xref: /netbsd-src/sys/kern/kern_descrip.c (revision 6d322f2f4598f0d8a138f10ea648ec4fabe41f8b)
1 /*	$NetBSD: kern_descrip.c,v 1.222 2013/09/15 13:03:59 martin 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 1.222 2013/09/15 13:03:59 martin 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 = true;
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 = false;
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_getsock1(unsigned fd, struct socket **sop, file_t **fp)
504 {
505 	*fp = fd_getfile(fd);
506 	if (__predict_false(*fp == NULL)) {
507 		return EBADF;
508 	}
509 	if (__predict_false((*fp)->f_type != DTYPE_SOCKET)) {
510 		fd_putfile(fd);
511 		return ENOTSOCK;
512 	}
513 	*sop = (*fp)->f_data;
514 	return 0;
515 }
516 
517 int
518 fd_getsock(unsigned fd, struct socket **sop)
519 {
520 	file_t *fp;
521 	return fd_getsock1(fd, sop, &fp);
522 }
523 
524 /*
525  * Look up the file structure corresponding to a file descriptor
526  * and return it with a reference held on the file, not the
527  * descriptor.
528  *
529  * This is heavyweight and only used when accessing descriptors
530  * from a foreign process.  The caller must ensure that `p' does
531  * not exit or fork across this call.
532  *
533  * To release the file (not descriptor) reference, use closef().
534  */
535 file_t *
536 fd_getfile2(proc_t *p, unsigned fd)
537 {
538 	filedesc_t *fdp;
539 	fdfile_t *ff;
540 	file_t *fp;
541 	fdtab_t *dt;
542 
543 	fdp = p->p_fd;
544 	mutex_enter(&fdp->fd_lock);
545 	dt = fdp->fd_dt;
546 	if (fd >= dt->dt_nfiles) {
547 		mutex_exit(&fdp->fd_lock);
548 		return NULL;
549 	}
550 	if ((ff = dt->dt_ff[fd]) == NULL) {
551 		mutex_exit(&fdp->fd_lock);
552 		return NULL;
553 	}
554 	if ((fp = ff->ff_file) == NULL) {
555 		mutex_exit(&fdp->fd_lock);
556 		return NULL;
557 	}
558 	mutex_enter(&fp->f_lock);
559 	fp->f_count++;
560 	mutex_exit(&fp->f_lock);
561 	mutex_exit(&fdp->fd_lock);
562 
563 	return fp;
564 }
565 
566 /*
567  * Internal form of close.  Must be called with a reference to the
568  * descriptor, and will drop the reference.  When all descriptor
569  * references are dropped, releases the descriptor slot and a single
570  * reference to the file structure.
571  */
572 int
573 fd_close(unsigned fd)
574 {
575 	struct flock lf;
576 	filedesc_t *fdp;
577 	fdfile_t *ff;
578 	file_t *fp;
579 	proc_t *p;
580 	lwp_t *l;
581 	u_int refcnt;
582 
583 	l = curlwp;
584 	p = l->l_proc;
585 	fdp = l->l_fd;
586 	ff = fdp->fd_dt->dt_ff[fd];
587 
588 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
589 
590 	mutex_enter(&fdp->fd_lock);
591 	KASSERT((ff->ff_refcnt & FR_MASK) > 0);
592 	if (__predict_false(ff->ff_file == NULL)) {
593 		/*
594 		 * Another user of the file is already closing, and is
595 		 * waiting for other users of the file to drain.  Release
596 		 * our reference, and wake up the closer.
597 		 */
598 		atomic_dec_uint(&ff->ff_refcnt);
599 		cv_broadcast(&ff->ff_closing);
600 		mutex_exit(&fdp->fd_lock);
601 
602 		/*
603 		 * An application error, so pretend that the descriptor
604 		 * was already closed.  We can't safely wait for it to
605 		 * be closed without potentially deadlocking.
606 		 */
607 		return (EBADF);
608 	}
609 	KASSERT((ff->ff_refcnt & FR_CLOSING) == 0);
610 
611 	/*
612 	 * There may be multiple users of this file within the process.
613 	 * Notify existing and new users that the file is closing.  This
614 	 * will prevent them from adding additional uses to this file
615 	 * while we are closing it.
616 	 */
617 	fp = ff->ff_file;
618 	ff->ff_file = NULL;
619 	ff->ff_exclose = false;
620 
621 	/*
622 	 * We expect the caller to hold a descriptor reference - drop it.
623 	 * The reference count may increase beyond zero at this point due
624 	 * to an erroneous descriptor reference by an application, but
625 	 * fd_getfile() will notice that the file is being closed and drop
626 	 * the reference again.
627 	 */
628 	if (fdp->fd_refcnt == 1) {
629 		/* Single threaded. */
630 		refcnt = --(ff->ff_refcnt);
631 	} else {
632 		/* Multi threaded. */
633 #ifndef __HAVE_ATOMIC_AS_MEMBAR
634 		membar_producer();
635 #endif
636 		refcnt = atomic_dec_uint_nv(&ff->ff_refcnt);
637 	}
638 	if (__predict_false(refcnt != 0)) {
639 		/*
640 		 * Wait for other references to drain.  This is typically
641 		 * an application error - the descriptor is being closed
642 		 * while still in use.
643 		 * (Or just a threaded application trying to unblock its
644 		 * thread that sleeps in (say) accept()).
645 		 */
646 		atomic_or_uint(&ff->ff_refcnt, FR_CLOSING);
647 
648 		/*
649 		 * Remove any knotes attached to the file.  A knote
650 		 * attached to the descriptor can hold references on it.
651 		 */
652 		mutex_exit(&fdp->fd_lock);
653 		if (!SLIST_EMPTY(&ff->ff_knlist)) {
654 			knote_fdclose(fd);
655 		}
656 
657 		/*
658 		 * Since the file system code doesn't know which fd
659 		 * each request came from (think dup()), we have to
660 		 * ask it to return ERESTART for any long-term blocks.
661 		 * The re-entry through read/write/etc will detect the
662 		 * closed fd and return EBAFD.
663 		 * Blocked partial writes may return a short length.
664 		 */
665 		(*fp->f_ops->fo_restart)(fp);
666 		mutex_enter(&fdp->fd_lock);
667 
668 		/*
669 		 * We need to see the count drop to zero at least once,
670 		 * in order to ensure that all pre-existing references
671 		 * have been drained.  New references past this point are
672 		 * of no interest.
673 		 * XXX (dsl) this may need to call fo_restart() after a
674 		 * timeout to guarantee that all the system calls exit.
675 		 */
676 		while ((ff->ff_refcnt & FR_MASK) != 0) {
677 			cv_wait(&ff->ff_closing, &fdp->fd_lock);
678 		}
679 		atomic_and_uint(&ff->ff_refcnt, ~FR_CLOSING);
680 	} else {
681 		/* If no references, there must be no knotes. */
682 		KASSERT(SLIST_EMPTY(&ff->ff_knlist));
683 	}
684 
685 	/*
686 	 * POSIX record locking dictates that any close releases ALL
687 	 * locks owned by this process.  This is handled by setting
688 	 * a flag in the unlock to free ONLY locks obeying POSIX
689 	 * semantics, and not to free BSD-style file locks.
690 	 * If the descriptor was in a message, POSIX-style locks
691 	 * aren't passed with the descriptor.
692 	 */
693 	if (__predict_false((p->p_flag & PK_ADVLOCK) != 0 &&
694 	    fp->f_type == DTYPE_VNODE)) {
695 		lf.l_whence = SEEK_SET;
696 		lf.l_start = 0;
697 		lf.l_len = 0;
698 		lf.l_type = F_UNLCK;
699 		mutex_exit(&fdp->fd_lock);
700 		(void)VOP_ADVLOCK(fp->f_data, p, F_UNLCK, &lf, F_POSIX);
701 		mutex_enter(&fdp->fd_lock);
702 	}
703 
704 	/* Free descriptor slot. */
705 	fd_unused(fdp, fd);
706 	mutex_exit(&fdp->fd_lock);
707 
708 	/* Now drop reference to the file itself. */
709 	return closef(fp);
710 }
711 
712 /*
713  * Duplicate a file descriptor.
714  */
715 int
716 fd_dup(file_t *fp, int minfd, int *newp, bool exclose)
717 {
718 	proc_t *p = curproc;
719 	int error;
720 
721 	while ((error = fd_alloc(p, minfd, newp)) != 0) {
722 		if (error != ENOSPC) {
723 			return error;
724 		}
725 		fd_tryexpand(p);
726 	}
727 
728 	curlwp->l_fd->fd_dt->dt_ff[*newp]->ff_exclose = exclose;
729 	fd_affix(p, fp, *newp);
730 	return 0;
731 }
732 
733 /*
734  * dup2 operation.
735  */
736 int
737 fd_dup2(file_t *fp, unsigned new, int flags)
738 {
739 	filedesc_t *fdp = curlwp->l_fd;
740 	fdfile_t *ff;
741 	fdtab_t *dt;
742 
743 	if (flags & ~(O_CLOEXEC|O_NONBLOCK))
744 		return EINVAL;
745 	/*
746 	 * Ensure there are enough slots in the descriptor table,
747 	 * and allocate an fdfile_t up front in case we need it.
748 	 */
749 	while (new >= fdp->fd_dt->dt_nfiles) {
750 		fd_tryexpand(curproc);
751 	}
752 	ff = pool_cache_get(fdfile_cache, PR_WAITOK);
753 
754 	/*
755 	 * If there is already a file open, close it.  If the file is
756 	 * half open, wait for it to be constructed before closing it.
757 	 * XXX Potential for deadlock here?
758 	 */
759 	mutex_enter(&fdp->fd_lock);
760 	while (fd_isused(fdp, new)) {
761 		mutex_exit(&fdp->fd_lock);
762 		if (fd_getfile(new) != NULL) {
763 			(void)fd_close(new);
764 		} else {
765 			/*
766 			 * Crummy, but unlikely to happen.
767 			 * Can occur if we interrupt another
768 			 * thread while it is opening a file.
769 			 */
770 			kpause("dup2", false, 1, NULL);
771 		}
772 		mutex_enter(&fdp->fd_lock);
773 	}
774 	dt = fdp->fd_dt;
775 	if (dt->dt_ff[new] == NULL) {
776 		KASSERT(new >= NDFDFILE);
777 		dt->dt_ff[new] = ff;
778 		ff = NULL;
779 	}
780 	fd_used(fdp, new);
781 	mutex_exit(&fdp->fd_lock);
782 
783 	dt->dt_ff[new]->ff_exclose = (flags & O_CLOEXEC) != 0;
784 	fp->f_flag |= flags & FNONBLOCK;
785 	/* Slot is now allocated.  Insert copy of the file. */
786 	fd_affix(curproc, fp, new);
787 	if (ff != NULL) {
788 		pool_cache_put(fdfile_cache, ff);
789 	}
790 	return 0;
791 }
792 
793 /*
794  * Drop reference to a file structure.
795  */
796 int
797 closef(file_t *fp)
798 {
799 	struct flock lf;
800 	int error;
801 
802 	/*
803 	 * Drop reference.  If referenced elsewhere it's still open
804 	 * and we have nothing more to do.
805 	 */
806 	mutex_enter(&fp->f_lock);
807 	KASSERT(fp->f_count > 0);
808 	if (--fp->f_count > 0) {
809 		mutex_exit(&fp->f_lock);
810 		return 0;
811 	}
812 	KASSERT(fp->f_count == 0);
813 	mutex_exit(&fp->f_lock);
814 
815 	/* We held the last reference - release locks, close and free. */
816 	if ((fp->f_flag & FHASLOCK) && fp->f_type == DTYPE_VNODE) {
817 		lf.l_whence = SEEK_SET;
818 		lf.l_start = 0;
819 		lf.l_len = 0;
820 		lf.l_type = F_UNLCK;
821 		(void)VOP_ADVLOCK(fp->f_data, fp, F_UNLCK, &lf, F_FLOCK);
822 	}
823 	if (fp->f_ops != NULL) {
824 		error = (*fp->f_ops->fo_close)(fp);
825 	} else {
826 		error = 0;
827 	}
828 	KASSERT(fp->f_count == 0);
829 	KASSERT(fp->f_cred != NULL);
830 	pool_cache_put(file_cache, fp);
831 
832 	return error;
833 }
834 
835 /*
836  * Allocate a file descriptor for the process.
837  */
838 int
839 fd_alloc(proc_t *p, int want, int *result)
840 {
841 	filedesc_t *fdp = p->p_fd;
842 	int i, lim, last, error;
843 	u_int off, new;
844 	fdtab_t *dt;
845 
846 	KASSERT(p == curproc || p == &proc0);
847 
848 	/*
849 	 * Search for a free descriptor starting at the higher
850 	 * of want or fd_freefile.
851 	 */
852 	mutex_enter(&fdp->fd_lock);
853 	fd_checkmaps(fdp);
854 	dt = fdp->fd_dt;
855 	KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
856 	lim = min((int)p->p_rlimit[RLIMIT_NOFILE].rlim_cur, maxfiles);
857 	last = min(dt->dt_nfiles, lim);
858 	for (;;) {
859 		if ((i = want) < fdp->fd_freefile)
860 			i = fdp->fd_freefile;
861 		off = i >> NDENTRYSHIFT;
862 		new = fd_next_zero(fdp, fdp->fd_himap, off,
863 		    (last + NDENTRIES - 1) >> NDENTRYSHIFT);
864 		if (new == -1)
865 			break;
866 		i = fd_next_zero(fdp, &fdp->fd_lomap[new],
867 		    new > off ? 0 : i & NDENTRYMASK, NDENTRIES);
868 		if (i == -1) {
869 			/*
870 			 * Free file descriptor in this block was
871 			 * below want, try again with higher want.
872 			 */
873 			want = (new + 1) << NDENTRYSHIFT;
874 			continue;
875 		}
876 		i += (new << NDENTRYSHIFT);
877 		if (i >= last) {
878 			break;
879 		}
880 		if (dt->dt_ff[i] == NULL) {
881 			KASSERT(i >= NDFDFILE);
882 			dt->dt_ff[i] = pool_cache_get(fdfile_cache, PR_WAITOK);
883 		}
884 		KASSERT(dt->dt_ff[i]->ff_file == NULL);
885 		fd_used(fdp, i);
886 		if (want <= fdp->fd_freefile) {
887 			fdp->fd_freefile = i;
888 		}
889 		*result = i;
890 		KASSERT(i >= NDFDFILE ||
891 		    dt->dt_ff[i] == (fdfile_t *)fdp->fd_dfdfile[i]);
892 		fd_checkmaps(fdp);
893 		mutex_exit(&fdp->fd_lock);
894 		return 0;
895 	}
896 
897 	/* No space in current array.  Let the caller expand and retry. */
898 	error = (dt->dt_nfiles >= lim) ? EMFILE : ENOSPC;
899 	mutex_exit(&fdp->fd_lock);
900 	return error;
901 }
902 
903 /*
904  * Allocate memory for a descriptor table.
905  */
906 static fdtab_t *
907 fd_dtab_alloc(int n)
908 {
909 	fdtab_t *dt;
910 	size_t sz;
911 
912 	KASSERT(n > NDFILE);
913 
914 	sz = sizeof(*dt) + (n - NDFILE) * sizeof(dt->dt_ff[0]);
915 	dt = kmem_alloc(sz, KM_SLEEP);
916 #ifdef DIAGNOSTIC
917 	memset(dt, 0xff, sz);
918 #endif
919 	dt->dt_nfiles = n;
920 	dt->dt_link = NULL;
921 	return dt;
922 }
923 
924 /*
925  * Free a descriptor table, and all tables linked for deferred free.
926  */
927 static void
928 fd_dtab_free(fdtab_t *dt)
929 {
930 	fdtab_t *next;
931 	size_t sz;
932 
933 	do {
934 		next = dt->dt_link;
935 		KASSERT(dt->dt_nfiles > NDFILE);
936 		sz = sizeof(*dt) +
937 		    (dt->dt_nfiles - NDFILE) * sizeof(dt->dt_ff[0]);
938 #ifdef DIAGNOSTIC
939 		memset(dt, 0xff, sz);
940 #endif
941 		kmem_free(dt, sz);
942 		dt = next;
943 	} while (dt != NULL);
944 }
945 
946 /*
947  * Allocate descriptor bitmap.
948  */
949 static void
950 fd_map_alloc(int n, uint32_t **lo, uint32_t **hi)
951 {
952 	uint8_t *ptr;
953 	size_t szlo, szhi;
954 
955 	KASSERT(n > NDENTRIES);
956 
957 	szlo = NDLOSLOTS(n) * sizeof(uint32_t);
958 	szhi = NDHISLOTS(n) * sizeof(uint32_t);
959 	ptr = kmem_alloc(szlo + szhi, KM_SLEEP);
960 	*lo = (uint32_t *)ptr;
961 	*hi = (uint32_t *)(ptr + szlo);
962 }
963 
964 /*
965  * Free descriptor bitmap.
966  */
967 static void
968 fd_map_free(int n, uint32_t *lo, uint32_t *hi)
969 {
970 	size_t szlo, szhi;
971 
972 	KASSERT(n > NDENTRIES);
973 
974 	szlo = NDLOSLOTS(n) * sizeof(uint32_t);
975 	szhi = NDHISLOTS(n) * sizeof(uint32_t);
976 	KASSERT(hi == (uint32_t *)((uint8_t *)lo + szlo));
977 	kmem_free(lo, szlo + szhi);
978 }
979 
980 /*
981  * Expand a process' descriptor table.
982  */
983 void
984 fd_tryexpand(proc_t *p)
985 {
986 	filedesc_t *fdp;
987 	int i, numfiles, oldnfiles;
988 	fdtab_t *newdt, *dt;
989 	uint32_t *newhimap, *newlomap;
990 
991 	KASSERT(p == curproc || p == &proc0);
992 
993 	fdp = p->p_fd;
994 	newhimap = NULL;
995 	newlomap = NULL;
996 	oldnfiles = fdp->fd_dt->dt_nfiles;
997 
998 	if (oldnfiles < NDEXTENT)
999 		numfiles = NDEXTENT;
1000 	else
1001 		numfiles = 2 * oldnfiles;
1002 
1003 	newdt = fd_dtab_alloc(numfiles);
1004 	if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1005 		fd_map_alloc(numfiles, &newlomap, &newhimap);
1006 	}
1007 
1008 	mutex_enter(&fdp->fd_lock);
1009 	dt = fdp->fd_dt;
1010 	KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1011 	if (dt->dt_nfiles != oldnfiles) {
1012 		/* fdp changed; caller must retry */
1013 		mutex_exit(&fdp->fd_lock);
1014 		fd_dtab_free(newdt);
1015 		if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1016 			fd_map_free(numfiles, newlomap, newhimap);
1017 		}
1018 		return;
1019 	}
1020 
1021 	/* Copy the existing descriptor table and zero the new portion. */
1022 	i = sizeof(fdfile_t *) * oldnfiles;
1023 	memcpy(newdt->dt_ff, dt->dt_ff, i);
1024 	memset((uint8_t *)newdt->dt_ff + i, 0,
1025 	    numfiles * sizeof(fdfile_t *) - i);
1026 
1027 	/*
1028 	 * Link old descriptor array into list to be discarded.  We defer
1029 	 * freeing until the last reference to the descriptor table goes
1030 	 * away (usually process exit).  This allows us to do lockless
1031 	 * lookups in fd_getfile().
1032 	 */
1033 	if (oldnfiles > NDFILE) {
1034 		if (fdp->fd_refcnt > 1) {
1035 			newdt->dt_link = dt;
1036 		} else {
1037 			fd_dtab_free(dt);
1038 		}
1039 	}
1040 
1041 	if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1042 		i = NDHISLOTS(oldnfiles) * sizeof(uint32_t);
1043 		memcpy(newhimap, fdp->fd_himap, i);
1044 		memset((uint8_t *)newhimap + i, 0,
1045 		    NDHISLOTS(numfiles) * sizeof(uint32_t) - i);
1046 
1047 		i = NDLOSLOTS(oldnfiles) * sizeof(uint32_t);
1048 		memcpy(newlomap, fdp->fd_lomap, i);
1049 		memset((uint8_t *)newlomap + i, 0,
1050 		    NDLOSLOTS(numfiles) * sizeof(uint32_t) - i);
1051 
1052 		if (NDHISLOTS(oldnfiles) > NDHISLOTS(NDFILE)) {
1053 			fd_map_free(oldnfiles, fdp->fd_lomap, fdp->fd_himap);
1054 		}
1055 		fdp->fd_himap = newhimap;
1056 		fdp->fd_lomap = newlomap;
1057 	}
1058 
1059 	/*
1060 	 * All other modifications must become globally visible before
1061 	 * the change to fd_dt.  See fd_getfile().
1062 	 */
1063 	membar_producer();
1064 	fdp->fd_dt = newdt;
1065 	KASSERT(newdt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1066 	fd_checkmaps(fdp);
1067 	mutex_exit(&fdp->fd_lock);
1068 }
1069 
1070 /*
1071  * Create a new open file structure and allocate a file descriptor
1072  * for the current process.
1073  */
1074 int
1075 fd_allocfile(file_t **resultfp, int *resultfd)
1076 {
1077 	proc_t *p = curproc;
1078 	kauth_cred_t cred;
1079 	file_t *fp;
1080 	int error;
1081 
1082 	while ((error = fd_alloc(p, 0, resultfd)) != 0) {
1083 		if (error != ENOSPC) {
1084 			return error;
1085 		}
1086 		fd_tryexpand(p);
1087 	}
1088 
1089 	fp = pool_cache_get(file_cache, PR_WAITOK);
1090 	if (fp == NULL) {
1091 		fd_abort(p, NULL, *resultfd);
1092 		return ENFILE;
1093 	}
1094 	KASSERT(fp->f_count == 0);
1095 	KASSERT(fp->f_msgcount == 0);
1096 	KASSERT(fp->f_unpcount == 0);
1097 
1098 	/* Replace cached credentials if not what we need. */
1099 	cred = curlwp->l_cred;
1100 	if (__predict_false(cred != fp->f_cred)) {
1101 		kauth_cred_free(fp->f_cred);
1102 		kauth_cred_hold(cred);
1103 		fp->f_cred = cred;
1104 	}
1105 
1106 	/*
1107 	 * Don't allow recycled files to be scanned.
1108 	 * See uipc_usrreq.c.
1109 	 */
1110 	if (__predict_false((fp->f_flag & FSCAN) != 0)) {
1111 		mutex_enter(&fp->f_lock);
1112 		atomic_and_uint(&fp->f_flag, ~FSCAN);
1113 		mutex_exit(&fp->f_lock);
1114 	}
1115 
1116 	fp->f_advice = 0;
1117 	fp->f_offset = 0;
1118 	*resultfp = fp;
1119 
1120 	return 0;
1121 }
1122 
1123 /*
1124  * Successful creation of a new descriptor: make visible to the process.
1125  */
1126 void
1127 fd_affix(proc_t *p, file_t *fp, unsigned fd)
1128 {
1129 	fdfile_t *ff;
1130 	filedesc_t *fdp;
1131 
1132 	KASSERT(p == curproc || p == &proc0);
1133 
1134 	/* Add a reference to the file structure. */
1135 	mutex_enter(&fp->f_lock);
1136 	fp->f_count++;
1137 	mutex_exit(&fp->f_lock);
1138 
1139 	/*
1140 	 * Insert the new file into the descriptor slot.
1141 	 *
1142 	 * The memory barriers provided by lock activity in this routine
1143 	 * ensure that any updates to the file structure become globally
1144 	 * visible before the file becomes visible to other LWPs in the
1145 	 * current process.
1146 	 */
1147 	fdp = p->p_fd;
1148 	ff = fdp->fd_dt->dt_ff[fd];
1149 
1150 	KASSERT(ff != NULL);
1151 	KASSERT(ff->ff_file == NULL);
1152 	KASSERT(ff->ff_allocated);
1153 	KASSERT(fd_isused(fdp, fd));
1154 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1155 
1156 	/* No need to lock in order to make file initially visible. */
1157 	ff->ff_file = fp;
1158 }
1159 
1160 /*
1161  * Abort creation of a new descriptor: free descriptor slot and file.
1162  */
1163 void
1164 fd_abort(proc_t *p, file_t *fp, unsigned fd)
1165 {
1166 	filedesc_t *fdp;
1167 	fdfile_t *ff;
1168 
1169 	KASSERT(p == curproc || p == &proc0);
1170 
1171 	fdp = p->p_fd;
1172 	ff = fdp->fd_dt->dt_ff[fd];
1173 	ff->ff_exclose = false;
1174 
1175 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1176 
1177 	mutex_enter(&fdp->fd_lock);
1178 	KASSERT(fd_isused(fdp, fd));
1179 	fd_unused(fdp, fd);
1180 	mutex_exit(&fdp->fd_lock);
1181 
1182 	if (fp != NULL) {
1183 		KASSERT(fp->f_count == 0);
1184 		KASSERT(fp->f_cred != NULL);
1185 		pool_cache_put(file_cache, fp);
1186 	}
1187 }
1188 
1189 static int
1190 file_ctor(void *arg, void *obj, int flags)
1191 {
1192 	file_t *fp = obj;
1193 
1194 	memset(fp, 0, sizeof(*fp));
1195 
1196 	mutex_enter(&filelist_lock);
1197 	if (__predict_false(nfiles >= maxfiles)) {
1198 		mutex_exit(&filelist_lock);
1199 		tablefull("file", "increase kern.maxfiles or MAXFILES");
1200 		return ENFILE;
1201 	}
1202 	nfiles++;
1203 	LIST_INSERT_HEAD(&filehead, fp, f_list);
1204 	mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1205 	fp->f_cred = curlwp->l_cred;
1206 	kauth_cred_hold(fp->f_cred);
1207 	mutex_exit(&filelist_lock);
1208 
1209 	return 0;
1210 }
1211 
1212 static void
1213 file_dtor(void *arg, void *obj)
1214 {
1215 	file_t *fp = obj;
1216 
1217 	mutex_enter(&filelist_lock);
1218 	nfiles--;
1219 	LIST_REMOVE(fp, f_list);
1220 	mutex_exit(&filelist_lock);
1221 
1222 	kauth_cred_free(fp->f_cred);
1223 	mutex_destroy(&fp->f_lock);
1224 }
1225 
1226 static int
1227 fdfile_ctor(void *arg, void *obj, int flags)
1228 {
1229 	fdfile_t *ff = obj;
1230 
1231 	memset(ff, 0, sizeof(*ff));
1232 	cv_init(&ff->ff_closing, "fdclose");
1233 
1234 	return 0;
1235 }
1236 
1237 static void
1238 fdfile_dtor(void *arg, void *obj)
1239 {
1240 	fdfile_t *ff = obj;
1241 
1242 	cv_destroy(&ff->ff_closing);
1243 }
1244 
1245 file_t *
1246 fgetdummy(void)
1247 {
1248 	file_t *fp;
1249 
1250 	fp = kmem_zalloc(sizeof(*fp), KM_SLEEP);
1251 	if (fp != NULL) {
1252 		mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1253 	}
1254 	return fp;
1255 }
1256 
1257 void
1258 fputdummy(file_t *fp)
1259 {
1260 
1261 	mutex_destroy(&fp->f_lock);
1262 	kmem_free(fp, sizeof(*fp));
1263 }
1264 
1265 /*
1266  * Create an initial filedesc structure.
1267  */
1268 filedesc_t *
1269 fd_init(filedesc_t *fdp)
1270 {
1271 #ifdef DIAGNOSTIC
1272 	unsigned fd;
1273 #endif
1274 
1275 	if (__predict_true(fdp == NULL)) {
1276 		fdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1277 	} else {
1278 		KASSERT(fdp == &filedesc0);
1279 		filedesc_ctor(NULL, fdp, PR_WAITOK);
1280 	}
1281 
1282 #ifdef DIAGNOSTIC
1283 	KASSERT(fdp->fd_lastfile == -1);
1284 	KASSERT(fdp->fd_lastkqfile == -1);
1285 	KASSERT(fdp->fd_knhash == NULL);
1286 	KASSERT(fdp->fd_freefile == 0);
1287 	KASSERT(fdp->fd_exclose == false);
1288 	KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1289 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1290 	for (fd = 0; fd < NDFDFILE; fd++) {
1291 		KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] ==
1292 		    (fdfile_t *)fdp->fd_dfdfile[fd]);
1293 	}
1294 	for (fd = NDFDFILE; fd < NDFILE; fd++) {
1295 		KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] == NULL);
1296 	}
1297 	KASSERT(fdp->fd_himap == fdp->fd_dhimap);
1298 	KASSERT(fdp->fd_lomap == fdp->fd_dlomap);
1299 #endif	/* DIAGNOSTIC */
1300 
1301 	fdp->fd_refcnt = 1;
1302 	fd_checkmaps(fdp);
1303 
1304 	return fdp;
1305 }
1306 
1307 /*
1308  * Initialize a file descriptor table.
1309  */
1310 static int
1311 filedesc_ctor(void *arg, void *obj, int flag)
1312 {
1313 	filedesc_t *fdp = obj;
1314 	fdfile_t **ffp;
1315 	int i;
1316 
1317 	memset(fdp, 0, sizeof(*fdp));
1318 	mutex_init(&fdp->fd_lock, MUTEX_DEFAULT, IPL_NONE);
1319 	fdp->fd_lastfile = -1;
1320 	fdp->fd_lastkqfile = -1;
1321 	fdp->fd_dt = &fdp->fd_dtbuiltin;
1322 	fdp->fd_dtbuiltin.dt_nfiles = NDFILE;
1323 	fdp->fd_himap = fdp->fd_dhimap;
1324 	fdp->fd_lomap = fdp->fd_dlomap;
1325 
1326 	CTASSERT(sizeof(fdp->fd_dfdfile[0]) >= sizeof(fdfile_t));
1327 	for (i = 0, ffp = fdp->fd_dt->dt_ff; i < NDFDFILE; i++, ffp++) {
1328 		*ffp = (fdfile_t *)fdp->fd_dfdfile[i];
1329 		(void)fdfile_ctor(NULL, fdp->fd_dfdfile[i], PR_WAITOK);
1330 	}
1331 
1332 	return 0;
1333 }
1334 
1335 static void
1336 filedesc_dtor(void *arg, void *obj)
1337 {
1338 	filedesc_t *fdp = obj;
1339 	int i;
1340 
1341 	for (i = 0; i < NDFDFILE; i++) {
1342 		fdfile_dtor(NULL, fdp->fd_dfdfile[i]);
1343 	}
1344 
1345 	mutex_destroy(&fdp->fd_lock);
1346 }
1347 
1348 /*
1349  * Make p share curproc's filedesc structure.
1350  */
1351 void
1352 fd_share(struct proc *p)
1353 {
1354 	filedesc_t *fdp;
1355 
1356 	fdp = curlwp->l_fd;
1357 	p->p_fd = fdp;
1358 	atomic_inc_uint(&fdp->fd_refcnt);
1359 }
1360 
1361 /*
1362  * Acquire a hold on a filedesc structure.
1363  */
1364 void
1365 fd_hold(lwp_t *l)
1366 {
1367 	filedesc_t *fdp = l->l_fd;
1368 
1369 	atomic_inc_uint(&fdp->fd_refcnt);
1370 }
1371 
1372 /*
1373  * Copy a filedesc structure.
1374  */
1375 filedesc_t *
1376 fd_copy(void)
1377 {
1378 	filedesc_t *newfdp, *fdp;
1379 	fdfile_t *ff, **ffp, **nffp, *ff2;
1380 	int i, j, numfiles, lastfile, newlast;
1381 	file_t *fp;
1382 	fdtab_t *newdt;
1383 
1384 	fdp = curproc->p_fd;
1385 	newfdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1386 	newfdp->fd_refcnt = 1;
1387 
1388 #ifdef DIAGNOSTIC
1389 	KASSERT(newfdp->fd_lastfile == -1);
1390 	KASSERT(newfdp->fd_lastkqfile == -1);
1391 	KASSERT(newfdp->fd_knhash == NULL);
1392 	KASSERT(newfdp->fd_freefile == 0);
1393 	KASSERT(newfdp->fd_exclose == false);
1394 	KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1395 	KASSERT(newfdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1396 	for (i = 0; i < NDFDFILE; i++) {
1397 		KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] ==
1398 		    (fdfile_t *)&newfdp->fd_dfdfile[i]);
1399 	}
1400 	for (i = NDFDFILE; i < NDFILE; i++) {
1401 		KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] == NULL);
1402 	}
1403 #endif	/* DIAGNOSTIC */
1404 
1405 	mutex_enter(&fdp->fd_lock);
1406 	fd_checkmaps(fdp);
1407 	numfiles = fdp->fd_dt->dt_nfiles;
1408 	lastfile = fdp->fd_lastfile;
1409 
1410 	/*
1411 	 * If the number of open files fits in the internal arrays
1412 	 * of the open file structure, use them, otherwise allocate
1413 	 * additional memory for the number of descriptors currently
1414 	 * in use.
1415 	 */
1416 	if (lastfile < NDFILE) {
1417 		i = NDFILE;
1418 		newdt = newfdp->fd_dt;
1419 		KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1420 	} else {
1421 		/*
1422 		 * Compute the smallest multiple of NDEXTENT needed
1423 		 * for the file descriptors currently in use,
1424 		 * allowing the table to shrink.
1425 		 */
1426 		i = numfiles;
1427 		while (i >= 2 * NDEXTENT && i > lastfile * 2) {
1428 			i /= 2;
1429 		}
1430 		KASSERT(i > NDFILE);
1431 		newdt = fd_dtab_alloc(i);
1432 		newfdp->fd_dt = newdt;
1433 		memcpy(newdt->dt_ff, newfdp->fd_dtbuiltin.dt_ff,
1434 		    NDFDFILE * sizeof(fdfile_t **));
1435 		memset(newdt->dt_ff + NDFDFILE, 0,
1436 		    (i - NDFDFILE) * sizeof(fdfile_t **));
1437 	}
1438 	if (NDHISLOTS(i) <= NDHISLOTS(NDFILE)) {
1439 		newfdp->fd_himap = newfdp->fd_dhimap;
1440 		newfdp->fd_lomap = newfdp->fd_dlomap;
1441 	} else {
1442 		fd_map_alloc(i, &newfdp->fd_lomap, &newfdp->fd_himap);
1443 		KASSERT(i >= NDENTRIES * NDENTRIES);
1444 		memset(newfdp->fd_himap, 0, NDHISLOTS(i)*sizeof(uint32_t));
1445 		memset(newfdp->fd_lomap, 0, NDLOSLOTS(i)*sizeof(uint32_t));
1446 	}
1447 	newfdp->fd_freefile = fdp->fd_freefile;
1448 	newfdp->fd_exclose = fdp->fd_exclose;
1449 
1450 	ffp = fdp->fd_dt->dt_ff;
1451 	nffp = newdt->dt_ff;
1452 	newlast = -1;
1453 	for (i = 0; i <= (int)lastfile; i++, ffp++, nffp++) {
1454 		KASSERT(i >= NDFDFILE ||
1455 		    *nffp == (fdfile_t *)newfdp->fd_dfdfile[i]);
1456 		ff = *ffp;
1457 		if (ff == NULL || (fp = ff->ff_file) == NULL) {
1458 			/* Descriptor unused, or descriptor half open. */
1459 			KASSERT(!fd_isused(newfdp, i));
1460 			continue;
1461 		}
1462 		if (__predict_false(fp->f_type == DTYPE_KQUEUE)) {
1463 			/* kqueue descriptors cannot be copied. */
1464 			if (i < newfdp->fd_freefile) {
1465 				newfdp->fd_freefile = i;
1466 			}
1467 			continue;
1468 		}
1469 		/* It's active: add a reference to the file. */
1470 		mutex_enter(&fp->f_lock);
1471 		fp->f_count++;
1472 		mutex_exit(&fp->f_lock);
1473 
1474 		/* Allocate an fdfile_t to represent it. */
1475 		if (i >= NDFDFILE) {
1476 			ff2 = pool_cache_get(fdfile_cache, PR_WAITOK);
1477 			*nffp = ff2;
1478 		} else {
1479 			ff2 = newdt->dt_ff[i];
1480 		}
1481 		ff2->ff_file = fp;
1482 		ff2->ff_exclose = ff->ff_exclose;
1483 		ff2->ff_allocated = true;
1484 
1485 		/* Fix up bitmaps. */
1486 		j = i >> NDENTRYSHIFT;
1487 		KASSERT((newfdp->fd_lomap[j] & (1 << (i & NDENTRYMASK))) == 0);
1488 		newfdp->fd_lomap[j] |= 1 << (i & NDENTRYMASK);
1489 		if (__predict_false(newfdp->fd_lomap[j] == ~0)) {
1490 			KASSERT((newfdp->fd_himap[j >> NDENTRYSHIFT] &
1491 			    (1 << (j & NDENTRYMASK))) == 0);
1492 			newfdp->fd_himap[j >> NDENTRYSHIFT] |=
1493 			    1 << (j & NDENTRYMASK);
1494 		}
1495 		newlast = i;
1496 	}
1497 	KASSERT(newdt->dt_ff[0] == (fdfile_t *)newfdp->fd_dfdfile[0]);
1498 	newfdp->fd_lastfile = newlast;
1499 	fd_checkmaps(newfdp);
1500 	mutex_exit(&fdp->fd_lock);
1501 
1502 	return newfdp;
1503 }
1504 
1505 /*
1506  * Release a filedesc structure.
1507  */
1508 void
1509 fd_free(void)
1510 {
1511 	fdfile_t *ff;
1512 	file_t *fp;
1513 	int fd, nf;
1514 	fdtab_t *dt;
1515 	lwp_t * const l = curlwp;
1516 	filedesc_t * const fdp = l->l_fd;
1517 	const bool noadvlock = (l->l_proc->p_flag & PK_ADVLOCK) == 0;
1518 
1519 	KASSERT(fdp->fd_dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1520 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1521 	KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1522 
1523 #ifndef __HAVE_ATOMIC_AS_MEMBAR
1524 	membar_exit();
1525 #endif
1526 	if (atomic_dec_uint_nv(&fdp->fd_refcnt) > 0)
1527 		return;
1528 
1529 	/*
1530 	 * Close any files that the process holds open.
1531 	 */
1532 	dt = fdp->fd_dt;
1533 	fd_checkmaps(fdp);
1534 #ifdef DEBUG
1535 	fdp->fd_refcnt = -1; /* see fd_checkmaps */
1536 #endif
1537 	for (fd = 0, nf = dt->dt_nfiles; fd < nf; fd++) {
1538 		ff = dt->dt_ff[fd];
1539 		KASSERT(fd >= NDFDFILE ||
1540 		    ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1541 		if (ff == NULL)
1542 			continue;
1543 		if ((fp = ff->ff_file) != NULL) {
1544 			/*
1545 			 * Must use fd_close() here if there is
1546 			 * a reference from kqueue or we might have posix
1547 			 * advisory locks.
1548 			 */
1549 			if (__predict_true(ff->ff_refcnt == 0) &&
1550 			    (noadvlock || fp->f_type != DTYPE_VNODE)) {
1551 				ff->ff_file = NULL;
1552 				ff->ff_exclose = false;
1553 				ff->ff_allocated = false;
1554 				closef(fp);
1555 			} else {
1556 				ff->ff_refcnt++;
1557 				fd_close(fd);
1558 			}
1559 		}
1560 		KASSERT(ff->ff_refcnt == 0);
1561 		KASSERT(ff->ff_file == NULL);
1562 		KASSERT(!ff->ff_exclose);
1563 		KASSERT(!ff->ff_allocated);
1564 		if (fd >= NDFDFILE) {
1565 			pool_cache_put(fdfile_cache, ff);
1566 			dt->dt_ff[fd] = NULL;
1567 		}
1568 	}
1569 
1570 	/*
1571 	 * Clean out the descriptor table for the next user and return
1572 	 * to the cache.
1573 	 */
1574 	if (__predict_false(dt != &fdp->fd_dtbuiltin)) {
1575 		fd_dtab_free(fdp->fd_dt);
1576 		/* Otherwise, done above. */
1577 		memset(&fdp->fd_dtbuiltin.dt_ff[NDFDFILE], 0,
1578 		    (NDFILE - NDFDFILE) * sizeof(fdp->fd_dtbuiltin.dt_ff[0]));
1579 		fdp->fd_dt = &fdp->fd_dtbuiltin;
1580 	}
1581 	if (__predict_false(NDHISLOTS(nf) > NDHISLOTS(NDFILE))) {
1582 		KASSERT(fdp->fd_himap != fdp->fd_dhimap);
1583 		KASSERT(fdp->fd_lomap != fdp->fd_dlomap);
1584 		fd_map_free(nf, fdp->fd_lomap, fdp->fd_himap);
1585 	}
1586 	if (__predict_false(fdp->fd_knhash != NULL)) {
1587 		hashdone(fdp->fd_knhash, HASH_LIST, fdp->fd_knhashmask);
1588 		fdp->fd_knhash = NULL;
1589 		fdp->fd_knhashmask = 0;
1590 	} else {
1591 		KASSERT(fdp->fd_knhashmask == 0);
1592 	}
1593 	fdp->fd_dt = &fdp->fd_dtbuiltin;
1594 	fdp->fd_lastkqfile = -1;
1595 	fdp->fd_lastfile = -1;
1596 	fdp->fd_freefile = 0;
1597 	fdp->fd_exclose = false;
1598 	memset(&fdp->fd_startzero, 0, sizeof(*fdp) -
1599 	    offsetof(filedesc_t, fd_startzero));
1600 	fdp->fd_himap = fdp->fd_dhimap;
1601 	fdp->fd_lomap = fdp->fd_dlomap;
1602 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1603 	KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1604 	KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1605 #ifdef DEBUG
1606 	fdp->fd_refcnt = 0; /* see fd_checkmaps */
1607 #endif
1608 	fd_checkmaps(fdp);
1609 	pool_cache_put(filedesc_cache, fdp);
1610 }
1611 
1612 /*
1613  * File Descriptor pseudo-device driver (/dev/fd/).
1614  *
1615  * Opening minor device N dup()s the file (if any) connected to file
1616  * descriptor N belonging to the calling process.  Note that this driver
1617  * consists of only the ``open()'' routine, because all subsequent
1618  * references to this file will be direct to the other driver.
1619  */
1620 static int
1621 filedescopen(dev_t dev, int mode, int type, lwp_t *l)
1622 {
1623 
1624 	/*
1625 	 * XXX Kludge: set dupfd to contain the value of the
1626 	 * the file descriptor being sought for duplication. The error
1627 	 * return ensures that the vnode for this device will be released
1628 	 * by vn_open. Open will detect this special error and take the
1629 	 * actions in fd_dupopen below. Other callers of vn_open or VOP_OPEN
1630 	 * will simply report the error.
1631 	 */
1632 	l->l_dupfd = minor(dev);	/* XXX */
1633 	return EDUPFD;
1634 }
1635 
1636 /*
1637  * Duplicate the specified descriptor to a free descriptor.
1638  */
1639 int
1640 fd_dupopen(int old, int *new, int mode, int error)
1641 {
1642 	filedesc_t *fdp;
1643 	fdfile_t *ff;
1644 	file_t *fp;
1645 	fdtab_t *dt;
1646 
1647 	if ((fp = fd_getfile(old)) == NULL) {
1648 		return EBADF;
1649 	}
1650 	fdp = curlwp->l_fd;
1651 	dt = fdp->fd_dt;
1652 	ff = dt->dt_ff[old];
1653 
1654 	/*
1655 	 * There are two cases of interest here.
1656 	 *
1657 	 * For EDUPFD simply dup (old) to file descriptor
1658 	 * (new) and return.
1659 	 *
1660 	 * For EMOVEFD steal away the file structure from (old) and
1661 	 * store it in (new).  (old) is effectively closed by
1662 	 * this operation.
1663 	 *
1664 	 * Any other error code is just returned.
1665 	 */
1666 	switch (error) {
1667 	case EDUPFD:
1668 		/*
1669 		 * Check that the mode the file is being opened for is a
1670 		 * subset of the mode of the existing descriptor.
1671 		 */
1672 		if (((mode & (FREAD|FWRITE)) | fp->f_flag) != fp->f_flag) {
1673 			error = EACCES;
1674 			break;
1675 		}
1676 
1677 		/* Copy it. */
1678 		error = fd_dup(fp, 0, new, ff->ff_exclose);
1679 		break;
1680 
1681 	case EMOVEFD:
1682 		/* Copy it. */
1683 		error = fd_dup(fp, 0, new, ff->ff_exclose);
1684 		if (error != 0) {
1685 			break;
1686 		}
1687 
1688 		/* Steal away the file pointer from 'old'. */
1689 		(void)fd_close(old);
1690 		return 0;
1691 	}
1692 
1693 	fd_putfile(old);
1694 	return error;
1695 }
1696 
1697 /*
1698  * Close open files on exec.
1699  */
1700 void
1701 fd_closeexec(void)
1702 {
1703 	proc_t *p;
1704 	filedesc_t *fdp;
1705 	fdfile_t *ff;
1706 	lwp_t *l;
1707 	fdtab_t *dt;
1708 	int fd;
1709 
1710 	l = curlwp;
1711 	p = l->l_proc;
1712 	fdp = p->p_fd;
1713 
1714 	if (fdp->fd_refcnt > 1) {
1715 		fdp = fd_copy();
1716 		fd_free();
1717 		p->p_fd = fdp;
1718 		l->l_fd = fdp;
1719 	}
1720 	if (!fdp->fd_exclose) {
1721 		return;
1722 	}
1723 	fdp->fd_exclose = false;
1724 	dt = fdp->fd_dt;
1725 
1726 	for (fd = 0; fd <= fdp->fd_lastfile; fd++) {
1727 		if ((ff = dt->dt_ff[fd]) == NULL) {
1728 			KASSERT(fd >= NDFDFILE);
1729 			continue;
1730 		}
1731 		KASSERT(fd >= NDFDFILE ||
1732 		    ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1733 		if (ff->ff_file == NULL)
1734 			continue;
1735 		if (ff->ff_exclose) {
1736 			/*
1737 			 * We need a reference to close the file.
1738 			 * No other threads can see the fdfile_t at
1739 			 * this point, so don't bother locking.
1740 			 */
1741 			KASSERT((ff->ff_refcnt & FR_CLOSING) == 0);
1742 			ff->ff_refcnt++;
1743 			fd_close(fd);
1744 		}
1745 	}
1746 }
1747 
1748 /*
1749  * Sets descriptor owner. If the owner is a process, 'pgid'
1750  * is set to positive value, process ID. If the owner is process group,
1751  * 'pgid' is set to -pg_id.
1752  */
1753 int
1754 fsetown(pid_t *pgid, u_long cmd, const void *data)
1755 {
1756 	pid_t id = *(const pid_t *)data;
1757 	int error;
1758 
1759 	switch (cmd) {
1760 	case TIOCSPGRP:
1761 		if (id < 0)
1762 			return EINVAL;
1763 		id = -id;
1764 		break;
1765 	default:
1766 		break;
1767 	}
1768 	if (id > 0) {
1769 		mutex_enter(proc_lock);
1770 		error = proc_find(id) ? 0 : ESRCH;
1771 		mutex_exit(proc_lock);
1772 	} else if (id < 0) {
1773 		error = pgid_in_session(curproc, -id);
1774 	} else {
1775 		error = 0;
1776 	}
1777 	if (!error) {
1778 		*pgid = id;
1779 	}
1780 	return error;
1781 }
1782 
1783 void
1784 fd_set_exclose(struct lwp *l, int fd, bool exclose)
1785 {
1786 	filedesc_t *fdp = l->l_fd;
1787 	fdfile_t *ff = fdp->fd_dt->dt_ff[fd];
1788 
1789 	ff->ff_exclose = exclose;
1790 	if (exclose)
1791 		fdp->fd_exclose = true;
1792 }
1793 
1794 /*
1795  * Return descriptor owner information. If the value is positive,
1796  * it's process ID. If it's negative, it's process group ID and
1797  * needs the sign removed before use.
1798  */
1799 int
1800 fgetown(pid_t pgid, u_long cmd, void *data)
1801 {
1802 
1803 	switch (cmd) {
1804 	case TIOCGPGRP:
1805 		*(int *)data = -pgid;
1806 		break;
1807 	default:
1808 		*(int *)data = pgid;
1809 		break;
1810 	}
1811 	return 0;
1812 }
1813 
1814 /*
1815  * Send signal to descriptor owner, either process or process group.
1816  */
1817 void
1818 fownsignal(pid_t pgid, int signo, int code, int band, void *fdescdata)
1819 {
1820 	ksiginfo_t ksi;
1821 
1822 	KASSERT(!cpu_intr_p());
1823 
1824 	if (pgid == 0) {
1825 		return;
1826 	}
1827 
1828 	KSI_INIT(&ksi);
1829 	ksi.ksi_signo = signo;
1830 	ksi.ksi_code = code;
1831 	ksi.ksi_band = band;
1832 
1833 	mutex_enter(proc_lock);
1834 	if (pgid > 0) {
1835 		struct proc *p1;
1836 
1837 		p1 = proc_find(pgid);
1838 		if (p1 != NULL) {
1839 			kpsignal(p1, &ksi, fdescdata);
1840 		}
1841 	} else {
1842 		struct pgrp *pgrp;
1843 
1844 		KASSERT(pgid < 0);
1845 		pgrp = pgrp_find(-pgid);
1846 		if (pgrp != NULL) {
1847 			kpgsignal(pgrp, &ksi, fdescdata, 0);
1848 		}
1849 	}
1850 	mutex_exit(proc_lock);
1851 }
1852 
1853 int
1854 fd_clone(file_t *fp, unsigned fd, int flag, const struct fileops *fops,
1855 	 void *data)
1856 {
1857 
1858 	fp->f_flag = flag;
1859 	fp->f_type = DTYPE_MISC;
1860 	fp->f_ops = fops;
1861 	fp->f_data = data;
1862 	curlwp->l_dupfd = fd;
1863 	fd_affix(curproc, fp, fd);
1864 
1865 	return EMOVEFD;
1866 }
1867 
1868 int
1869 fnullop_fcntl(file_t *fp, u_int cmd, void *data)
1870 {
1871 
1872 	if (cmd == F_SETFL)
1873 		return 0;
1874 
1875 	return EOPNOTSUPP;
1876 }
1877 
1878 int
1879 fnullop_poll(file_t *fp, int which)
1880 {
1881 
1882 	return 0;
1883 }
1884 
1885 int
1886 fnullop_kqfilter(file_t *fp, struct knote *kn)
1887 {
1888 
1889 	return EOPNOTSUPP;
1890 }
1891 
1892 void
1893 fnullop_restart(file_t *fp)
1894 {
1895 
1896 }
1897 
1898 int
1899 fbadop_read(file_t *fp, off_t *offset, struct uio *uio,
1900 	    kauth_cred_t cred, int flags)
1901 {
1902 
1903 	return EOPNOTSUPP;
1904 }
1905 
1906 int
1907 fbadop_write(file_t *fp, off_t *offset, struct uio *uio,
1908 	     kauth_cred_t cred, int flags)
1909 {
1910 
1911 	return EOPNOTSUPP;
1912 }
1913 
1914 int
1915 fbadop_ioctl(file_t *fp, u_long com, void *data)
1916 {
1917 
1918 	return EOPNOTSUPP;
1919 }
1920 
1921 int
1922 fbadop_stat(file_t *fp, struct stat *sb)
1923 {
1924 
1925 	return EOPNOTSUPP;
1926 }
1927 
1928 int
1929 fbadop_close(file_t *fp)
1930 {
1931 
1932 	return EOPNOTSUPP;
1933 }
1934 
1935 /*
1936  * sysctl routines pertaining to file descriptors
1937  */
1938 
1939 /* Initialized in sysctl_init() for now... */
1940 extern kmutex_t sysctl_file_marker_lock;
1941 static u_int sysctl_file_marker = 1;
1942 
1943 /*
1944  * Expects to be called with proc_lock and sysctl_file_marker_lock locked.
1945  */
1946 static void
1947 sysctl_file_marker_reset(void)
1948 {
1949 	struct proc *p;
1950 
1951 	PROCLIST_FOREACH(p, &allproc) {
1952 		struct filedesc *fd = p->p_fd;
1953 		fdtab_t *dt;
1954 		u_int i;
1955 
1956 		mutex_enter(&fd->fd_lock);
1957 		dt = fd->fd_dt;
1958 		for (i = 0; i < dt->dt_nfiles; i++) {
1959 			struct file *fp;
1960 			fdfile_t *ff;
1961 
1962 			if ((ff = dt->dt_ff[i]) == NULL) {
1963 				continue;
1964 			}
1965 			if ((fp = ff->ff_file) == NULL) {
1966 				continue;
1967 			}
1968 			fp->f_marker = 0;
1969 		}
1970 		mutex_exit(&fd->fd_lock);
1971 	}
1972 }
1973 
1974 /*
1975  * sysctl helper routine for kern.file pseudo-subtree.
1976  */
1977 static int
1978 sysctl_kern_file(SYSCTLFN_ARGS)
1979 {
1980 	int error;
1981 	size_t buflen;
1982 	struct file *fp, fbuf;
1983 	char *start, *where;
1984 	struct proc *p;
1985 
1986 	start = where = oldp;
1987 	buflen = *oldlenp;
1988 
1989 	if (where == NULL) {
1990 		/*
1991 		 * overestimate by 10 files
1992 		 */
1993 		*oldlenp = sizeof(filehead) + (nfiles + 10) *
1994 		    sizeof(struct file);
1995 		return 0;
1996 	}
1997 
1998 	/*
1999 	 * first sysctl_copyout filehead
2000 	 */
2001 	if (buflen < sizeof(filehead)) {
2002 		*oldlenp = 0;
2003 		return 0;
2004 	}
2005 	sysctl_unlock();
2006 	error = sysctl_copyout(l, &filehead, where, sizeof(filehead));
2007 	if (error) {
2008 		sysctl_relock();
2009 		return error;
2010 	}
2011 	buflen -= sizeof(filehead);
2012 	where += sizeof(filehead);
2013 
2014 	/*
2015 	 * followed by an array of file structures
2016 	 */
2017 	mutex_enter(&sysctl_file_marker_lock);
2018 	mutex_enter(proc_lock);
2019 	PROCLIST_FOREACH(p, &allproc) {
2020 		struct filedesc *fd;
2021 		fdtab_t *dt;
2022 		u_int i;
2023 
2024 		if (p->p_stat == SIDL) {
2025 			/* skip embryonic processes */
2026 			continue;
2027 		}
2028 		mutex_enter(p->p_lock);
2029 		error = kauth_authorize_process(l->l_cred,
2030 		    KAUTH_PROCESS_CANSEE, p,
2031 		    KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2032 		    NULL, NULL);
2033 		mutex_exit(p->p_lock);
2034 		if (error != 0) {
2035 			/*
2036 			 * Don't leak kauth retval if we're silently
2037 			 * skipping this entry.
2038 			 */
2039 			error = 0;
2040 			continue;
2041 		}
2042 
2043 		/*
2044 		 * Grab a hold on the process.
2045 		 */
2046 		if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2047 			continue;
2048 		}
2049 		mutex_exit(proc_lock);
2050 
2051 		fd = p->p_fd;
2052 		mutex_enter(&fd->fd_lock);
2053 		dt = fd->fd_dt;
2054 		for (i = 0; i < dt->dt_nfiles; i++) {
2055 			fdfile_t *ff;
2056 
2057 			if ((ff = dt->dt_ff[i]) == NULL) {
2058 				continue;
2059 			}
2060 			if ((fp = ff->ff_file) == NULL) {
2061 				continue;
2062 			}
2063 
2064 			mutex_enter(&fp->f_lock);
2065 
2066 			if ((fp->f_count == 0) ||
2067 			    (fp->f_marker == sysctl_file_marker)) {
2068 				mutex_exit(&fp->f_lock);
2069 				continue;
2070 			}
2071 
2072 			/* Check that we have enough space. */
2073 			if (buflen < sizeof(struct file)) {
2074 				*oldlenp = where - start;
2075 				mutex_exit(&fp->f_lock);
2076 				error = ENOMEM;
2077 				break;
2078 			}
2079 
2080 			memcpy(&fbuf, fp, sizeof(fbuf));
2081 			mutex_exit(&fp->f_lock);
2082 			error = sysctl_copyout(l, &fbuf, where, sizeof(fbuf));
2083 			if (error) {
2084 				break;
2085 			}
2086 			buflen -= sizeof(struct file);
2087 			where += sizeof(struct file);
2088 
2089 			fp->f_marker = sysctl_file_marker;
2090 		}
2091 		mutex_exit(&fd->fd_lock);
2092 
2093 		/*
2094 		 * Release reference to process.
2095 		 */
2096 		mutex_enter(proc_lock);
2097 		rw_exit(&p->p_reflock);
2098 
2099 		if (error)
2100 			break;
2101 	}
2102 
2103 	sysctl_file_marker++;
2104 	/* Reset all markers if wrapped. */
2105 	if (sysctl_file_marker == 0) {
2106 		sysctl_file_marker_reset();
2107 		sysctl_file_marker++;
2108 	}
2109 
2110 	mutex_exit(proc_lock);
2111 	mutex_exit(&sysctl_file_marker_lock);
2112 
2113 	*oldlenp = where - start;
2114 	sysctl_relock();
2115 	return error;
2116 }
2117 
2118 /*
2119  * sysctl helper function for kern.file2
2120  */
2121 static int
2122 sysctl_kern_file2(SYSCTLFN_ARGS)
2123 {
2124 	struct proc *p;
2125 	struct file *fp;
2126 	struct filedesc *fd;
2127 	struct kinfo_file kf;
2128 	char *dp;
2129 	u_int i, op;
2130 	size_t len, needed, elem_size, out_size;
2131 	int error, arg, elem_count;
2132 	fdfile_t *ff;
2133 	fdtab_t *dt;
2134 
2135 	if (namelen == 1 && name[0] == CTL_QUERY)
2136 		return sysctl_query(SYSCTLFN_CALL(rnode));
2137 
2138 	if (namelen != 4)
2139 		return EINVAL;
2140 
2141 	error = 0;
2142 	dp = oldp;
2143 	len = (oldp != NULL) ? *oldlenp : 0;
2144 	op = name[0];
2145 	arg = name[1];
2146 	elem_size = name[2];
2147 	elem_count = name[3];
2148 	out_size = MIN(sizeof(kf), elem_size);
2149 	needed = 0;
2150 
2151 	if (elem_size < 1 || elem_count < 0)
2152 		return EINVAL;
2153 
2154 	switch (op) {
2155 	case KERN_FILE_BYFILE:
2156 	case KERN_FILE_BYPID:
2157 		/*
2158 		 * We're traversing the process list in both cases; the BYFILE
2159 		 * case does additional work of keeping track of files already
2160 		 * looked at.
2161 		 */
2162 
2163 		/* doesn't use arg so it must be zero */
2164 		if ((op == KERN_FILE_BYFILE) && (arg != 0))
2165 			return EINVAL;
2166 
2167 		if ((op == KERN_FILE_BYPID) && (arg < -1))
2168 			/* -1 means all processes */
2169 			return EINVAL;
2170 
2171 		sysctl_unlock();
2172 		if (op == KERN_FILE_BYFILE)
2173 			mutex_enter(&sysctl_file_marker_lock);
2174 		mutex_enter(proc_lock);
2175 		PROCLIST_FOREACH(p, &allproc) {
2176 			if (p->p_stat == SIDL) {
2177 				/* skip embryonic processes */
2178 				continue;
2179 			}
2180 			if (arg > 0 && p->p_pid != arg) {
2181 				/* pick only the one we want */
2182 				/* XXX want 0 to mean "kernel files" */
2183 				continue;
2184 			}
2185 			mutex_enter(p->p_lock);
2186 			error = kauth_authorize_process(l->l_cred,
2187 			    KAUTH_PROCESS_CANSEE, p,
2188 			    KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2189 			    NULL, NULL);
2190 			mutex_exit(p->p_lock);
2191 			if (error != 0) {
2192 				/*
2193 				 * Don't leak kauth retval if we're silently
2194 				 * skipping this entry.
2195 				 */
2196 				error = 0;
2197 				continue;
2198 			}
2199 
2200 			/*
2201 			 * Grab a hold on the process.
2202 			 */
2203 			if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2204 				continue;
2205 			}
2206 			mutex_exit(proc_lock);
2207 
2208 			fd = p->p_fd;
2209 			mutex_enter(&fd->fd_lock);
2210 			dt = fd->fd_dt;
2211 			for (i = 0; i < dt->dt_nfiles; i++) {
2212 				if ((ff = dt->dt_ff[i]) == NULL) {
2213 					continue;
2214 				}
2215 				if ((fp = ff->ff_file) == NULL) {
2216 					continue;
2217 				}
2218 
2219 				if ((op == KERN_FILE_BYFILE) &&
2220 				    (fp->f_marker == sysctl_file_marker)) {
2221 					continue;
2222 				}
2223 				if (len >= elem_size && elem_count > 0) {
2224 					mutex_enter(&fp->f_lock);
2225 					fill_file(&kf, fp, ff, i, p->p_pid);
2226 					mutex_exit(&fp->f_lock);
2227 					mutex_exit(&fd->fd_lock);
2228 					error = sysctl_copyout(l,
2229 					    &kf, dp, out_size);
2230 					mutex_enter(&fd->fd_lock);
2231 					if (error)
2232 						break;
2233 					dp += elem_size;
2234 					len -= elem_size;
2235 				}
2236 				if (op == KERN_FILE_BYFILE)
2237 					fp->f_marker = sysctl_file_marker;
2238 				needed += elem_size;
2239 				if (elem_count > 0 && elem_count != INT_MAX)
2240 					elem_count--;
2241 			}
2242 			mutex_exit(&fd->fd_lock);
2243 
2244 			/*
2245 			 * Release reference to process.
2246 			 */
2247 			mutex_enter(proc_lock);
2248 			rw_exit(&p->p_reflock);
2249 		}
2250 		if (op == KERN_FILE_BYFILE) {
2251 			sysctl_file_marker++;
2252 
2253 			/* Reset all markers if wrapped. */
2254 			if (sysctl_file_marker == 0) {
2255 				sysctl_file_marker_reset();
2256 				sysctl_file_marker++;
2257 			}
2258 		}
2259 		mutex_exit(proc_lock);
2260 		if (op == KERN_FILE_BYFILE)
2261 			mutex_exit(&sysctl_file_marker_lock);
2262 		sysctl_relock();
2263 		break;
2264 	default:
2265 		return EINVAL;
2266 	}
2267 
2268 	if (oldp == NULL)
2269 		needed += KERN_FILESLOP * elem_size;
2270 	*oldlenp = needed;
2271 
2272 	return error;
2273 }
2274 
2275 static void
2276 fill_file(struct kinfo_file *kp, const file_t *fp, const fdfile_t *ff,
2277 	  int i, pid_t pid)
2278 {
2279 
2280 	memset(kp, 0, sizeof(*kp));
2281 
2282 	kp->ki_fileaddr =	PTRTOUINT64(fp);
2283 	kp->ki_flag =		fp->f_flag;
2284 	kp->ki_iflags =		0;
2285 	kp->ki_ftype =		fp->f_type;
2286 	kp->ki_count =		fp->f_count;
2287 	kp->ki_msgcount =	fp->f_msgcount;
2288 	kp->ki_fucred =		PTRTOUINT64(fp->f_cred);
2289 	kp->ki_fuid =		kauth_cred_geteuid(fp->f_cred);
2290 	kp->ki_fgid =		kauth_cred_getegid(fp->f_cred);
2291 	kp->ki_fops =		PTRTOUINT64(fp->f_ops);
2292 	kp->ki_foffset =	fp->f_offset;
2293 	kp->ki_fdata =		PTRTOUINT64(fp->f_data);
2294 
2295 	/* vnode information to glue this file to something */
2296 	if (fp->f_type == DTYPE_VNODE) {
2297 		struct vnode *vp = (struct vnode *)fp->f_data;
2298 
2299 		kp->ki_vun =	PTRTOUINT64(vp->v_un.vu_socket);
2300 		kp->ki_vsize =	vp->v_size;
2301 		kp->ki_vtype =	vp->v_type;
2302 		kp->ki_vtag =	vp->v_tag;
2303 		kp->ki_vdata =	PTRTOUINT64(vp->v_data);
2304 	}
2305 
2306 	/* process information when retrieved via KERN_FILE_BYPID */
2307 	if (ff != NULL) {
2308 		kp->ki_pid =		pid;
2309 		kp->ki_fd =		i;
2310 		kp->ki_ofileflags =	ff->ff_exclose;
2311 		kp->ki_usecount =	ff->ff_refcnt;
2312 	}
2313 }
2314