  * Free a descriptor table, and all tables linked for deferred free.
919  */
920 static void
921 fd_dtab_free(fdtab_t *dt)
922 {
923 	fdtab_t *next;
924 	size_t sz;
925 
926 	do {
927 		next = dt->dt_link;
928 		KASSERT(dt->dt_nfiles > NDFILE);
929 		sz = sizeof(*dt) +
930 		    (dt->dt_nfiles - NDFILE) * sizeof(dt->dt_ff[0]);
931 #ifdef DIAGNOSTIC
932 		memset(dt, 0xff, sz);
933 #endif
934 		kmem_free(dt, sz);
935 		dt = next;
936 	} while (dt != NULL);
937 }
938 
939 /*
940  * Allocate descriptor bitmap.
941  */
942 static void
943 fd_map_alloc(int n, uint32_t **lo, uint32_t **hi)
944 {
945 	uint8_t *ptr;
946 	size_t szlo, szhi;
947 
948 	KASSERT(n > NDENTRIES);
949 
950 	szlo = NDLOSLOTS(n) * sizeof(uint32_t);
951 	szhi = NDHISLOTS(n) * sizeof(uint32_t);
952 	ptr = kmem_alloc(szlo + szhi, KM_SLEEP);
953 	*lo = (uint32_t *)ptr;
954 	*hi = (uint32_t *)(ptr + szlo);
955 }
956 
957 /*
958  * Free descriptor bitmap.
959  */
960 static void
961 fd_map_free(int n, uint32_t *lo, uint32_t *hi)
962 {
963 	size_t szlo, szhi;
964 
965 	KASSERT(n > NDENTRIES);
966 
967 	szlo = NDLOSLOTS(n) * sizeof(uint32_t);
968 	szhi = NDHISLOTS(n) * sizeof(uint32_t);
969 	KASSERT(hi == (uint32_t *)((uint8_t *)lo + szlo));
970 	kmem_free(lo, szlo + szhi);
971 }
972 
973 /*
974  * Expand a process' descriptor table.
975  */
976 void
977 fd_tryexpand(proc_t *p)
978 {
979 	filedesc_t *fdp;
980 	int i, numfiles, oldnfiles;
981 	fdtab_t *newdt, *dt;
982 	uint32_t *newhimap, *newlomap;
983 
984 	KASSERT(p == curproc || p == &proc0);
985 
986 	fdp = p->p_fd;
987 	newhimap = NULL;
988 	newlomap = NULL;
989 	oldnfiles = fdp->fd_dt->dt_nfiles;
990 
991 	if (oldnfiles < NDEXTENT)
992 		numfiles = NDEXTENT;
993 	else
994 		numfiles = 2 * oldnfiles;
995 
996 	newdt = fd_dtab_alloc(numfiles);
997 	if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
998 		fd_map_alloc(numfiles, &newlomap, &newhimap);
999 	}
1000 
1001 	mutex_enter(&fdp->fd_lock);
1002 	dt = fdp->fd_dt;
1003 	KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1004 	if (dt->dt_nfiles != oldnfiles) {
1005 		/* fdp changed; caller must retry */
1006 		mutex_exit(&fdp->fd_lock);
1007 		fd_dtab_free(newdt);
1008 		if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1009 			fd_map_free(numfiles, newlomap, newhimap);
1010 		}
1011 		return;
1012 	}
1013 
1014 	/* Copy the existing descriptor table and zero the new portion. */
1015 	i = sizeof(fdfile_t *) * oldnfiles;
1016 	memcpy(newdt->dt_ff, dt->dt_ff, i);
1017 	memset((uint8_t *)newdt->dt_ff + i, 0,
1018 	    numfiles * sizeof(fdfile_t *) - i);
1019 
1020 	/*
1021 	 * Link old descriptor array into list to be discarded.  We defer
1022 	 * freeing until the last reference to the descriptor table goes
1023 	 * away (usually process exit).  This allows us to do lockless
1024 	 * lookups in fd_getfile().
1025 	 */
1026 	if (oldnfiles > NDFILE) {
1027 		if (fdp->fd_refcnt > 1) {
1028 			newdt->dt_link = dt;
1029 		} else {
1030 			fd_dtab_free(dt);
1031 		}
1032 	}
1033 
1034 	if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1035 		i = NDHISLOTS(oldnfiles) * sizeof(uint32_t);
1036 		memcpy(newhimap, fdp->fd_himap, i);
1037 		memset((uint8_t *)newhimap + i, 0,
1038 		    NDHISLOTS(numfiles) * sizeof(uint32_t) - i);
1039 
1040 		i = NDLOSLOTS(oldnfiles) * sizeof(uint32_t);
1041 		memcpy(newlomap, fdp->fd_lomap, i);
1042 		memset((uint8_t *)newlomap + i, 0,
1043 		    NDLOSLOTS(numfiles) * sizeof(uint32_t) - i);
1044 
1045 		if (NDHISLOTS(oldnfiles) > NDHISLOTS(NDFILE)) {
1046 			fd_map_free(oldnfiles, fdp->fd_lomap, fdp->fd_himap);
1047 		}
1048 		fdp->fd_himap = newhimap;
1049 		fdp->fd_lomap = newlomap;
1050 	}
1051 
1052 	/*
1053 	 * All other modifications must become globally visible before
1054 	 * the change to fd_dt.  See fd_getfile().
1055 	 */
1056 	membar_producer();
1057 	fdp->fd_dt = newdt;
1058 	KASSERT(newdt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1059 	fd_checkmaps(fdp);
1060 	mutex_exit(&fdp->fd_lock);
1061 }
1062 
1063 /*
1064  * Create a new open file structure and allocate a file descriptor
1065  * for the current process.
1066  */
1067 int
1068 fd_allocfile(file_t **resultfp, int *resultfd)
1069 {
1070 	proc_t *p = curproc;
1071 	kauth_cred_t cred;
1072 	file_t *fp;
1073 	int error;
1074 
1075 	while ((error = fd_alloc(p, 0, resultfd)) != 0) {
1076 		if (error != ENOSPC) {
1077 			return error;
1078 		}
1079 		fd_tryexpand(p);
1080 	}
1081 
1082 	fp = pool_cache_get(file_cache, PR_WAITOK);
1083 	if (fp == NULL) {
1084 		return ENFILE;
1085 	}
1086 	KASSERT(fp->f_count == 0);
1087 	KASSERT(fp->f_msgcount == 0);
1088 	KASSERT(fp->f_unpcount == 0);
1089 
1090 	/* Replace cached credentials if not what we need. */
1091 	cred = curlwp->l_cred;
1092 	if (__predict_false(cred != fp->f_cred)) {
1093 		kauth_cred_free(fp->f_cred);
1094 		kauth_cred_hold(cred);
1095 		fp->f_cred = cred;
1096 	}
1097 
1098 	/*
1099 	 * Don't allow recycled files to be scanned.
1100 	 * See uipc_usrreq.c.
1101 	 */
1102 	if (__predict_false((fp->f_flag & FSCAN) != 0)) {
1103 		mutex_enter(&fp->f_lock);
1104 		atomic_and_uint(&fp->f_flag, ~FSCAN);
1105 		mutex_exit(&fp->f_lock);
1106 	}
1107 
1108 	fp->f_advice = 0;
1109 	fp->f_offset = 0;
1110 	*resultfp = fp;
1111 
1112 	return 0;
1113 }
1114 
1115 /*
1116  * Successful creation of a new descriptor: make visible to the process.
1117  */
1118 void
1119 fd_affix(proc_t *p, file_t *fp, unsigned fd)
1120 {
1121 	fdfile_t *ff;
1122 	filedesc_t *fdp;
1123 
1124 	KASSERT(p == curproc || p == &proc0);
1125 
1126 	/* Add a reference to the file structure. */
1127 	mutex_enter(&fp->f_lock);
1128 	fp->f_count++;
1129 	mutex_exit(&fp->f_lock);
1130 
1131 	/*
1132 	 * Insert the new file into the descriptor slot.
1133 	 *
1134 	 * The memory barriers provided by lock activity in this routine
1135 	 * ensure that any updates to the file structure become globally
1136 	 * visible before the file becomes visible to other LWPs in the
1137 	 * current process.
1138 	 */
1139 	fdp = p->p_fd;
1140 	ff = fdp->fd_dt->dt_ff[fd];
1141 
1142 	KASSERT(ff != NULL);
1143 	KASSERT(ff->ff_file == NULL);
1144 	KASSERT(ff->ff_allocated);
1145 	KASSERT(fd_isused(fdp, fd));
1146 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1147 
1148 	/* No need to lock in order to make file initially visible. */
1149 	ff->ff_file = fp;
1150 }
1151 
1152 /*
1153  * Abort creation of a new descriptor: free descriptor slot and file.
1154  */
1155 void
1156 fd_abort(proc_t *p, file_t *fp, unsigned fd)
1157 {
1158 	filedesc_t *fdp;
1159 	fdfile_t *ff;
1160 
1161 	KASSERT(p == curproc || p == &proc0);
1162 
1163 	fdp = p->p_fd;
1164 	ff = fdp->fd_dt->dt_ff[fd];
1165 
1166 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1167 
1168 	mutex_enter(&fdp->fd_lock);
1169 	KASSERT(fd_isused(fdp, fd));
1170 	fd_unused(fdp, fd);
1171 	mutex_exit(&fdp->fd_lock);
1172 
1173 	if (fp != NULL) {
1174 		KASSERT(fp->f_count == 0);
1175 		KASSERT(fp->f_cred != NULL);
1176 		pool_cache_put(file_cache, fp);
1177 	}
1178 }
1179 
1180 static int
1181 file_ctor(void *arg, void *obj, int flags)
1182 {
1183 	file_t *fp = obj;
1184 
1185 	memset(fp, 0, sizeof(*fp));
1186 
1187 	mutex_enter(&filelist_lock);
1188 	if (__predict_false(nfiles >= maxfiles)) {
1189 		mutex_exit(&filelist_lock);
1190 		tablefull("file", "increase kern.maxfiles or MAXFILES");
1191 		return ENFILE;
1192 	}
1193 	nfiles++;
1194 	LIST_INSERT_HEAD(&filehead, fp, f_list);
1195 	mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1196 	fp->f_cred = curlwp->l_cred;
1197 	kauth_cred_hold(fp->f_cred);
1198 	mutex_exit(&filelist_lock);
1199 
1200 	return 0;
1201 }
1202 
1203 static void
1204 file_dtor(void *arg, void *obj)
1205 {
1206 	file_t *fp = obj;
1207 
1208 	mutex_enter(&filelist_lock);
1209 	nfiles--;
1210 	LIST_REMOVE(fp, f_list);
1211 	mutex_exit(&filelist_lock);
1212 
1213 	kauth_cred_free(fp->f_cred);
1214 	mutex_destroy(&fp->f_lock);
1215 }
1216 
1217 static int
1218 fdfile_ctor(void *arg, void *obj, int flags)
1219 {
1220 	fdfile_t *ff = obj;
1221 
1222 	memset(ff, 0, sizeof(*ff));
1223 	cv_init(&ff->ff_closing, "fdclose");
1224 
1225 	return 0;
1226 }
1227 
1228 static void
1229 fdfile_dtor(void *arg, void *obj)
1230 {
1231 	fdfile_t *ff = obj;
1232 
1233 	cv_destroy(&ff->ff_closing);
1234 }
1235 
1236 file_t *
1237 fgetdummy(void)
1238 {
1239 	file_t *fp;
1240 
1241 	fp = kmem_zalloc(sizeof(*fp), KM_SLEEP);
1242 	if (fp != NULL) {
1243 		mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1244 	}
1245 	return fp;
1246 }
1247 
1248 void
1249 fputdummy(file_t *fp)
1250 {
1251 
1252 	mutex_destroy(&fp->f_lock);
1253 	kmem_free(fp, sizeof(*fp));
1254 }
1255 
1256 /*
1257  * Create an initial filedesc structure.
1258  */
1259 filedesc_t *
1260 fd_init(filedesc_t *fdp)
1261 {
1262 #ifdef DIAGNOSTIC
1263 	unsigned fd;
1264 #endif
1265 
1266 	if (__predict_true(fdp == NULL)) {
1267 		fdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1268 	} else {
1269 		KASSERT(fdp == &filedesc0);
1270 		filedesc_ctor(NULL, fdp, PR_WAITOK);
1271 	}
1272 
1273 #ifdef DIAGNOSTIC
1274 	KASSERT(fdp->fd_lastfile == -1);
1275 	KASSERT(fdp->fd_lastkqfile == -1);
1276 	KASSERT(fdp->fd_knhash == NULL);
1277 	KASSERT(fdp->fd_freefile == 0);
1278 	KASSERT(fdp->fd_exclose == false);
1279 	KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1280 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1281 	for (fd = 0; fd < NDFDFILE; fd++) {
1282 		KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] ==
1283 		    (fdfile_t *)fdp->fd_dfdfile[fd]);
1284 	}
1285 	for (fd = NDFDFILE; fd < NDFILE; fd++) {
1286 		KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] == NULL);
1287 	}
1288 	KASSERT(fdp->fd_himap == fdp->fd_dhimap);
1289 	KASSERT(fdp->fd_lomap == fdp->fd_dlomap);
1290 #endif	/* DIAGNOSTIC */
1291 
1292 	fdp->fd_refcnt = 1;
1293 	fd_checkmaps(fdp);
1294 
1295 	return fdp;
1296 }
1297 
1298 /*
1299  * Initialize a file descriptor table.
1300  */
1301 static int
1302 filedesc_ctor(void *arg, void *obj, int flag)
1303 {
1304 	filedesc_t *fdp = obj;
1305 	fdfile_t **ffp;
1306 	int i;
1307 
1308 	memset(fdp, 0, sizeof(*fdp));
1309 	mutex_init(&fdp->fd_lock, MUTEX_DEFAULT, IPL_NONE);
1310 	fdp->fd_lastfile = -1;
1311 	fdp->fd_lastkqfile = -1;
1312 	fdp->fd_dt = &fdp->fd_dtbuiltin;
1313 	fdp->fd_dtbuiltin.dt_nfiles = NDFILE;
1314 	fdp->fd_himap = fdp->fd_dhimap;
1315 	fdp->fd_lomap = fdp->fd_dlomap;
1316 
1317 	CTASSERT(sizeof(fdp->fd_dfdfile[0]) >= sizeof(fdfile_t));
1318 	for (i = 0, ffp = fdp->fd_dt->dt_ff; i < NDFDFILE; i++, ffp++) {
1319 		*ffp = (fdfile_t *)fdp->fd_dfdfile[i];
1320 		(void)fdfile_ctor(NULL, fdp->fd_dfdfile[i], PR_WAITOK);
1321 	}
1322 
1323 	return 0;
1324 }
1325 
1326 static void
1327 filedesc_dtor(void *arg, void *obj)
1328 {
1329 	filedesc_t *fdp = obj;
1330 	int i;
1331 
1332 	for (i = 0; i < NDFDFILE; i++) {
1333 		fdfile_dtor(NULL, fdp->fd_dfdfile[i]);
1334 	}
1335 
1336 	mutex_destroy(&fdp->fd_lock);
1337 }
1338 
1339 /*
1340  * Make p share curproc's filedesc structure.
1341  */
1342 void
1343 fd_share(struct proc *p)
1344 {
1345 	filedesc_t *fdp;
1346 
1347 	fdp = curlwp->l_fd;
1348 	p->p_fd = fdp;
1349 	atomic_inc_uint(&fdp->fd_refcnt);
1350 }
1351 
1352 /*
1353  * Acquire a hold on a filedesc structure.
1354  */
1355 void
1356 fd_hold(lwp_t *l)
1357 {
1358 	filedesc_t *fdp = l->l_fd;
1359 
1360 	atomic_inc_uint(&fdp->fd_refcnt);
1361 }
1362 
1363 /*
1364  * Copy a filedesc structure.
1365  */
1366 filedesc_t *
1367 fd_copy(void)
1368 {
1369 	filedesc_t *newfdp, *fdp;
1370 	fdfile_t *ff, **ffp, **nffp, *ff2;
1371 	int i, j, numfiles, lastfile, newlast;
1372 	file_t *fp;
1373 	fdtab_t *newdt;
1374 
1375 	fdp = curproc->p_fd;
1376 	newfdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1377 	newfdp->fd_refcnt = 1;
1378 
1379 #ifdef DIAGNOSTIC
1380 	KASSERT(newfdp->fd_lastfile == -1);
1381 	KASSERT(newfdp->fd_lastkqfile == -1);
1382 	KASSERT(newfdp->fd_knhash == NULL);
1383 	KASSERT(newfdp->fd_freefile == 0);
1384 	KASSERT(newfdp->fd_exclose == false);
1385 	KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1386 	KASSERT(newfdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1387 	for (i = 0; i < NDFDFILE; i++) {
1388 		KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] ==
1389 		    (fdfile_t *)&newfdp->fd_dfdfile[i]);
1390 	}
1391 	for (i = NDFDFILE; i < NDFILE; i++) {
1392 		KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] == NULL);
1393 	}
1394 #endif	/* DIAGNOSTIC */
1395 
1396 	mutex_enter(&fdp->fd_lock);
1397 	fd_checkmaps(fdp);
1398 	numfiles = fdp->fd_dt->dt_nfiles;
1399 	lastfile = fdp->fd_lastfile;
1400 
1401 	/*
1402 	 * If the number of open files fits in the internal arrays
1403 	 * of the open file structure, use them, otherwise allocate
1404 	 * additional memory for the number of descriptors currently
1405 	 * in use.
1406 	 */
1407 	if (lastfile < NDFILE) {
1408 		i = NDFILE;
1409 		newdt = newfdp->fd_dt;
1410 		KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1411 	} else {
1412 		/*
1413 		 * Compute the smallest multiple of NDEXTENT needed
1414 		 * for the file descriptors currently in use,
1415 		 * allowing the table to shrink.
1416 		 */
1417 		i = numfiles;
1418 		while (i >= 2 * NDEXTENT && i > lastfile * 2) {
1419 			i /= 2;
1420 		}
1421 		KASSERT(i > NDFILE);
1422 		newdt = fd_dtab_alloc(i);
1423 		newfdp->fd_dt = newdt;
1424 		memcpy(newdt->dt_ff, newfdp->fd_dtbuiltin.dt_ff,
1425 		    NDFDFILE * sizeof(fdfile_t **));
1426 		memset(newdt->dt_ff + NDFDFILE, 0,
1427 		    (i - NDFDFILE) * sizeof(fdfile_t **));
1428 	}
1429 	if (NDHISLOTS(i) <= NDHISLOTS(NDFILE)) {
1430 		newfdp->fd_himap = newfdp->fd_dhimap;
1431 		newfdp->fd_lomap = newfdp->fd_dlomap;
1432 	} else {
1433 		fd_map_alloc(i, &newfdp->fd_lomap, &newfdp->fd_himap);
1434 		KASSERT(i >= NDENTRIES * NDENTRIES);
1435 		memset(newfdp->fd_himap, 0, NDHISLOTS(i)*sizeof(uint32_t));
1436 		memset(newfdp->fd_lomap, 0, NDLOSLOTS(i)*sizeof(uint32_t));
1437 	}
1438 	newfdp->fd_freefile = fdp->fd_freefile;
1439 	newfdp->fd_exclose = fdp->fd_exclose;
1440 
1441 	ffp = fdp->fd_dt->dt_ff;
1442 	nffp = newdt->dt_ff;
1443 	newlast = -1;
1444 	for (i = 0; i <= (int)lastfile; i++, ffp++, nffp++) {
1445 		KASSERT(i >= NDFDFILE ||
1446 		    *nffp == (fdfile_t *)newfdp->fd_dfdfile[i]);
1447 		ff = *ffp;
1448 		if (ff == NULL || (fp = ff->ff_file) == NULL) {
1449 			/* Descriptor unused, or descriptor half open. */
1450 			KASSERT(!fd_isused(newfdp, i));
1451 			continue;
1452 		}
1453 		if (__predict_false(fp->f_type == DTYPE_KQUEUE)) {
1454 			/* kqueue descriptors cannot be copied. */
1455 			if (i < newfdp->fd_freefile) {
1456 				newfdp->fd_freefile = i;
1457 			}
1458 			continue;
1459 		}
1460 		/* It's active: add a reference to the file. */
1461 		mutex_enter(&fp->f_lock);
1462 		fp->f_count++;
1463 		mutex_exit(&fp->f_lock);
1464 
1465 		/* Allocate an fdfile_t to represent it. */
1466 		if (i >= NDFDFILE) {
1467 			ff2 = pool_cache_get(fdfile_cache, PR_WAITOK);
1468 			*nffp = ff2;
1469 		} else {
1470 			ff2 = newdt->dt_ff[i];
1471 		}
1472 		ff2->ff_file = fp;
1473 		ff2->ff_exclose = ff->ff_exclose;
1474 		ff2->ff_allocated = true;
1475 
1476 		/* Fix up bitmaps. */
1477 		j = i >> NDENTRYSHIFT;
1478 		KASSERT((newfdp->fd_lomap[j] & (1 << (i & NDENTRYMASK))) == 0);
1479 		newfdp->fd_lomap[j] |= 1 << (i & NDENTRYMASK);
1480 		if (__predict_false(newfdp->fd_lomap[j] == ~0)) {
1481 			KASSERT((newfdp->fd_himap[j >> NDENTRYSHIFT] &
1482 			    (1 << (j & NDENTRYMASK))) == 0);
1483 			newfdp->fd_himap[j >> NDENTRYSHIFT] |=
1484 			    1 << (j & NDENTRYMASK);
1485 		}
1486 		newlast = i;
1487 	}
1488 	KASSERT(newdt->dt_ff[0] == (fdfile_t *)newfdp->fd_dfdfile[0]);
1489 	newfdp->fd_lastfile = newlast;
1490 	fd_checkmaps(newfdp);
1491 	mutex_exit(&fdp->fd_lock);
1492 
1493 	return newfdp;
1494 }
1495 
1496 /*
1497  * Release a filedesc structure.
1498  */
1499 void
1500 fd_free(void)
1501 {
1502 	fdfile_t *ff;
1503 	file_t *fp;
1504 	int fd, nf;
1505 	fdtab_t *dt;
1506 	lwp_t * const l = curlwp;
1507 	filedesc_t * const fdp = l->l_fd;
1508 	const bool noadvlock = (l->l_proc->p_flag & PK_ADVLOCK) == 0;
1509 
1510 	KASSERT(fdp->fd_dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1511 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1512 	KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1513 
1514 #ifndef __HAVE_ATOMIC_AS_MEMBAR
1515 	membar_exit();
1516 #endif
1517 	if (atomic_dec_uint_nv(&fdp->fd_refcnt) > 0)
1518 		return;
1519 
1520 	/*
1521 	 * Close any files that the process holds open.
1522 	 */
1523 	dt = fdp->fd_dt;
1524 	fd_checkmaps(fdp);
1525 #ifdef DEBUG
1526 	fdp->fd_refcnt = -1; /* see fd_checkmaps */
1527 #endif
1528 	for (fd = 0, nf = dt->dt_nfiles; fd < nf; fd++) {
1529 		ff = dt->dt_ff[fd];
1530 		KASSERT(fd >= NDFDFILE ||
1531 		    ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1532 		if (ff == NULL)
1533 			continue;
1534 		if ((fp = ff->ff_file) != NULL) {
1535 			/*
1536 			 * Must use fd_close() here if there is
1537 			 * a reference from kqueue or we might have posix
1538 			 * advisory locks.
1539 			 */
1540 			if (__predict_true(ff->ff_refcnt == 0) &&
1541 			    (noadvlock || fp->f_type != DTYPE_VNODE)) {
1542 				ff->ff_file = NULL;
1543 				ff->ff_exclose = false;
1544 				ff->ff_allocated = false;
1545 				closef(fp);
1546 			} else {
1547 				ff->ff_refcnt++;
1548 				fd_close(fd);
1549 			}
1550 		}
1551 		KASSERT(ff->ff_refcnt == 0);
1552 		KASSERT(ff->ff_file == NULL);
1553 		KASSERT(!ff->ff_exclose);
1554 		KASSERT(!ff->ff_allocated);
1555 		if (fd >= NDFDFILE) {
1556 			pool_cache_put(fdfile_cache, ff);
1557 			dt->dt_ff[fd] = NULL;
1558 		}
1559 	}
1560 
1561 	/*
1562 	 * Clean out the descriptor table for the next user and return
1563 	 * to the cache.
1564 	 */
1565 	if (__predict_false(dt != &fdp->fd_dtbuiltin)) {
1566 		fd_dtab_free(fdp->fd_dt);
1567 		/* Otherwise, done above. */
1568 		memset(&fdp->fd_dtbuiltin.dt_ff[NDFDFILE], 0,
1569 		    (NDFILE - NDFDFILE) * sizeof(fdp->fd_dtbuiltin.dt_ff[0]));
1570 		fdp->fd_dt = &fdp->fd_dtbuiltin;
1571 	}
1572 	if (__predict_false(NDHISLOTS(nf) > NDHISLOTS(NDFILE))) {
1573 		KASSERT(fdp->fd_himap != fdp->fd_dhimap);
1574 		KASSERT(fdp->fd_lomap != fdp->fd_dlomap);
1575 		fd_map_free(nf, fdp->fd_lomap, fdp->fd_himap);
1576 	}
1577 	if (__predict_false(fdp->fd_knhash != NULL)) {
1578 		hashdone(fdp->fd_knhash, HASH_LIST, fdp->fd_knhashmask);
1579 		fdp->fd_knhash = NULL;
1580 		fdp->fd_knhashmask = 0;
1581 	} else {
1582 		KASSERT(fdp->fd_knhashmask == 0);
1583 	}
1584 	fdp->fd_dt = &fdp->fd_dtbuiltin;
1585 	fdp->fd_lastkqfile = -1;
1586 	fdp->fd_lastfile = -1;
1587 	fdp->fd_freefile = 0;
1588 	fdp->fd_exclose = false;
1589 	memset(&fdp->fd_startzero, 0, sizeof(*fdp) -
1590 	    offsetof(filedesc_t, fd_startzero));
1591 	fdp->fd_himap = fdp->fd_dhimap;
1592 	fdp->fd_lomap = fdp->fd_dlomap;
1593 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1594 	KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1595 	KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1596 #ifdef DEBUG
1597 	fdp->fd_refcnt = 0; /* see fd_checkmaps */
1598 #endif
1599 	fd_checkmaps(fdp);
1600 	pool_cache_put(filedesc_cache, fdp);
1601 }
1602 
1603 /*
1604  * File Descriptor pseudo-device driver (/dev/fd/).
1605  *
1606  * Opening minor device N dup()s the file (if any) connected to file
1607  * descriptor N belonging to the calling process.  Note that this driver
1608  * consists of only the ``open()'' routine, because all subsequent
1609  * references to this file will be direct to the other driver.
1610  */
1611 static int
1612 filedescopen(dev_t dev, int mode, int type, lwp_t *l)
1613 {
1614 
1615 	/*
1616 	 * XXX Kludge: set dupfd to contain the value of the
1617 	 * the file descriptor being sought for duplication. The error
1618 	 * return ensures that the vnode for this device will be released
1619 	 * by vn_open. Open will detect this special error and take the
1620 	 * actions in fd_dupopen below. Other callers of vn_open or VOP_OPEN
1621 	 * will simply report the error.
1622 	 */
1623 	l->l_dupfd = minor(dev);	/* XXX */
1624 	return EDUPFD;
1625 }
1626 
1627 /*
1628  * Duplicate the specified descriptor to a free descriptor.
1629  */
1630 int
1631 fd_dupopen(int old, int *new, int mode, int error)
1632 {
1633 	filedesc_t *fdp;
1634 	fdfile_t *ff;
1635 	file_t *fp;
1636 	fdtab_t *dt;
1637 
1638 	if ((fp = fd_getfile(old)) == NULL) {
1639 		return EBADF;
1640 	}
1641 	fdp = curlwp->l_fd;
1642 	dt = fdp->fd_dt;
1643 	ff = dt->dt_ff[old];
1644 
1645 	/*
1646 	 * There are two cases of interest here.
1647 	 *
1648 	 * For EDUPFD simply dup (old) to file descriptor
1649 	 * (new) and return.
1650 	 *
1651 	 * For EMOVEFD steal away the file structure from (old) and
1652 	 * store it in (new).  (old) is effectively closed by
1653 	 * this operation.
1654 	 *
1655 	 * Any other error code is just returned.
1656 	 */
1657 	switch (error) {
1658 	case EDUPFD:
1659 		/*
1660 		 * Check that the mode the file is being opened for is a
1661 		 * subset of the mode of the existing descriptor.
1662 		 */
1663 		if (((mode & (FREAD|FWRITE)) | fp->f_flag) != fp->f_flag) {
1664 			error = EACCES;
1665 			break;
1666 		}
1667 
1668 		/* Copy it. */
1669 		error = fd_dup(fp, 0, new, ff->ff_exclose);
1670 		break;
1671 
1672 	case EMOVEFD:
1673 		/* Copy it. */
1674 		error = fd_dup(fp, 0, new, ff->ff_exclose);
1675 		if (error != 0) {
1676 			break;
1677 		}
1678 
1679 		/* Steal away the file pointer from 'old'. */
1680 		(void)fd_close(old);
1681 		return 0;
1682 	}
1683 
1684 	fd_putfile(old);
1685 	return error;
1686 }
1687 
1688 /*
1689  * Close open files on exec.
1690  */
1691 void
1692 fd_closeexec(void)
1693 {
1694 	proc_t *p;
1695 	filedesc_t *fdp;
1696 	fdfile_t *ff;
1697 	lwp_t *l;
1698 	fdtab_t *dt;
1699 	int fd;
1700 
1701 	l = curlwp;
1702 	p = l->l_proc;
1703 	fdp = p->p_fd;
1704 
1705 	if (fdp->fd_refcnt > 1) {
1706 		fdp = fd_copy();
1707 		fd_free();
1708 		p->p_fd = fdp;
1709 		l->l_fd = fdp;
1710 	}
1711 	if (!fdp->fd_exclose) {
1712 		return;
1713 	}
1714 	fdp->fd_exclose = false;
1715 	dt = fdp->fd_dt;
1716 
1717 	for (fd = 0; fd <= fdp->fd_lastfile; fd++) {
1718 		if ((ff = dt->dt_ff[fd]) == NULL) {
1719 			KASSERT(fd >= NDFDFILE);
1720 			continue;
1721 		}
1722 		KASSERT(fd >= NDFDFILE ||
1723 		    ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1724 		if (ff->ff_file == NULL)
1725 			continue;
1726 		if (ff->ff_exclose) {
1727 			/*
1728 			 * We need a reference to close the file.
1729 			 * No other threads can see the fdfile_t at
1730 			 * this point, so don't bother locking.
1731 			 */
1732 			KASSERT((ff->ff_refcnt & FR_CLOSING) == 0);
1733 			ff->ff_refcnt++;
1734 			fd_close(fd);
1735 		}
1736 	}
1737 }
1738 
1739 /*
1740  * Sets descriptor owner. If the owner is a process, 'pgid'
1741  * is set to positive value, process ID. If the owner is process group,
1742  * 'pgid' is set to -pg_id.
1743  */
1744 int
1745 fsetown(pid_t *pgid, u_long cmd, const void *data)
1746 {
1747 	pid_t id = *(const pid_t *)data;
1748 	int error;
1749 
1750 	switch (cmd) {
1751 	case TIOCSPGRP:
1752 		if (id < 0)
1753 			return EINVAL;
1754 		id = -id;
1755 		break;
1756 	default:
1757 		break;
1758 	}
1759 	if (id > 0) {
1760 		mutex_enter(proc_lock);
1761 		error = proc_find(id) ? 0 : ESRCH;
1762 		mutex_exit(proc_lock);
1763 	} else if (id < 0) {
1764 		error = pgid_in_session(curproc, -id);
1765 	} else {
1766 		error = 0;
1767 	}
1768 	if (!error) {
1769 		*pgid = id;
1770 	}
1771 	return error;
1772 }
1773 
1774 void
1775 fd_set_exclose(struct lwp *l, int fd, bool exclose)
1776 {
1777 	filedesc_t *fdp = l->l_fd;
1778 	fdfile_t *ff = fdp->fd_dt->dt_ff[fd];
1779 
1780 	ff->ff_exclose = exclose;
1781 	if (exclose)
1782 		fdp->fd_exclose = true;
1783 }
1784 
1785 /*
1786  * Return descriptor owner information. If the value is positive,
1787  * it's process ID. If it's negative, it's process group ID and
1788  * needs the sign removed before use.
1789  */
1790 int
1791 fgetown(pid_t pgid, u_long cmd, void *data)
1792 {
1793 
1794 	switch (cmd) {
1795 	case TIOCGPGRP:
1796 		*(int *)data = -pgid;
1797 		break;
1798 	default:
1799 		*(int *)data = pgid;
1800 		break;
1801 	}
1802 	return 0;
1803 }
1804 
1805 /*
1806  * Send signal to descriptor owner, either process or process group.
1807  */
1808 void
1809 fownsignal(pid_t pgid, int signo, int code, int band, void *fdescdata)
1810 {
1811 	ksiginfo_t ksi;
1812 
1813 	KASSERT(!cpu_intr_p());
1814 
1815 	if (pgid == 0) {
1816 		return;
1817 	}
1818 
1819 	KSI_INIT(&ksi);
1820 	ksi.ksi_signo = signo;
1821 	ksi.ksi_code = code;
1822 	ksi.ksi_band = band;
1823 
1824 	mutex_enter(proc_lock);
1825 	if (pgid > 0) {
1826 		struct proc *p1;
1827 
1828 		p1 = proc_find(pgid);
1829 		if (p1 != NULL) {
1830 			kpsignal(p1, &ksi, fdescdata);
1831 		}
1832 	} else {
1833 		struct pgrp *pgrp;
1834 
1835 		KASSERT(pgid < 0);
1836 		pgrp = pgrp_find(-pgid);
1837 		if (pgrp != NULL) {
1838 			kpgsignal(pgrp, &ksi, fdescdata, 0);
1839 		}
1840 	}
1841 	mutex_exit(proc_lock);
1842 }
1843 
1844 int
1845 fd_clone(file_t *fp, unsigned fd, int flag, const struct fileops *fops,
1846 	 void *data)
1847 {
1848 
1849 	fp->f_flag = flag;
1850 	fp->f_type = DTYPE_MISC;
1851 	fp->f_ops = fops;
1852 	fp->f_data = data;
1853 	curlwp->l_dupfd = fd;
1854 	fd_affix(curproc, fp, fd);
1855 
1856 	return EMOVEFD;
1857 }
1858 
1859 int
1860 fnullop_fcntl(file_t *fp, u_int cmd, void *data)
1861 {
1862 
1863 	if (cmd == F_SETFL)
1864 		return 0;
1865 
1866 	return EOPNOTSUPP;
1867 }
1868 
1869 int
1870 fnullop_poll(file_t *fp, int which)
1871 {
1872 
1873 	return 0;
1874 }
1875 
1876 int
1877 fnullop_kqfilter(file_t *fp, struct knote *kn)
1878 {
1879 
1880 	return 0;
1881 }
1882 
1883 void
1884 fnullop_restart(file_t *fp)
1885 {
1886 
1887 }
1888 
1889 int
1890 fbadop_read(file_t *fp, off_t *offset, struct uio *uio,
1891 	    kauth_cred_t cred, int flags)
1892 {
1893 
1894 	return EOPNOTSUPP;
1895 }
1896 
1897 int
1898 fbadop_write(file_t *fp, off_t *offset, struct uio *uio,
1899 	     kauth_cred_t cred, int flags)
1900 {
1901 
1902 	return EOPNOTSUPP;
1903 }
1904 
1905 int
1906 fbadop_ioctl(file_t *fp, u_long com, void *data)
1907 {
1908 
1909 	return EOPNOTSUPP;
1910 }
1911 
1912 int
1913 fbadop_stat(file_t *fp, struct stat *sb)
1914 {
1915 
1916 	return EOPNOTSUPP;
1917 }
1918 
1919 int
1920 fbadop_close(file_t *fp)
1921 {
1922 
1923 	return EOPNOTSUPP;
1924 }
1925 
1926 /*
1927  * sysctl routines pertaining to file descriptors
1928  */
1929 
1930 /* Initialized in sysctl_init() for now... */
1931 extern kmutex_t sysctl_file_marker_lock;
1932 static u_int sysctl_file_marker = 1;
1933 
1934 /*
1935  * Expects to be called with proc_lock and sysctl_file_marker_lock locked.
1936  */
1937 static void
1938 sysctl_file_marker_reset(void)
1939 {
1940 	struct proc *p;
1941 
1942 	PROCLIST_FOREACH(p, &allproc) {
1943 		struct filedesc *fd = p->p_fd;
1944 		fdtab_t *dt;
1945 		u_int i;
1946 
1947 		mutex_enter(&fd->fd_lock);
1948 		dt = fd->fd_dt;
1949 		for (i = 0; i < dt->dt_nfiles; i++) {
1950 			struct file *fp;
1951 			fdfile_t *ff;
1952 
1953 			if ((ff = dt->dt_ff[i]) == NULL) {
1954 				continue;
1955 			}
1956 			if ((fp = ff->ff_file) == NULL) {
1957 				continue;
1958 			}
1959 			fp->f_marker = 0;
1960 		}
1961 		mutex_exit(&fd->fd_lock);
1962 	}
1963 }
1964 
1965 /*
1966  * sysctl helper routine for kern.file pseudo-subtree.
1967  */
1968 static int
1969 sysctl_kern_file(SYSCTLFN_ARGS)
1970 {
1971 	int error;
1972 	size_t buflen;
1973 	struct file *fp, fbuf;
1974 	char *start, *where;
1975 	struct proc *p;
1976 
1977 	start = where = oldp;
1978 	buflen = *oldlenp;
1979 
1980 	if (where == NULL) {
1981 		/*
1982 		 * overestimate by 10 files
1983 		 */
1984 		*oldlenp = sizeof(filehead) + (nfiles + 10) *
1985 		    sizeof(struct file);
1986 		return 0;
1987 	}
1988 
1989 	/*
1990 	 * first sysctl_copyout filehead
1991 	 */
1992 	if (buflen < sizeof(filehead)) {
1993 		*oldlenp = 0;
1994 		return 0;
1995 	}
1996 	sysctl_unlock();
1997 	error = sysctl_copyout(l, &filehead, where, sizeof(filehead));
1998 	if (error) {
1999 		sysctl_relock();
2000 		return error;
2001 	}
2002 	buflen -= sizeof(filehead);
2003 	where += sizeof(filehead);
2004 
2005 	/*
2006 	 * followed by an array of file structures
2007 	 */
2008 	mutex_enter(&sysctl_file_marker_lock);
2009 	mutex_enter(proc_lock);
2010 	PROCLIST_FOREACH(p, &allproc) {
2011 		struct filedesc *fd;
2012 		fdtab_t *dt;
2013 		u_int i;
2014 
2015 		if (p->p_stat == SIDL) {
2016 			/* skip embryonic processes */
2017 			continue;
2018 		}
2019 		mutex_enter(p->p_lock);
2020 		error = kauth_authorize_process(l->l_cred,
2021 		    KAUTH_PROCESS_CANSEE, p,
2022 		    KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2023 		    NULL, NULL);
2024 		mutex_exit(p->p_lock);
2025 		if (error != 0) {
2026 			/*
2027 			 * Don't leak kauth retval if we're silently
2028 			 * skipping this entry.
2029 			 */
2030 			error = 0;
2031 			continue;
2032 		}
2033 
2034 		/*
2035 		 * Grab a hold on the process.
2036 		 */
2037 		if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2038 			continue;
2039 		}
2040 		mutex_exit(proc_lock);
2041 
2042 		fd = p->p_fd;
2043 		mutex_enter(&fd->fd_lock);
2044 		dt = fd->fd_dt;
2045 		for (i = 0; i < dt->dt_nfiles; i++) {
2046 			fdfile_t *ff;
2047 
2048 			if ((ff = dt->dt_ff[i]) == NULL) {
2049 				continue;
2050 			}
2051 			if ((fp = ff->ff_file) == NULL) {
2052 				continue;
2053 			}
2054 
2055 			mutex_enter(&fp->f_lock);
2056 
2057 			if ((fp->f_count == 0) ||
2058 			    (fp->f_marker == sysctl_file_marker)) {
2059 				mutex_exit(&fp->f_lock);
2060 				continue;
2061 			}
2062 
2063 			/* Check that we have enough space. */
2064 			if (buflen < sizeof(struct file)) {
2065 				*oldlenp = where - start;
2066 				mutex_exit(&fp->f_lock);
2067 				error = ENOMEM;
2068 				break;
2069 			}
2070 
2071 			memcpy(&fbuf, fp, sizeof(fbuf));
2072 			mutex_exit(&fp->f_lock);
2073 			error = sysctl_copyout(l, &fbuf, where, sizeof(fbuf));
2074 			if (error) {
2075 				break;
2076 			}
2077 			buflen -= sizeof(struct file);
2078 			where += sizeof(struct file);
2079 
2080 			fp->f_marker = sysctl_file_marker;
2081 		}
2082 		mutex_exit(&fd->fd_lock);
2083 
2084 		/*
2085 		 * Release reference to process.
2086 		 */
2087 		mutex_enter(proc_lock);
2088 		rw_exit(&p->p_reflock);
2089 
2090 		if (error)
2091 			break;
2092 	}
2093 
2094 	sysctl_file_marker++;
2095 	/* Reset all markers if wrapped. */
2096 	if (sysctl_file_marker == 0) {
2097 		sysctl_file_marker_reset();
2098 		sysctl_file_marker++;
2099 	}
2100 
2101 	mutex_exit(proc_lock);
2102 	mutex_exit(&sysctl_file_marker_lock);
2103 
2104 	*oldlenp = where - start;
2105 	sysctl_relock();
2106 	return error;
2107 }
2108 
2109 /*
2110  * sysctl helper function for kern.file2
2111  */
2112 static int
2113 sysctl_kern_file2(SYSCTLFN_ARGS)
2114 {
2115 	struct proc *p;
2116 	struct file *fp;
2117 	struct filedesc *fd;
2118 	struct kinfo_file kf;
2119 	char *dp;
2120 	u_int i, op;
2121 	size_t len, needed, elem_size, out_size;
2122 	int error, arg, elem_count;
2123 	fdfile_t *ff;
2124 	fdtab_t *dt;
2125 
2126 	if (namelen == 1 && name[0] == CTL_QUERY)
2127 		return sysctl_query(SYSCTLFN_CALL(rnode));
2128 
2129 	if (namelen != 4)
2130 		return EINVAL;
2131 
2132 	error = 0;
2133 	dp = oldp;
2134 	len = (oldp != NULL) ? *oldlenp : 0;
2135 	op = name[0];
2136 	arg = name[1];
2137 	elem_size = name[2];
2138 	elem_count = name[3];
2139 	out_size = MIN(sizeof(kf), elem_size);
2140 	needed = 0;
2141 
2142 	if (elem_size < 1 || elem_count < 0)
2143 		return EINVAL;
2144 
2145 	switch (op) {
2146 	case KERN_FILE_BYFILE:
2147 	case KERN_FILE_BYPID:
2148 		/*
2149 		 * We're traversing the process list in both cases; the BYFILE
2150 		 * case does additional work of keeping track of files already
2151 		 * looked at.
2152 		 */
2153 
2154 		/* doesn't use arg so it must be zero */
2155 		if ((op == KERN_FILE_BYFILE) && (arg != 0))
2156 			return EINVAL;
2157 
2158 		if ((op == KERN_FILE_BYPID) && (arg < -1))
2159 			/* -1 means all processes */
2160 			return EINVAL;
2161 
2162 		sysctl_unlock();
2163 		if (op == KERN_FILE_BYFILE)
2164 			mutex_enter(&sysctl_file_marker_lock);
2165 		mutex_enter(proc_lock);
2166 		PROCLIST_FOREACH(p, &allproc) {
2167 			if (p->p_stat == SIDL) {
2168 				/* skip embryonic processes */
2169 				continue;
2170 			}
2171 			if (arg > 0 && p->p_pid != arg) {
2172 				/* pick only the one we want */
2173 				/* XXX want 0 to mean "kernel files" */
2174 				continue;
2175 			}
2176 			mutex_enter(p->p_lock);
2177 			error = kauth_authorize_process(l->l_cred,
2178 			    KAUTH_PROCESS_CANSEE, p,
2179 			    KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2180 			    NULL, NULL);
2181 			mutex_exit(p->p_lock);
2182 			if (error != 0) {
2183 				/*
2184 				 * Don't leak kauth retval if we're silently
2185 				 * skipping this entry.
2186 				 */
2187 				error = 0;
2188 				continue;
2189 			}
2190 
2191 			/*
2192 			 * Grab a hold on the process.
2193 			 */
2194 			if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2195 				continue;
2196 			}
2197 			mutex_exit(proc_lock);
2198 
2199 			fd = p->p_fd;
2200 			mutex_enter(&fd->fd_lock);
2201 			dt = fd->fd_dt;
2202 			for (i = 0; i < dt->dt_nfiles; i++) {
2203 				if ((ff = dt->dt_ff[i]) == NULL) {
2204 					continue;
2205 				}
2206 				if ((fp = ff->ff_file) == NULL) {
2207 					continue;
2208 				}
2209 
2210 				if ((op == KERN_FILE_BYFILE) &&
2211 				    (fp->f_marker == sysctl_file_marker)) {
2212 					continue;
2213 				}
2214 				if (len >= elem_size && elem_count > 0) {
2215 					mutex_enter(&fp->f_lock);
2216 					fill_file(&kf, fp, ff, i, p->p_pid);
2217 					mutex_exit(&fp->f_lock);
2218 					mutex_exit(&fd->fd_lock);
2219 					error = sysctl_copyout(l,
2220 					    &kf, dp, out_size);
2221 					mutex_enter(&fd->fd_lock);
2222 					if (error)
2223 						break;
2224 					dp += elem_size;
2225 					len -= elem_size;
2226 				}
2227 				if (op == KERN_FILE_BYFILE)
2228 					fp->f_marker = sysctl_file_marker;
2229 				needed += elem_size;
2230 				if (elem_count > 0 && elem_count != INT_MAX)
2231 					elem_count--;
2232 			}
2233 			mutex_exit(&fd->fd_lock);
2234 
2235 			/*
2236 			 * Release reference to process.
2237 			 */
2238 			mutex_enter(proc_lock);
2239 			rw_exit(&p->p_reflock);
2240 		}
2241 		if (op == KERN_FILE_BYFILE) {
2242 			sysctl_file_marker++;
2243 
2244 			/* Reset all markers if wrapped. */
2245 			if (sysctl_file_marker == 0) {
2246 				sysctl_file_marker_reset();
2247 				sysctl_file_marker++;
2248 			}
2249 		}
2250 		mutex_exit(proc_lock);
2251 		if (op == KERN_FILE_BYFILE)
2252 			mutex_exit(&sysctl_file_marker_lock);
2253 		sysctl_relock();
2254 		break;
2255 	default:
2256 		return EINVAL;
2257 	}
2258 
2259 	if (oldp == NULL)
2260 		needed += KERN_FILESLOP * elem_size;
2261 	*oldlenp = needed;
2262 
2263 	return error;
2264 }
2265 
2266 static void
2267 fill_file(struct kinfo_file *kp, const file_t *fp, const fdfile_t *ff,
2268 	  int i, pid_t pid)
2269 {
2270 
2271 	memset(kp, 0, sizeof(*kp));
2272 
2273 	kp->ki_fileaddr =	PTRTOUINT64(fp);
2274 	kp->ki_flag =		fp->f_flag;
2275 	kp->ki_iflags =		0;
2276 	kp->ki_ftype =		fp->f_type;
2277 	kp->ki_count =		fp->f_count;
2278 	kp->ki_msgcount =	fp->f_msgcount;
2279 	kp->ki_fucred =		PTRTOUINT64(fp->f_cred);
2280 	kp->ki_fuid =		kauth_cred_geteuid(fp->f_cred);
2281 	kp->ki_fgid =		kauth_cred_getegid(fp->f_cred);
2282 	kp->ki_fops =		PTRTOUINT64(fp->f_ops);
2283 	kp->ki_foffset =	fp->f_offset;
2284 	kp->ki_fdata =		PTRTOUINT64(fp->f_data);
2285 
2286 	/* vnode information to glue this file to something */
2287 	if (fp->f_type == DTYPE_VNODE) {
2288 		struct vnode *vp = (struct vnode *)fp->f_data;
2289 
2290 		kp->ki_vun =	PTRTOUINT64(vp->v_un.vu_socket);
2291 		kp->ki_vsize =	vp->v_size;
2292 		kp->ki_vtype =	vp->v_type;
2293 		kp->ki_vtag =	vp->v_tag;
2294 		kp->ki_vdata =	PTRTOUINT64(vp->v_data);
2295 	}
2296 
2297 	/* process information when retrieved via KERN_FILE_BYPID */
2298 	if (ff != NULL) {
2299 		kp->ki_pid =		pid;
2300 		kp->ki_fd =		i;
2301 		kp->ki_ofileflags =	ff->ff_exclose;
2302 		kp->ki_usecount =	ff->ff_refcnt;
2303 	}
2304 }
2305