xref: /netbsd-src/sys/kern/kern_descrip.c (revision 181254a7b1bdde6873432bffef2d2decc4b5c22f)
1 /*	$NetBSD: kern_descrip.c,v 1.249 2020/08/28 10:20:14 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 1.249 2020/08/28 10:20:14 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 file *, const struct file *);
122 static void fill_file2(struct kinfo_file *, const file_t *, const fdfile_t *,
123 		      int, pid_t);
124 
125 const struct cdevsw filedesc_cdevsw = {
126 	.d_open = filedescopen,
127 	.d_close = noclose,
128 	.d_read = noread,
129 	.d_write = nowrite,
130 	.d_ioctl = noioctl,
131 	.d_stop = nostop,
132 	.d_tty = notty,
133 	.d_poll = nopoll,
134 	.d_mmap = nommap,
135 	.d_kqfilter = nokqfilter,
136 	.d_discard = nodiscard,
137 	.d_flag = D_OTHER | D_MPSAFE
138 };
139 
140 /* For ease of reading. */
141 __strong_alias(fd_putvnode,fd_putfile)
142 __strong_alias(fd_putsock,fd_putfile)
143 
144 /*
145  * Initialize the descriptor system.
146  */
147 void
148 fd_sys_init(void)
149 {
150 	static struct sysctllog *clog;
151 
152 	mutex_init(&filelist_lock, MUTEX_DEFAULT, IPL_NONE);
153 
154 	LIST_INIT(&filehead);
155 
156 	file_cache = pool_cache_init(sizeof(file_t), coherency_unit, 0,
157 	    0, "file", NULL, IPL_NONE, file_ctor, file_dtor, NULL);
158 	KASSERT(file_cache != NULL);
159 
160 	fdfile_cache = pool_cache_init(sizeof(fdfile_t), coherency_unit, 0,
161 	    PR_LARGECACHE, "fdfile", NULL, IPL_NONE, fdfile_ctor, fdfile_dtor,
162 	    NULL);
163 	KASSERT(fdfile_cache != NULL);
164 
165 	filedesc_cache = pool_cache_init(sizeof(filedesc_t), coherency_unit,
166 	    0, 0, "filedesc", NULL, IPL_NONE, filedesc_ctor, filedesc_dtor,
167 	    NULL);
168 	KASSERT(filedesc_cache != NULL);
169 
170 	sysctl_createv(&clog, 0, NULL, NULL,
171 		       CTLFLAG_PERMANENT,
172 		       CTLTYPE_STRUCT, "file",
173 		       SYSCTL_DESCR("System open file table"),
174 		       sysctl_kern_file, 0, NULL, 0,
175 		       CTL_KERN, KERN_FILE, CTL_EOL);
176 	sysctl_createv(&clog, 0, NULL, NULL,
177 		       CTLFLAG_PERMANENT,
178 		       CTLTYPE_STRUCT, "file2",
179 		       SYSCTL_DESCR("System open file table"),
180 		       sysctl_kern_file2, 0, NULL, 0,
181 		       CTL_KERN, KERN_FILE2, CTL_EOL);
182 }
183 
184 static bool
185 fd_isused(filedesc_t *fdp, unsigned fd)
186 {
187 	u_int off = fd >> NDENTRYSHIFT;
188 
189 	KASSERT(fd < atomic_load_consume(&fdp->fd_dt)->dt_nfiles);
190 
191 	return (fdp->fd_lomap[off] & (1U << (fd & NDENTRYMASK))) != 0;
192 }
193 
194 /*
195  * Verify that the bitmaps match the descriptor table.
196  */
197 static inline void
198 fd_checkmaps(filedesc_t *fdp)
199 {
200 #ifdef DEBUG
201 	fdtab_t *dt;
202 	u_int fd;
203 
204 	KASSERT(fdp->fd_refcnt <= 1 || mutex_owned(&fdp->fd_lock));
205 
206 	dt = fdp->fd_dt;
207 	if (fdp->fd_refcnt == -1) {
208 		/*
209 		 * fd_free tears down the table without maintaining its bitmap.
210 		 */
211 		return;
212 	}
213 	for (fd = 0; fd < dt->dt_nfiles; fd++) {
214 		if (fd < NDFDFILE) {
215 			KASSERT(dt->dt_ff[fd] ==
216 			    (fdfile_t *)fdp->fd_dfdfile[fd]);
217 		}
218 		if (dt->dt_ff[fd] == NULL) {
219 			KASSERT(!fd_isused(fdp, fd));
220 		} else if (dt->dt_ff[fd]->ff_file != NULL) {
221 			KASSERT(fd_isused(fdp, fd));
222 		}
223 	}
224 #endif
225 }
226 
227 static int
228 fd_next_zero(filedesc_t *fdp, uint32_t *bitmap, int want, u_int bits)
229 {
230 	int i, off, maxoff;
231 	uint32_t sub;
232 
233 	KASSERT(mutex_owned(&fdp->fd_lock));
234 
235 	fd_checkmaps(fdp);
236 
237 	if (want > bits)
238 		return -1;
239 
240 	off = want >> NDENTRYSHIFT;
241 	i = want & NDENTRYMASK;
242 	if (i) {
243 		sub = bitmap[off] | ((u_int)~0 >> (NDENTRIES - i));
244 		if (sub != ~0)
245 			goto found;
246 		off++;
247 	}
248 
249 	maxoff = NDLOSLOTS(bits);
250 	while (off < maxoff) {
251 		if ((sub = bitmap[off]) != ~0)
252 			goto found;
253 		off++;
254 	}
255 
256 	return -1;
257 
258  found:
259 	return (off << NDENTRYSHIFT) + ffs(~sub) - 1;
260 }
261 
262 static int
263 fd_last_set(filedesc_t *fd, int last)
264 {
265 	int off, i;
266 	fdfile_t **ff = fd->fd_dt->dt_ff;
267 	uint32_t *bitmap = fd->fd_lomap;
268 
269 	KASSERT(mutex_owned(&fd->fd_lock));
270 
271 	fd_checkmaps(fd);
272 
273 	off = (last - 1) >> NDENTRYSHIFT;
274 
275 	while (off >= 0 && !bitmap[off])
276 		off--;
277 
278 	if (off < 0)
279 		return -1;
280 
281 	i = ((off + 1) << NDENTRYSHIFT) - 1;
282 	if (i >= last)
283 		i = last - 1;
284 
285 	/* XXX should use bitmap */
286 	while (i > 0 && (ff[i] == NULL || !ff[i]->ff_allocated))
287 		i--;
288 
289 	return i;
290 }
291 
292 static inline void
293 fd_used(filedesc_t *fdp, unsigned fd)
294 {
295 	u_int off = fd >> NDENTRYSHIFT;
296 	fdfile_t *ff;
297 
298 	ff = fdp->fd_dt->dt_ff[fd];
299 
300 	KASSERT(mutex_owned(&fdp->fd_lock));
301 	KASSERT((fdp->fd_lomap[off] & (1U << (fd & NDENTRYMASK))) == 0);
302 	KASSERT(ff != NULL);
303 	KASSERT(ff->ff_file == NULL);
304 	KASSERT(!ff->ff_allocated);
305 
306 	ff->ff_allocated = true;
307 	fdp->fd_lomap[off] |= 1U << (fd & NDENTRYMASK);
308 	if (__predict_false(fdp->fd_lomap[off] == ~0)) {
309 		KASSERT((fdp->fd_himap[off >> NDENTRYSHIFT] &
310 		    (1U << (off & NDENTRYMASK))) == 0);
311 		fdp->fd_himap[off >> NDENTRYSHIFT] |= 1U << (off & NDENTRYMASK);
312 	}
313 
314 	if ((int)fd > fdp->fd_lastfile) {
315 		fdp->fd_lastfile = fd;
316 	}
317 
318 	fd_checkmaps(fdp);
319 }
320 
321 static inline void
322 fd_unused(filedesc_t *fdp, unsigned fd)
323 {
324 	u_int off = fd >> NDENTRYSHIFT;
325 	fdfile_t *ff;
326 
327 	ff = fdp->fd_dt->dt_ff[fd];
328 
329 	KASSERT(mutex_owned(&fdp->fd_lock));
330 	KASSERT(ff != NULL);
331 	KASSERT(ff->ff_file == NULL);
332 	KASSERT(ff->ff_allocated);
333 
334 	if (fd < fdp->fd_freefile) {
335 		fdp->fd_freefile = fd;
336 	}
337 
338 	if (fdp->fd_lomap[off] == ~0) {
339 		KASSERT((fdp->fd_himap[off >> NDENTRYSHIFT] &
340 		    (1U << (off & NDENTRYMASK))) != 0);
341 		fdp->fd_himap[off >> NDENTRYSHIFT] &=
342 		    ~(1U << (off & NDENTRYMASK));
343 	}
344 	KASSERT((fdp->fd_lomap[off] & (1U << (fd & NDENTRYMASK))) != 0);
345 	fdp->fd_lomap[off] &= ~(1U << (fd & NDENTRYMASK));
346 	ff->ff_allocated = false;
347 
348 	KASSERT(fd <= fdp->fd_lastfile);
349 	if (fd == fdp->fd_lastfile) {
350 		fdp->fd_lastfile = fd_last_set(fdp, fd);
351 	}
352 	fd_checkmaps(fdp);
353 }
354 
355 /*
356  * Look up the file structure corresponding to a file descriptor
357  * and return the file, holding a reference on the descriptor.
358  */
359 file_t *
360 fd_getfile(unsigned fd)
361 {
362 	filedesc_t *fdp;
363 	fdfile_t *ff;
364 	file_t *fp;
365 	fdtab_t *dt;
366 
367 	/*
368 	 * Look up the fdfile structure representing this descriptor.
369 	 * We are doing this unlocked.  See fd_tryexpand().
370 	 */
371 	fdp = curlwp->l_fd;
372 	dt = atomic_load_consume(&fdp->fd_dt);
373 	if (__predict_false(fd >= dt->dt_nfiles)) {
374 		return NULL;
375 	}
376 	ff = dt->dt_ff[fd];
377 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
378 	if (__predict_false(ff == NULL)) {
379 		return NULL;
380 	}
381 
382 	/* Now get a reference to the descriptor. */
383 	if (fdp->fd_refcnt == 1) {
384 		/*
385 		 * Single threaded: don't need to worry about concurrent
386 		 * access (other than earlier calls to kqueue, which may
387 		 * hold a reference to the descriptor).
388 		 */
389 		ff->ff_refcnt++;
390 	} else {
391 		/*
392 		 * Multi threaded: issue a memory barrier to ensure that we
393 		 * acquire the file pointer _after_ adding a reference.  If
394 		 * no memory barrier, we could fetch a stale pointer.
395 		 */
396 		atomic_inc_uint(&ff->ff_refcnt);
397 #ifndef __HAVE_ATOMIC_AS_MEMBAR
398 		membar_enter();
399 #endif
400 	}
401 
402 	/*
403 	 * If the file is not open or is being closed then put the
404 	 * reference back.
405 	 */
406 	fp = atomic_load_consume(&ff->ff_file);
407 	if (__predict_true(fp != NULL)) {
408 		return fp;
409 	}
410 	fd_putfile(fd);
411 	return NULL;
412 }
413 
414 /*
415  * Release a reference to a file descriptor acquired with fd_getfile().
416  */
417 void
418 fd_putfile(unsigned fd)
419 {
420 	filedesc_t *fdp;
421 	fdfile_t *ff;
422 	u_int u, v;
423 
424 	fdp = curlwp->l_fd;
425 	KASSERT(fd < atomic_load_consume(&fdp->fd_dt)->dt_nfiles);
426 	ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
427 
428 	KASSERT(ff != NULL);
429 	KASSERT((ff->ff_refcnt & FR_MASK) > 0);
430 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
431 
432 	if (fdp->fd_refcnt == 1) {
433 		/*
434 		 * Single threaded: don't need to worry about concurrent
435 		 * access (other than earlier calls to kqueue, which may
436 		 * hold a reference to the descriptor).
437 		 */
438 		if (__predict_false((ff->ff_refcnt & FR_CLOSING) != 0)) {
439 			fd_close(fd);
440 			return;
441 		}
442 		ff->ff_refcnt--;
443 		return;
444 	}
445 
446 	/*
447 	 * Ensure that any use of the file is complete and globally
448 	 * visible before dropping the final reference.  If no membar,
449 	 * the current CPU could still access memory associated with
450 	 * the file after it has been freed or recycled by another
451 	 * CPU.
452 	 */
453 #ifndef __HAVE_ATOMIC_AS_MEMBAR
454 	membar_exit();
455 #endif
456 
457 	/*
458 	 * Be optimistic and start out with the assumption that no other
459 	 * threads are trying to close the descriptor.  If the CAS fails,
460 	 * we lost a race and/or it's being closed.
461 	 */
462 	for (u = ff->ff_refcnt & FR_MASK;; u = v) {
463 		v = atomic_cas_uint(&ff->ff_refcnt, u, u - 1);
464 		if (__predict_true(u == v)) {
465 			return;
466 		}
467 		if (__predict_false((v & FR_CLOSING) != 0)) {
468 			break;
469 		}
470 	}
471 
472 	/* Another thread is waiting to close the file: join it. */
473 	(void)fd_close(fd);
474 }
475 
476 /*
477  * Convenience wrapper around fd_getfile() that returns reference
478  * to a vnode.
479  */
480 int
481 fd_getvnode(unsigned fd, file_t **fpp)
482 {
483 	vnode_t *vp;
484 	file_t *fp;
485 
486 	fp = fd_getfile(fd);
487 	if (__predict_false(fp == NULL)) {
488 		return EBADF;
489 	}
490 	if (__predict_false(fp->f_type != DTYPE_VNODE)) {
491 		fd_putfile(fd);
492 		return EINVAL;
493 	}
494 	vp = fp->f_vnode;
495 	if (__predict_false(vp->v_type == VBAD)) {
496 		/* XXX Is this case really necessary? */
497 		fd_putfile(fd);
498 		return EBADF;
499 	}
500 	*fpp = fp;
501 	return 0;
502 }
503 
504 /*
505  * Convenience wrapper around fd_getfile() that returns reference
506  * to a socket.
507  */
508 int
509 fd_getsock1(unsigned fd, struct socket **sop, file_t **fp)
510 {
511 	*fp = fd_getfile(fd);
512 	if (__predict_false(*fp == NULL)) {
513 		return EBADF;
514 	}
515 	if (__predict_false((*fp)->f_type != DTYPE_SOCKET)) {
516 		fd_putfile(fd);
517 		return ENOTSOCK;
518 	}
519 	*sop = (*fp)->f_socket;
520 	return 0;
521 }
522 
523 int
524 fd_getsock(unsigned fd, struct socket **sop)
525 {
526 	file_t *fp;
527 	return fd_getsock1(fd, sop, &fp);
528 }
529 
530 /*
531  * Look up the file structure corresponding to a file descriptor
532  * and return it with a reference held on the file, not the
533  * descriptor.
534  *
535  * This is heavyweight and only used when accessing descriptors
536  * from a foreign process.  The caller must ensure that `p' does
537  * not exit or fork across this call.
538  *
539  * To release the file (not descriptor) reference, use closef().
540  */
541 file_t *
542 fd_getfile2(proc_t *p, unsigned fd)
543 {
544 	filedesc_t *fdp;
545 	fdfile_t *ff;
546 	file_t *fp;
547 	fdtab_t *dt;
548 
549 	fdp = p->p_fd;
550 	mutex_enter(&fdp->fd_lock);
551 	dt = fdp->fd_dt;
552 	if (fd >= dt->dt_nfiles) {
553 		mutex_exit(&fdp->fd_lock);
554 		return NULL;
555 	}
556 	if ((ff = dt->dt_ff[fd]) == NULL) {
557 		mutex_exit(&fdp->fd_lock);
558 		return NULL;
559 	}
560 	if ((fp = atomic_load_consume(&ff->ff_file)) == NULL) {
561 		mutex_exit(&fdp->fd_lock);
562 		return NULL;
563 	}
564 	mutex_enter(&fp->f_lock);
565 	fp->f_count++;
566 	mutex_exit(&fp->f_lock);
567 	mutex_exit(&fdp->fd_lock);
568 
569 	return fp;
570 }
571 
572 /*
573  * Internal form of close.  Must be called with a reference to the
574  * descriptor, and will drop the reference.  When all descriptor
575  * references are dropped, releases the descriptor slot and a single
576  * reference to the file structure.
577  */
578 int
579 fd_close(unsigned fd)
580 {
581 	struct flock lf;
582 	filedesc_t *fdp;
583 	fdfile_t *ff;
584 	file_t *fp;
585 	proc_t *p;
586 	lwp_t *l;
587 	u_int refcnt;
588 
589 	l = curlwp;
590 	p = l->l_proc;
591 	fdp = l->l_fd;
592 	ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
593 
594 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
595 
596 	mutex_enter(&fdp->fd_lock);
597 	KASSERT((ff->ff_refcnt & FR_MASK) > 0);
598 	fp = atomic_load_consume(&ff->ff_file);
599 	if (__predict_false(fp == NULL)) {
600 		/*
601 		 * Another user of the file is already closing, and is
602 		 * waiting for other users of the file to drain.  Release
603 		 * our reference, and wake up the closer.
604 		 */
605 		atomic_dec_uint(&ff->ff_refcnt);
606 		cv_broadcast(&ff->ff_closing);
607 		mutex_exit(&fdp->fd_lock);
608 
609 		/*
610 		 * An application error, so pretend that the descriptor
611 		 * was already closed.  We can't safely wait for it to
612 		 * be closed without potentially deadlocking.
613 		 */
614 		return (EBADF);
615 	}
616 	KASSERT((ff->ff_refcnt & FR_CLOSING) == 0);
617 
618 	/*
619 	 * There may be multiple users of this file within the process.
620 	 * Notify existing and new users that the file is closing.  This
621 	 * will prevent them from adding additional uses to this file
622 	 * while we are closing it.
623 	 */
624 	ff->ff_file = NULL;
625 	ff->ff_exclose = false;
626 
627 	/*
628 	 * We expect the caller to hold a descriptor reference - drop it.
629 	 * The reference count may increase beyond zero at this point due
630 	 * to an erroneous descriptor reference by an application, but
631 	 * fd_getfile() will notice that the file is being closed and drop
632 	 * the reference again.
633 	 */
634 	if (fdp->fd_refcnt == 1) {
635 		/* Single threaded. */
636 		refcnt = --(ff->ff_refcnt);
637 	} else {
638 		/* Multi threaded. */
639 #ifndef __HAVE_ATOMIC_AS_MEMBAR
640 		membar_producer();
641 #endif
642 		refcnt = atomic_dec_uint_nv(&ff->ff_refcnt);
643 	}
644 	if (__predict_false(refcnt != 0)) {
645 		/*
646 		 * Wait for other references to drain.  This is typically
647 		 * an application error - the descriptor is being closed
648 		 * while still in use.
649 		 * (Or just a threaded application trying to unblock its
650 		 * thread that sleeps in (say) accept()).
651 		 */
652 		atomic_or_uint(&ff->ff_refcnt, FR_CLOSING);
653 
654 		/*
655 		 * Remove any knotes attached to the file.  A knote
656 		 * attached to the descriptor can hold references on it.
657 		 */
658 		mutex_exit(&fdp->fd_lock);
659 		if (!SLIST_EMPTY(&ff->ff_knlist)) {
660 			knote_fdclose(fd);
661 		}
662 
663 		/*
664 		 * Since the file system code doesn't know which fd
665 		 * each request came from (think dup()), we have to
666 		 * ask it to return ERESTART for any long-term blocks.
667 		 * The re-entry through read/write/etc will detect the
668 		 * closed fd and return EBAFD.
669 		 * Blocked partial writes may return a short length.
670 		 */
671 		(*fp->f_ops->fo_restart)(fp);
672 		mutex_enter(&fdp->fd_lock);
673 
674 		/*
675 		 * We need to see the count drop to zero at least once,
676 		 * in order to ensure that all pre-existing references
677 		 * have been drained.  New references past this point are
678 		 * of no interest.
679 		 * XXX (dsl) this may need to call fo_restart() after a
680 		 * timeout to guarantee that all the system calls exit.
681 		 */
682 		while ((ff->ff_refcnt & FR_MASK) != 0) {
683 			cv_wait(&ff->ff_closing, &fdp->fd_lock);
684 		}
685 		atomic_and_uint(&ff->ff_refcnt, ~FR_CLOSING);
686 	} else {
687 		/* If no references, there must be no knotes. */
688 		KASSERT(SLIST_EMPTY(&ff->ff_knlist));
689 	}
690 
691 	/*
692 	 * POSIX record locking dictates that any close releases ALL
693 	 * locks owned by this process.  This is handled by setting
694 	 * a flag in the unlock to free ONLY locks obeying POSIX
695 	 * semantics, and not to free BSD-style file locks.
696 	 * If the descriptor was in a message, POSIX-style locks
697 	 * aren't passed with the descriptor.
698 	 */
699 	if (__predict_false((p->p_flag & PK_ADVLOCK) != 0 &&
700 	    fp->f_type == DTYPE_VNODE)) {
701 		lf.l_whence = SEEK_SET;
702 		lf.l_start = 0;
703 		lf.l_len = 0;
704 		lf.l_type = F_UNLCK;
705 		mutex_exit(&fdp->fd_lock);
706 		(void)VOP_ADVLOCK(fp->f_vnode, p, F_UNLCK, &lf, F_POSIX);
707 		mutex_enter(&fdp->fd_lock);
708 	}
709 
710 	/* Free descriptor slot. */
711 	fd_unused(fdp, fd);
712 	mutex_exit(&fdp->fd_lock);
713 
714 	/* Now drop reference to the file itself. */
715 	return closef(fp);
716 }
717 
718 /*
719  * Duplicate a file descriptor.
720  */
721 int
722 fd_dup(file_t *fp, int minfd, int *newp, bool exclose)
723 {
724 	proc_t *p = curproc;
725 	fdtab_t *dt;
726 	int error;
727 
728 	while ((error = fd_alloc(p, minfd, newp)) != 0) {
729 		if (error != ENOSPC) {
730 			return error;
731 		}
732 		fd_tryexpand(p);
733 	}
734 
735 	dt = atomic_load_consume(&curlwp->l_fd->fd_dt);
736 	dt->dt_ff[*newp]->ff_exclose = exclose;
737 	fd_affix(p, fp, *newp);
738 	return 0;
739 }
740 
741 /*
742  * dup2 operation.
743  */
744 int
745 fd_dup2(file_t *fp, unsigned newfd, int flags)
746 {
747 	filedesc_t *fdp = curlwp->l_fd;
748 	fdfile_t *ff;
749 	fdtab_t *dt;
750 
751 	if (flags & ~(O_CLOEXEC|O_NONBLOCK|O_NOSIGPIPE))
752 		return EINVAL;
753 	/*
754 	 * Ensure there are enough slots in the descriptor table,
755 	 * and allocate an fdfile_t up front in case we need it.
756 	 */
757 	while (newfd >= atomic_load_consume(&fdp->fd_dt)->dt_nfiles) {
758 		fd_tryexpand(curproc);
759 	}
760 	ff = pool_cache_get(fdfile_cache, PR_WAITOK);
761 
762 	/*
763 	 * If there is already a file open, close it.  If the file is
764 	 * half open, wait for it to be constructed before closing it.
765 	 * XXX Potential for deadlock here?
766 	 */
767 	mutex_enter(&fdp->fd_lock);
768 	while (fd_isused(fdp, newfd)) {
769 		mutex_exit(&fdp->fd_lock);
770 		if (fd_getfile(newfd) != NULL) {
771 			(void)fd_close(newfd);
772 		} else {
773 			/*
774 			 * Crummy, but unlikely to happen.
775 			 * Can occur if we interrupt another
776 			 * thread while it is opening a file.
777 			 */
778 			kpause("dup2", false, 1, NULL);
779 		}
780 		mutex_enter(&fdp->fd_lock);
781 	}
782 	dt = fdp->fd_dt;
783 	if (dt->dt_ff[newfd] == NULL) {
784 		KASSERT(newfd >= NDFDFILE);
785 		dt->dt_ff[newfd] = ff;
786 		ff = NULL;
787 	}
788 	fd_used(fdp, newfd);
789 	mutex_exit(&fdp->fd_lock);
790 
791 	dt->dt_ff[newfd]->ff_exclose = (flags & O_CLOEXEC) != 0;
792 	fp->f_flag |= flags & (FNONBLOCK|FNOSIGPIPE);
793 	/* Slot is now allocated.  Insert copy of the file. */
794 	fd_affix(curproc, fp, newfd);
795 	if (ff != NULL) {
796 		pool_cache_put(fdfile_cache, ff);
797 	}
798 	return 0;
799 }
800 
801 /*
802  * Drop reference to a file structure.
803  */
804 int
805 closef(file_t *fp)
806 {
807 	struct flock lf;
808 	int error;
809 
810 	/*
811 	 * Drop reference.  If referenced elsewhere it's still open
812 	 * and we have nothing more to do.
813 	 */
814 	mutex_enter(&fp->f_lock);
815 	KASSERT(fp->f_count > 0);
816 	if (--fp->f_count > 0) {
817 		mutex_exit(&fp->f_lock);
818 		return 0;
819 	}
820 	KASSERT(fp->f_count == 0);
821 	mutex_exit(&fp->f_lock);
822 
823 	/* We held the last reference - release locks, close and free. */
824 	if ((fp->f_flag & FHASLOCK) && fp->f_type == DTYPE_VNODE) {
825 		lf.l_whence = SEEK_SET;
826 		lf.l_start = 0;
827 		lf.l_len = 0;
828 		lf.l_type = F_UNLCK;
829 		(void)VOP_ADVLOCK(fp->f_vnode, fp, F_UNLCK, &lf, F_FLOCK);
830 	}
831 	if (fp->f_ops != NULL) {
832 		error = (*fp->f_ops->fo_close)(fp);
833 	} else {
834 		error = 0;
835 	}
836 	KASSERT(fp->f_count == 0);
837 	KASSERT(fp->f_cred != NULL);
838 	pool_cache_put(file_cache, fp);
839 
840 	return error;
841 }
842 
843 /*
844  * Allocate a file descriptor for the process.
845  */
846 int
847 fd_alloc(proc_t *p, int want, int *result)
848 {
849 	filedesc_t *fdp = p->p_fd;
850 	int i, lim, last, error, hi;
851 	u_int off;
852 	fdtab_t *dt;
853 
854 	KASSERT(p == curproc || p == &proc0);
855 
856 	/*
857 	 * Search for a free descriptor starting at the higher
858 	 * of want or fd_freefile.
859 	 */
860 	mutex_enter(&fdp->fd_lock);
861 	fd_checkmaps(fdp);
862 	dt = fdp->fd_dt;
863 	KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
864 	lim = uimin((int)p->p_rlimit[RLIMIT_NOFILE].rlim_cur, maxfiles);
865 	last = uimin(dt->dt_nfiles, lim);
866 	for (;;) {
867 		if ((i = want) < fdp->fd_freefile)
868 			i = fdp->fd_freefile;
869 		off = i >> NDENTRYSHIFT;
870 		hi = fd_next_zero(fdp, fdp->fd_himap, off,
871 		    (last + NDENTRIES - 1) >> NDENTRYSHIFT);
872 		if (hi == -1)
873 			break;
874 		i = fd_next_zero(fdp, &fdp->fd_lomap[hi],
875 		    hi > off ? 0 : i & NDENTRYMASK, NDENTRIES);
876 		if (i == -1) {
877 			/*
878 			 * Free file descriptor in this block was
879 			 * below want, try again with higher want.
880 			 */
881 			want = (hi + 1) << NDENTRYSHIFT;
882 			continue;
883 		}
884 		i += (hi << NDENTRYSHIFT);
885 		if (i >= last) {
886 			break;
887 		}
888 		if (dt->dt_ff[i] == NULL) {
889 			KASSERT(i >= NDFDFILE);
890 			dt->dt_ff[i] = pool_cache_get(fdfile_cache, PR_WAITOK);
891 		}
892 		KASSERT(dt->dt_ff[i]->ff_file == NULL);
893 		fd_used(fdp, i);
894 		if (want <= fdp->fd_freefile) {
895 			fdp->fd_freefile = i;
896 		}
897 		*result = i;
898 		KASSERT(i >= NDFDFILE ||
899 		    dt->dt_ff[i] == (fdfile_t *)fdp->fd_dfdfile[i]);
900 		fd_checkmaps(fdp);
901 		mutex_exit(&fdp->fd_lock);
902 		return 0;
903 	}
904 
905 	/* No space in current array.  Let the caller expand and retry. */
906 	error = (dt->dt_nfiles >= lim) ? EMFILE : ENOSPC;
907 	mutex_exit(&fdp->fd_lock);
908 	return error;
909 }
910 
911 /*
912  * Allocate memory for a descriptor table.
913  */
914 static fdtab_t *
915 fd_dtab_alloc(int n)
916 {
917 	fdtab_t *dt;
918 	size_t sz;
919 
920 	KASSERT(n > NDFILE);
921 
922 	sz = sizeof(*dt) + (n - NDFILE) * sizeof(dt->dt_ff[0]);
923 	dt = kmem_alloc(sz, KM_SLEEP);
924 #ifdef DIAGNOSTIC
925 	memset(dt, 0xff, sz);
926 #endif
927 	dt->dt_nfiles = n;
928 	dt->dt_link = NULL;
929 	return dt;
930 }
931 
932 /*
933  * Free a descriptor table, and all tables linked for deferred free.
934  */
935 static void
936 fd_dtab_free(fdtab_t *dt)
937 {
938 	fdtab_t *next;
939 	size_t sz;
940 
941 	do {
942 		next = dt->dt_link;
943 		KASSERT(dt->dt_nfiles > NDFILE);
944 		sz = sizeof(*dt) +
945 		    (dt->dt_nfiles - NDFILE) * sizeof(dt->dt_ff[0]);
946 #ifdef DIAGNOSTIC
947 		memset(dt, 0xff, sz);
948 #endif
949 		kmem_free(dt, sz);
950 		dt = next;
951 	} while (dt != NULL);
952 }
953 
954 /*
955  * Allocate descriptor bitmap.
956  */
957 static void
958 fd_map_alloc(int n, uint32_t **lo, uint32_t **hi)
959 {
960 	uint8_t *ptr;
961 	size_t szlo, szhi;
962 
963 	KASSERT(n > NDENTRIES);
964 
965 	szlo = NDLOSLOTS(n) * sizeof(uint32_t);
966 	szhi = NDHISLOTS(n) * sizeof(uint32_t);
967 	ptr = kmem_alloc(szlo + szhi, KM_SLEEP);
968 	*lo = (uint32_t *)ptr;
969 	*hi = (uint32_t *)(ptr + szlo);
970 }
971 
972 /*
973  * Free descriptor bitmap.
974  */
975 static void
976 fd_map_free(int n, uint32_t *lo, uint32_t *hi)
977 {
978 	size_t szlo, szhi;
979 
980 	KASSERT(n > NDENTRIES);
981 
982 	szlo = NDLOSLOTS(n) * sizeof(uint32_t);
983 	szhi = NDHISLOTS(n) * sizeof(uint32_t);
984 	KASSERT(hi == (uint32_t *)((uint8_t *)lo + szlo));
985 	kmem_free(lo, szlo + szhi);
986 }
987 
988 /*
989  * Expand a process' descriptor table.
990  */
991 void
992 fd_tryexpand(proc_t *p)
993 {
994 	filedesc_t *fdp;
995 	int i, numfiles, oldnfiles;
996 	fdtab_t *newdt, *dt;
997 	uint32_t *newhimap, *newlomap;
998 
999 	KASSERT(p == curproc || p == &proc0);
1000 
1001 	fdp = p->p_fd;
1002 	newhimap = NULL;
1003 	newlomap = NULL;
1004 	oldnfiles = atomic_load_consume(&fdp->fd_dt)->dt_nfiles;
1005 
1006 	if (oldnfiles < NDEXTENT)
1007 		numfiles = NDEXTENT;
1008 	else
1009 		numfiles = 2 * oldnfiles;
1010 
1011 	newdt = fd_dtab_alloc(numfiles);
1012 	if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1013 		fd_map_alloc(numfiles, &newlomap, &newhimap);
1014 	}
1015 
1016 	mutex_enter(&fdp->fd_lock);
1017 	dt = fdp->fd_dt;
1018 	KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1019 	if (dt->dt_nfiles != oldnfiles) {
1020 		/* fdp changed; caller must retry */
1021 		mutex_exit(&fdp->fd_lock);
1022 		fd_dtab_free(newdt);
1023 		if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1024 			fd_map_free(numfiles, newlomap, newhimap);
1025 		}
1026 		return;
1027 	}
1028 
1029 	/* Copy the existing descriptor table and zero the new portion. */
1030 	i = sizeof(fdfile_t *) * oldnfiles;
1031 	memcpy(newdt->dt_ff, dt->dt_ff, i);
1032 	memset((uint8_t *)newdt->dt_ff + i, 0,
1033 	    numfiles * sizeof(fdfile_t *) - i);
1034 
1035 	/*
1036 	 * Link old descriptor array into list to be discarded.  We defer
1037 	 * freeing until the last reference to the descriptor table goes
1038 	 * away (usually process exit).  This allows us to do lockless
1039 	 * lookups in fd_getfile().
1040 	 */
1041 	if (oldnfiles > NDFILE) {
1042 		if (fdp->fd_refcnt > 1) {
1043 			newdt->dt_link = dt;
1044 		} else {
1045 			fd_dtab_free(dt);
1046 		}
1047 	}
1048 
1049 	if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1050 		i = NDHISLOTS(oldnfiles) * sizeof(uint32_t);
1051 		memcpy(newhimap, fdp->fd_himap, i);
1052 		memset((uint8_t *)newhimap + i, 0,
1053 		    NDHISLOTS(numfiles) * sizeof(uint32_t) - i);
1054 
1055 		i = NDLOSLOTS(oldnfiles) * sizeof(uint32_t);
1056 		memcpy(newlomap, fdp->fd_lomap, i);
1057 		memset((uint8_t *)newlomap + i, 0,
1058 		    NDLOSLOTS(numfiles) * sizeof(uint32_t) - i);
1059 
1060 		if (NDHISLOTS(oldnfiles) > NDHISLOTS(NDFILE)) {
1061 			fd_map_free(oldnfiles, fdp->fd_lomap, fdp->fd_himap);
1062 		}
1063 		fdp->fd_himap = newhimap;
1064 		fdp->fd_lomap = newlomap;
1065 	}
1066 
1067 	/*
1068 	 * All other modifications must become globally visible before
1069 	 * the change to fd_dt.  See fd_getfile().
1070 	 */
1071 	atomic_store_release(&fdp->fd_dt, newdt);
1072 	KASSERT(newdt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1073 	fd_checkmaps(fdp);
1074 	mutex_exit(&fdp->fd_lock);
1075 }
1076 
1077 /*
1078  * Create a new open file structure and allocate a file descriptor
1079  * for the current process.
1080  */
1081 int
1082 fd_allocfile(file_t **resultfp, int *resultfd)
1083 {
1084 	proc_t *p = curproc;
1085 	kauth_cred_t cred;
1086 	file_t *fp;
1087 	int error;
1088 
1089 	while ((error = fd_alloc(p, 0, resultfd)) != 0) {
1090 		if (error != ENOSPC) {
1091 			return error;
1092 		}
1093 		fd_tryexpand(p);
1094 	}
1095 
1096 	fp = pool_cache_get(file_cache, PR_WAITOK);
1097 	if (fp == NULL) {
1098 		fd_abort(p, NULL, *resultfd);
1099 		return ENFILE;
1100 	}
1101 	KASSERT(fp->f_count == 0);
1102 	KASSERT(fp->f_msgcount == 0);
1103 	KASSERT(fp->f_unpcount == 0);
1104 
1105 	/* Replace cached credentials if not what we need. */
1106 	cred = curlwp->l_cred;
1107 	if (__predict_false(cred != fp->f_cred)) {
1108 		kauth_cred_free(fp->f_cred);
1109 		kauth_cred_hold(cred);
1110 		fp->f_cred = cred;
1111 	}
1112 
1113 	/*
1114 	 * Don't allow recycled files to be scanned.
1115 	 * See uipc_usrreq.c.
1116 	 */
1117 	if (__predict_false((fp->f_flag & FSCAN) != 0)) {
1118 		mutex_enter(&fp->f_lock);
1119 		atomic_and_uint(&fp->f_flag, ~FSCAN);
1120 		mutex_exit(&fp->f_lock);
1121 	}
1122 
1123 	fp->f_advice = 0;
1124 	fp->f_offset = 0;
1125 	*resultfp = fp;
1126 
1127 	return 0;
1128 }
1129 
1130 /*
1131  * Successful creation of a new descriptor: make visible to the process.
1132  */
1133 void
1134 fd_affix(proc_t *p, file_t *fp, unsigned fd)
1135 {
1136 	fdfile_t *ff;
1137 	filedesc_t *fdp;
1138 	fdtab_t *dt;
1139 
1140 	KASSERT(p == curproc || p == &proc0);
1141 
1142 	/* Add a reference to the file structure. */
1143 	mutex_enter(&fp->f_lock);
1144 	fp->f_count++;
1145 	mutex_exit(&fp->f_lock);
1146 
1147 	/*
1148 	 * Insert the new file into the descriptor slot.
1149 	 *
1150 	 * The memory barriers provided by lock activity in this routine
1151 	 * ensure that any updates to the file structure become globally
1152 	 * visible before the file becomes visible to other LWPs in the
1153 	 * current process; otherwise we would set ff->ff_file with
1154 	 * atomic_store_release(&ff->ff_file, fp) at the bottom.
1155 	 */
1156 	fdp = p->p_fd;
1157 	dt = atomic_load_consume(&fdp->fd_dt);
1158 	ff = dt->dt_ff[fd];
1159 
1160 	KASSERT(ff != NULL);
1161 	KASSERT(ff->ff_file == NULL);
1162 	KASSERT(ff->ff_allocated);
1163 	KASSERT(fd_isused(fdp, fd));
1164 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1165 
1166 	/* No need to lock in order to make file initially visible. */
1167 	ff->ff_file = fp;
1168 }
1169 
1170 /*
1171  * Abort creation of a new descriptor: free descriptor slot and file.
1172  */
1173 void
1174 fd_abort(proc_t *p, file_t *fp, unsigned fd)
1175 {
1176 	filedesc_t *fdp;
1177 	fdfile_t *ff;
1178 
1179 	KASSERT(p == curproc || p == &proc0);
1180 
1181 	fdp = p->p_fd;
1182 	ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
1183 	ff->ff_exclose = false;
1184 
1185 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1186 
1187 	mutex_enter(&fdp->fd_lock);
1188 	KASSERT(fd_isused(fdp, fd));
1189 	fd_unused(fdp, fd);
1190 	mutex_exit(&fdp->fd_lock);
1191 
1192 	if (fp != NULL) {
1193 		KASSERT(fp->f_count == 0);
1194 		KASSERT(fp->f_cred != NULL);
1195 		pool_cache_put(file_cache, fp);
1196 	}
1197 }
1198 
1199 static int
1200 file_ctor(void *arg, void *obj, int flags)
1201 {
1202 	file_t *fp = obj;
1203 
1204 	memset(fp, 0, sizeof(*fp));
1205 
1206 	mutex_enter(&filelist_lock);
1207 	if (__predict_false(nfiles >= maxfiles)) {
1208 		mutex_exit(&filelist_lock);
1209 		tablefull("file", "increase kern.maxfiles or MAXFILES");
1210 		return ENFILE;
1211 	}
1212 	nfiles++;
1213 	LIST_INSERT_HEAD(&filehead, fp, f_list);
1214 	mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1215 	fp->f_cred = curlwp->l_cred;
1216 	kauth_cred_hold(fp->f_cred);
1217 	mutex_exit(&filelist_lock);
1218 
1219 	return 0;
1220 }
1221 
1222 static void
1223 file_dtor(void *arg, void *obj)
1224 {
1225 	file_t *fp = obj;
1226 
1227 	mutex_enter(&filelist_lock);
1228 	nfiles--;
1229 	LIST_REMOVE(fp, f_list);
1230 	mutex_exit(&filelist_lock);
1231 
1232 	KASSERT(fp->f_count == 0);
1233 	kauth_cred_free(fp->f_cred);
1234 	mutex_destroy(&fp->f_lock);
1235 }
1236 
1237 static int
1238 fdfile_ctor(void *arg, void *obj, int flags)
1239 {
1240 	fdfile_t *ff = obj;
1241 
1242 	memset(ff, 0, sizeof(*ff));
1243 	cv_init(&ff->ff_closing, "fdclose");
1244 
1245 	return 0;
1246 }
1247 
1248 static void
1249 fdfile_dtor(void *arg, void *obj)
1250 {
1251 	fdfile_t *ff = obj;
1252 
1253 	cv_destroy(&ff->ff_closing);
1254 }
1255 
1256 file_t *
1257 fgetdummy(void)
1258 {
1259 	file_t *fp;
1260 
1261 	fp = kmem_zalloc(sizeof(*fp), KM_SLEEP);
1262 	mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1263 	return fp;
1264 }
1265 
1266 void
1267 fputdummy(file_t *fp)
1268 {
1269 
1270 	mutex_destroy(&fp->f_lock);
1271 	kmem_free(fp, sizeof(*fp));
1272 }
1273 
1274 /*
1275  * Create an initial filedesc structure.
1276  */
1277 filedesc_t *
1278 fd_init(filedesc_t *fdp)
1279 {
1280 #ifdef DIAGNOSTIC
1281 	unsigned fd;
1282 #endif
1283 
1284 	if (__predict_true(fdp == NULL)) {
1285 		fdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1286 	} else {
1287 		KASSERT(fdp == &filedesc0);
1288 		filedesc_ctor(NULL, fdp, PR_WAITOK);
1289 	}
1290 
1291 #ifdef DIAGNOSTIC
1292 	KASSERT(fdp->fd_lastfile == -1);
1293 	KASSERT(fdp->fd_lastkqfile == -1);
1294 	KASSERT(fdp->fd_knhash == NULL);
1295 	KASSERT(fdp->fd_freefile == 0);
1296 	KASSERT(fdp->fd_exclose == false);
1297 	KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1298 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1299 	for (fd = 0; fd < NDFDFILE; fd++) {
1300 		KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] ==
1301 		    (fdfile_t *)fdp->fd_dfdfile[fd]);
1302 	}
1303 	for (fd = NDFDFILE; fd < NDFILE; fd++) {
1304 		KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] == NULL);
1305 	}
1306 	KASSERT(fdp->fd_himap == fdp->fd_dhimap);
1307 	KASSERT(fdp->fd_lomap == fdp->fd_dlomap);
1308 #endif	/* DIAGNOSTIC */
1309 
1310 	fdp->fd_refcnt = 1;
1311 	fd_checkmaps(fdp);
1312 
1313 	return fdp;
1314 }
1315 
1316 /*
1317  * Initialize a file descriptor table.
1318  */
1319 static int
1320 filedesc_ctor(void *arg, void *obj, int flag)
1321 {
1322 	filedesc_t *fdp = obj;
1323 	fdfile_t **ffp;
1324 	int i;
1325 
1326 	memset(fdp, 0, sizeof(*fdp));
1327 	mutex_init(&fdp->fd_lock, MUTEX_DEFAULT, IPL_NONE);
1328 	fdp->fd_lastfile = -1;
1329 	fdp->fd_lastkqfile = -1;
1330 	fdp->fd_dt = &fdp->fd_dtbuiltin;
1331 	fdp->fd_dtbuiltin.dt_nfiles = NDFILE;
1332 	fdp->fd_himap = fdp->fd_dhimap;
1333 	fdp->fd_lomap = fdp->fd_dlomap;
1334 
1335 	CTASSERT(sizeof(fdp->fd_dfdfile[0]) >= sizeof(fdfile_t));
1336 	for (i = 0, ffp = fdp->fd_dt->dt_ff; i < NDFDFILE; i++, ffp++) {
1337 		*ffp = (fdfile_t *)fdp->fd_dfdfile[i];
1338 		(void)fdfile_ctor(NULL, fdp->fd_dfdfile[i], PR_WAITOK);
1339 	}
1340 
1341 	return 0;
1342 }
1343 
1344 static void
1345 filedesc_dtor(void *arg, void *obj)
1346 {
1347 	filedesc_t *fdp = obj;
1348 	int i;
1349 
1350 	for (i = 0; i < NDFDFILE; i++) {
1351 		fdfile_dtor(NULL, fdp->fd_dfdfile[i]);
1352 	}
1353 
1354 	mutex_destroy(&fdp->fd_lock);
1355 }
1356 
1357 /*
1358  * Make p share curproc's filedesc structure.
1359  */
1360 void
1361 fd_share(struct proc *p)
1362 {
1363 	filedesc_t *fdp;
1364 
1365 	fdp = curlwp->l_fd;
1366 	p->p_fd = fdp;
1367 	atomic_inc_uint(&fdp->fd_refcnt);
1368 }
1369 
1370 /*
1371  * Acquire a hold on a filedesc structure.
1372  */
1373 void
1374 fd_hold(lwp_t *l)
1375 {
1376 	filedesc_t *fdp = l->l_fd;
1377 
1378 	atomic_inc_uint(&fdp->fd_refcnt);
1379 }
1380 
1381 /*
1382  * Copy a filedesc structure.
1383  */
1384 filedesc_t *
1385 fd_copy(void)
1386 {
1387 	filedesc_t *newfdp, *fdp;
1388 	fdfile_t *ff, **ffp, **nffp, *ff2;
1389 	int i, j, numfiles, lastfile, newlast;
1390 	file_t *fp;
1391 	fdtab_t *newdt;
1392 
1393 	fdp = curproc->p_fd;
1394 	newfdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1395 	newfdp->fd_refcnt = 1;
1396 
1397 #ifdef DIAGNOSTIC
1398 	KASSERT(newfdp->fd_lastfile == -1);
1399 	KASSERT(newfdp->fd_lastkqfile == -1);
1400 	KASSERT(newfdp->fd_knhash == NULL);
1401 	KASSERT(newfdp->fd_freefile == 0);
1402 	KASSERT(newfdp->fd_exclose == false);
1403 	KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1404 	KASSERT(newfdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1405 	for (i = 0; i < NDFDFILE; i++) {
1406 		KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] ==
1407 		    (fdfile_t *)&newfdp->fd_dfdfile[i]);
1408 	}
1409 	for (i = NDFDFILE; i < NDFILE; i++) {
1410 		KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] == NULL);
1411 	}
1412 #endif	/* DIAGNOSTIC */
1413 
1414 	mutex_enter(&fdp->fd_lock);
1415 	fd_checkmaps(fdp);
1416 	numfiles = fdp->fd_dt->dt_nfiles;
1417 	lastfile = fdp->fd_lastfile;
1418 
1419 	/*
1420 	 * If the number of open files fits in the internal arrays
1421 	 * of the open file structure, use them, otherwise allocate
1422 	 * additional memory for the number of descriptors currently
1423 	 * in use.
1424 	 */
1425 	if (lastfile < NDFILE) {
1426 		i = NDFILE;
1427 		newdt = newfdp->fd_dt;
1428 		KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1429 	} else {
1430 		/*
1431 		 * Compute the smallest multiple of NDEXTENT needed
1432 		 * for the file descriptors currently in use,
1433 		 * allowing the table to shrink.
1434 		 */
1435 		i = numfiles;
1436 		while (i >= 2 * NDEXTENT && i > lastfile * 2) {
1437 			i /= 2;
1438 		}
1439 		KASSERT(i > NDFILE);
1440 		newdt = fd_dtab_alloc(i);
1441 		newfdp->fd_dt = newdt;
1442 		memcpy(newdt->dt_ff, newfdp->fd_dtbuiltin.dt_ff,
1443 		    NDFDFILE * sizeof(fdfile_t **));
1444 		memset(newdt->dt_ff + NDFDFILE, 0,
1445 		    (i - NDFDFILE) * sizeof(fdfile_t **));
1446 	}
1447 	if (NDHISLOTS(i) <= NDHISLOTS(NDFILE)) {
1448 		newfdp->fd_himap = newfdp->fd_dhimap;
1449 		newfdp->fd_lomap = newfdp->fd_dlomap;
1450 	} else {
1451 		fd_map_alloc(i, &newfdp->fd_lomap, &newfdp->fd_himap);
1452 		KASSERT(i >= NDENTRIES * NDENTRIES);
1453 		memset(newfdp->fd_himap, 0, NDHISLOTS(i)*sizeof(uint32_t));
1454 		memset(newfdp->fd_lomap, 0, NDLOSLOTS(i)*sizeof(uint32_t));
1455 	}
1456 	newfdp->fd_freefile = fdp->fd_freefile;
1457 	newfdp->fd_exclose = fdp->fd_exclose;
1458 
1459 	ffp = fdp->fd_dt->dt_ff;
1460 	nffp = newdt->dt_ff;
1461 	newlast = -1;
1462 	for (i = 0; i <= lastfile; i++, ffp++, nffp++) {
1463 		KASSERT(i >= NDFDFILE ||
1464 		    *nffp == (fdfile_t *)newfdp->fd_dfdfile[i]);
1465 		ff = *ffp;
1466 		if (ff == NULL ||
1467 		    (fp = atomic_load_consume(&ff->ff_file)) == NULL) {
1468 			/* Descriptor unused, or descriptor half open. */
1469 			KASSERT(!fd_isused(newfdp, i));
1470 			continue;
1471 		}
1472 		if (__predict_false(fp->f_type == DTYPE_KQUEUE)) {
1473 			/* kqueue descriptors cannot be copied. */
1474 			if (i < newfdp->fd_freefile) {
1475 				newfdp->fd_freefile = i;
1476 			}
1477 			continue;
1478 		}
1479 		/* It's active: add a reference to the file. */
1480 		mutex_enter(&fp->f_lock);
1481 		fp->f_count++;
1482 		mutex_exit(&fp->f_lock);
1483 
1484 		/* Allocate an fdfile_t to represent it. */
1485 		if (i >= NDFDFILE) {
1486 			ff2 = pool_cache_get(fdfile_cache, PR_WAITOK);
1487 			*nffp = ff2;
1488 		} else {
1489 			ff2 = newdt->dt_ff[i];
1490 		}
1491 		ff2->ff_file = fp;
1492 		ff2->ff_exclose = ff->ff_exclose;
1493 		ff2->ff_allocated = true;
1494 
1495 		/* Fix up bitmaps. */
1496 		j = i >> NDENTRYSHIFT;
1497 		KASSERT((newfdp->fd_lomap[j] & (1U << (i & NDENTRYMASK))) == 0);
1498 		newfdp->fd_lomap[j] |= 1U << (i & NDENTRYMASK);
1499 		if (__predict_false(newfdp->fd_lomap[j] == ~0)) {
1500 			KASSERT((newfdp->fd_himap[j >> NDENTRYSHIFT] &
1501 			    (1U << (j & NDENTRYMASK))) == 0);
1502 			newfdp->fd_himap[j >> NDENTRYSHIFT] |=
1503 			    1U << (j & NDENTRYMASK);
1504 		}
1505 		newlast = i;
1506 	}
1507 	KASSERT(newdt->dt_ff[0] == (fdfile_t *)newfdp->fd_dfdfile[0]);
1508 	newfdp->fd_lastfile = newlast;
1509 	fd_checkmaps(newfdp);
1510 	mutex_exit(&fdp->fd_lock);
1511 
1512 	return newfdp;
1513 }
1514 
1515 /*
1516  * Release a filedesc structure.
1517  */
1518 void
1519 fd_free(void)
1520 {
1521 	fdfile_t *ff;
1522 	file_t *fp;
1523 	int fd, nf;
1524 	fdtab_t *dt;
1525 	lwp_t * const l = curlwp;
1526 	filedesc_t * const fdp = l->l_fd;
1527 	const bool noadvlock = (l->l_proc->p_flag & PK_ADVLOCK) == 0;
1528 
1529 	KASSERT(atomic_load_consume(&fdp->fd_dt)->dt_ff[0] ==
1530 	    (fdfile_t *)fdp->fd_dfdfile[0]);
1531 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1532 	KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1533 
1534 #ifndef __HAVE_ATOMIC_AS_MEMBAR
1535 	membar_exit();
1536 #endif
1537 	if (atomic_dec_uint_nv(&fdp->fd_refcnt) > 0)
1538 		return;
1539 
1540 	/*
1541 	 * Close any files that the process holds open.
1542 	 */
1543 	dt = fdp->fd_dt;
1544 	fd_checkmaps(fdp);
1545 #ifdef DEBUG
1546 	fdp->fd_refcnt = -1; /* see fd_checkmaps */
1547 #endif
1548 	for (fd = 0, nf = dt->dt_nfiles; fd < nf; fd++) {
1549 		ff = dt->dt_ff[fd];
1550 		KASSERT(fd >= NDFDFILE ||
1551 		    ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1552 		if (ff == NULL)
1553 			continue;
1554 		if ((fp = atomic_load_consume(&ff->ff_file)) != NULL) {
1555 			/*
1556 			 * Must use fd_close() here if there is
1557 			 * a reference from kqueue or we might have posix
1558 			 * advisory locks.
1559 			 */
1560 			if (__predict_true(ff->ff_refcnt == 0) &&
1561 			    (noadvlock || fp->f_type != DTYPE_VNODE)) {
1562 				ff->ff_file = NULL;
1563 				ff->ff_exclose = false;
1564 				ff->ff_allocated = false;
1565 				closef(fp);
1566 			} else {
1567 				ff->ff_refcnt++;
1568 				fd_close(fd);
1569 			}
1570 		}
1571 		KASSERT(ff->ff_refcnt == 0);
1572 		KASSERT(ff->ff_file == NULL);
1573 		KASSERT(!ff->ff_exclose);
1574 		KASSERT(!ff->ff_allocated);
1575 		if (fd >= NDFDFILE) {
1576 			pool_cache_put(fdfile_cache, ff);
1577 			dt->dt_ff[fd] = NULL;
1578 		}
1579 	}
1580 
1581 	/*
1582 	 * Clean out the descriptor table for the next user and return
1583 	 * to the cache.
1584 	 */
1585 	if (__predict_false(dt != &fdp->fd_dtbuiltin)) {
1586 		fd_dtab_free(fdp->fd_dt);
1587 		/* Otherwise, done above. */
1588 		memset(&fdp->fd_dtbuiltin.dt_ff[NDFDFILE], 0,
1589 		    (NDFILE - NDFDFILE) * sizeof(fdp->fd_dtbuiltin.dt_ff[0]));
1590 		fdp->fd_dt = &fdp->fd_dtbuiltin;
1591 	}
1592 	if (__predict_false(NDHISLOTS(nf) > NDHISLOTS(NDFILE))) {
1593 		KASSERT(fdp->fd_himap != fdp->fd_dhimap);
1594 		KASSERT(fdp->fd_lomap != fdp->fd_dlomap);
1595 		fd_map_free(nf, fdp->fd_lomap, fdp->fd_himap);
1596 	}
1597 	if (__predict_false(fdp->fd_knhash != NULL)) {
1598 		hashdone(fdp->fd_knhash, HASH_LIST, fdp->fd_knhashmask);
1599 		fdp->fd_knhash = NULL;
1600 		fdp->fd_knhashmask = 0;
1601 	} else {
1602 		KASSERT(fdp->fd_knhashmask == 0);
1603 	}
1604 	fdp->fd_dt = &fdp->fd_dtbuiltin;
1605 	fdp->fd_lastkqfile = -1;
1606 	fdp->fd_lastfile = -1;
1607 	fdp->fd_freefile = 0;
1608 	fdp->fd_exclose = false;
1609 	memset(&fdp->fd_startzero, 0, sizeof(*fdp) -
1610 	    offsetof(filedesc_t, fd_startzero));
1611 	fdp->fd_himap = fdp->fd_dhimap;
1612 	fdp->fd_lomap = fdp->fd_dlomap;
1613 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1614 	KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1615 	KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1616 #ifdef DEBUG
1617 	fdp->fd_refcnt = 0; /* see fd_checkmaps */
1618 #endif
1619 	fd_checkmaps(fdp);
1620 	pool_cache_put(filedesc_cache, fdp);
1621 }
1622 
1623 /*
1624  * File Descriptor pseudo-device driver (/dev/fd/).
1625  *
1626  * Opening minor device N dup()s the file (if any) connected to file
1627  * descriptor N belonging to the calling process.  Note that this driver
1628  * consists of only the ``open()'' routine, because all subsequent
1629  * references to this file will be direct to the other driver.
1630  */
1631 static int
1632 filedescopen(dev_t dev, int mode, int type, lwp_t *l)
1633 {
1634 
1635 	/*
1636 	 * XXX Kludge: set dupfd to contain the value of the
1637 	 * the file descriptor being sought for duplication. The error
1638 	 * return ensures that the vnode for this device will be released
1639 	 * by vn_open. Open will detect this special error and take the
1640 	 * actions in fd_dupopen below. Other callers of vn_open or VOP_OPEN
1641 	 * will simply report the error.
1642 	 */
1643 	l->l_dupfd = minor(dev);	/* XXX */
1644 	return EDUPFD;
1645 }
1646 
1647 /*
1648  * Duplicate the specified descriptor to a free descriptor.
1649  */
1650 int
1651 fd_dupopen(int old, int *newp, int mode, int error)
1652 {
1653 	filedesc_t *fdp;
1654 	fdfile_t *ff;
1655 	file_t *fp;
1656 	fdtab_t *dt;
1657 
1658 	if ((fp = fd_getfile(old)) == NULL) {
1659 		return EBADF;
1660 	}
1661 	fdp = curlwp->l_fd;
1662 	dt = atomic_load_consume(&fdp->fd_dt);
1663 	ff = dt->dt_ff[old];
1664 
1665 	/*
1666 	 * There are two cases of interest here.
1667 	 *
1668 	 * For EDUPFD simply dup (old) to file descriptor
1669 	 * (new) and return.
1670 	 *
1671 	 * For EMOVEFD steal away the file structure from (old) and
1672 	 * store it in (new).  (old) is effectively closed by
1673 	 * this operation.
1674 	 *
1675 	 * Any other error code is just returned.
1676 	 */
1677 	switch (error) {
1678 	case EDUPFD:
1679 		/*
1680 		 * Check that the mode the file is being opened for is a
1681 		 * subset of the mode of the existing descriptor.
1682 		 */
1683 		if (((mode & (FREAD|FWRITE)) | fp->f_flag) != fp->f_flag) {
1684 			error = EACCES;
1685 			break;
1686 		}
1687 
1688 		/* Copy it. */
1689 		error = fd_dup(fp, 0, newp, ff->ff_exclose);
1690 		break;
1691 
1692 	case EMOVEFD:
1693 		/* Copy it. */
1694 		error = fd_dup(fp, 0, newp, ff->ff_exclose);
1695 		if (error != 0) {
1696 			break;
1697 		}
1698 
1699 		/* Steal away the file pointer from 'old'. */
1700 		(void)fd_close(old);
1701 		return 0;
1702 	}
1703 
1704 	fd_putfile(old);
1705 	return error;
1706 }
1707 
1708 /*
1709  * Close open files on exec.
1710  */
1711 void
1712 fd_closeexec(void)
1713 {
1714 	proc_t *p;
1715 	filedesc_t *fdp;
1716 	fdfile_t *ff;
1717 	lwp_t *l;
1718 	fdtab_t *dt;
1719 	int fd;
1720 
1721 	l = curlwp;
1722 	p = l->l_proc;
1723 	fdp = p->p_fd;
1724 
1725 	if (fdp->fd_refcnt > 1) {
1726 		fdp = fd_copy();
1727 		fd_free();
1728 		p->p_fd = fdp;
1729 		l->l_fd = fdp;
1730 	}
1731 	if (!fdp->fd_exclose) {
1732 		return;
1733 	}
1734 	fdp->fd_exclose = false;
1735 	dt = atomic_load_consume(&fdp->fd_dt);
1736 
1737 	for (fd = 0; fd <= fdp->fd_lastfile; fd++) {
1738 		if ((ff = dt->dt_ff[fd]) == NULL) {
1739 			KASSERT(fd >= NDFDFILE);
1740 			continue;
1741 		}
1742 		KASSERT(fd >= NDFDFILE ||
1743 		    ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1744 		if (ff->ff_file == NULL)
1745 			continue;
1746 		if (ff->ff_exclose) {
1747 			/*
1748 			 * We need a reference to close the file.
1749 			 * No other threads can see the fdfile_t at
1750 			 * this point, so don't bother locking.
1751 			 */
1752 			KASSERT((ff->ff_refcnt & FR_CLOSING) == 0);
1753 			ff->ff_refcnt++;
1754 			fd_close(fd);
1755 		}
1756 	}
1757 }
1758 
1759 /*
1760  * Sets descriptor owner. If the owner is a process, 'pgid'
1761  * is set to positive value, process ID. If the owner is process group,
1762  * 'pgid' is set to -pg_id.
1763  */
1764 int
1765 fsetown(pid_t *pgid, u_long cmd, const void *data)
1766 {
1767 	pid_t id = *(const pid_t *)data;
1768 	int error;
1769 
1770 	switch (cmd) {
1771 	case TIOCSPGRP:
1772 		if (id < 0)
1773 			return EINVAL;
1774 		id = -id;
1775 		break;
1776 	default:
1777 		break;
1778 	}
1779 	if (id > 0) {
1780 		mutex_enter(&proc_lock);
1781 		error = proc_find(id) ? 0 : ESRCH;
1782 		mutex_exit(&proc_lock);
1783 	} else if (id < 0) {
1784 		error = pgid_in_session(curproc, -id);
1785 	} else {
1786 		error = 0;
1787 	}
1788 	if (!error) {
1789 		*pgid = id;
1790 	}
1791 	return error;
1792 }
1793 
1794 void
1795 fd_set_exclose(struct lwp *l, int fd, bool exclose)
1796 {
1797 	filedesc_t *fdp = l->l_fd;
1798 	fdfile_t *ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
1799 
1800 	ff->ff_exclose = exclose;
1801 	if (exclose)
1802 		fdp->fd_exclose = true;
1803 }
1804 
1805 /*
1806  * Return descriptor owner information. If the value is positive,
1807  * it's process ID. If it's negative, it's process group ID and
1808  * needs the sign removed before use.
1809  */
1810 int
1811 fgetown(pid_t pgid, u_long cmd, void *data)
1812 {
1813 
1814 	switch (cmd) {
1815 	case TIOCGPGRP:
1816 		*(int *)data = -pgid;
1817 		break;
1818 	default:
1819 		*(int *)data = pgid;
1820 		break;
1821 	}
1822 	return 0;
1823 }
1824 
1825 /*
1826  * Send signal to descriptor owner, either process or process group.
1827  */
1828 void
1829 fownsignal(pid_t pgid, int signo, int code, int band, void *fdescdata)
1830 {
1831 	ksiginfo_t ksi;
1832 
1833 	KASSERT(!cpu_intr_p());
1834 
1835 	if (pgid == 0) {
1836 		return;
1837 	}
1838 
1839 	KSI_INIT(&ksi);
1840 	ksi.ksi_signo = signo;
1841 	ksi.ksi_code = code;
1842 	ksi.ksi_band = band;
1843 
1844 	mutex_enter(&proc_lock);
1845 	if (pgid > 0) {
1846 		struct proc *p1;
1847 
1848 		p1 = proc_find(pgid);
1849 		if (p1 != NULL) {
1850 			kpsignal(p1, &ksi, fdescdata);
1851 		}
1852 	} else {
1853 		struct pgrp *pgrp;
1854 
1855 		KASSERT(pgid < 0);
1856 		pgrp = pgrp_find(-pgid);
1857 		if (pgrp != NULL) {
1858 			kpgsignal(pgrp, &ksi, fdescdata, 0);
1859 		}
1860 	}
1861 	mutex_exit(&proc_lock);
1862 }
1863 
1864 int
1865 fd_clone(file_t *fp, unsigned fd, int flag, const struct fileops *fops,
1866 	 void *data)
1867 {
1868 	fdfile_t *ff;
1869 	filedesc_t *fdp;
1870 
1871 	fp->f_flag = flag & FMASK;
1872 	fdp = curproc->p_fd;
1873 	ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
1874 	KASSERT(ff != NULL);
1875 	ff->ff_exclose = (flag & O_CLOEXEC) != 0;
1876 	fp->f_type = DTYPE_MISC;
1877 	fp->f_ops = fops;
1878 	fp->f_data = data;
1879 	curlwp->l_dupfd = fd;
1880 	fd_affix(curproc, fp, fd);
1881 
1882 	return EMOVEFD;
1883 }
1884 
1885 int
1886 fnullop_fcntl(file_t *fp, u_int cmd, void *data)
1887 {
1888 
1889 	if (cmd == F_SETFL)
1890 		return 0;
1891 
1892 	return EOPNOTSUPP;
1893 }
1894 
1895 int
1896 fnullop_poll(file_t *fp, int which)
1897 {
1898 
1899 	return 0;
1900 }
1901 
1902 int
1903 fnullop_kqfilter(file_t *fp, struct knote *kn)
1904 {
1905 
1906 	return EOPNOTSUPP;
1907 }
1908 
1909 void
1910 fnullop_restart(file_t *fp)
1911 {
1912 
1913 }
1914 
1915 int
1916 fbadop_read(file_t *fp, off_t *offset, struct uio *uio,
1917 	    kauth_cred_t cred, int flags)
1918 {
1919 
1920 	return EOPNOTSUPP;
1921 }
1922 
1923 int
1924 fbadop_write(file_t *fp, off_t *offset, struct uio *uio,
1925 	     kauth_cred_t cred, int flags)
1926 {
1927 
1928 	return EOPNOTSUPP;
1929 }
1930 
1931 int
1932 fbadop_ioctl(file_t *fp, u_long com, void *data)
1933 {
1934 
1935 	return EOPNOTSUPP;
1936 }
1937 
1938 int
1939 fbadop_stat(file_t *fp, struct stat *sb)
1940 {
1941 
1942 	return EOPNOTSUPP;
1943 }
1944 
1945 int
1946 fbadop_close(file_t *fp)
1947 {
1948 
1949 	return EOPNOTSUPP;
1950 }
1951 
1952 /*
1953  * sysctl routines pertaining to file descriptors
1954  */
1955 
1956 /* Initialized in sysctl_init() for now... */
1957 extern kmutex_t sysctl_file_marker_lock;
1958 static u_int sysctl_file_marker = 1;
1959 
1960 /*
1961  * Expects to be called with proc_lock and sysctl_file_marker_lock locked.
1962  */
1963 static void
1964 sysctl_file_marker_reset(void)
1965 {
1966 	struct proc *p;
1967 
1968 	PROCLIST_FOREACH(p, &allproc) {
1969 		struct filedesc *fd = p->p_fd;
1970 		fdtab_t *dt;
1971 		u_int i;
1972 
1973 		mutex_enter(&fd->fd_lock);
1974 		dt = fd->fd_dt;
1975 		for (i = 0; i < dt->dt_nfiles; i++) {
1976 			struct file *fp;
1977 			fdfile_t *ff;
1978 
1979 			if ((ff = dt->dt_ff[i]) == NULL) {
1980 				continue;
1981 			}
1982 			if ((fp = atomic_load_consume(&ff->ff_file)) == NULL) {
1983 				continue;
1984 			}
1985 			fp->f_marker = 0;
1986 		}
1987 		mutex_exit(&fd->fd_lock);
1988 	}
1989 }
1990 
1991 /*
1992  * sysctl helper routine for kern.file pseudo-subtree.
1993  */
1994 static int
1995 sysctl_kern_file(SYSCTLFN_ARGS)
1996 {
1997 	const bool allowaddr = get_expose_address(curproc);
1998 	struct filelist flist;
1999 	int error;
2000 	size_t buflen;
2001 	struct file *fp, fbuf;
2002 	char *start, *where;
2003 	struct proc *p;
2004 
2005 	start = where = oldp;
2006 	buflen = *oldlenp;
2007 
2008 	if (where == NULL) {
2009 		/*
2010 		 * overestimate by 10 files
2011 		 */
2012 		*oldlenp = sizeof(filehead) + (nfiles + 10) *
2013 		    sizeof(struct file);
2014 		return 0;
2015 	}
2016 
2017 	/*
2018 	 * first sysctl_copyout filehead
2019 	 */
2020 	if (buflen < sizeof(filehead)) {
2021 		*oldlenp = 0;
2022 		return 0;
2023 	}
2024 	sysctl_unlock();
2025 	if (allowaddr) {
2026 		memcpy(&flist, &filehead, sizeof(flist));
2027 	} else {
2028 		memset(&flist, 0, sizeof(flist));
2029 	}
2030 	error = sysctl_copyout(l, &flist, where, sizeof(flist));
2031 	if (error) {
2032 		sysctl_relock();
2033 		return error;
2034 	}
2035 	buflen -= sizeof(flist);
2036 	where += sizeof(flist);
2037 
2038 	/*
2039 	 * followed by an array of file structures
2040 	 */
2041 	mutex_enter(&sysctl_file_marker_lock);
2042 	mutex_enter(&proc_lock);
2043 	PROCLIST_FOREACH(p, &allproc) {
2044 		struct filedesc *fd;
2045 		fdtab_t *dt;
2046 		u_int i;
2047 
2048 		if (p->p_stat == SIDL) {
2049 			/* skip embryonic processes */
2050 			continue;
2051 		}
2052 		mutex_enter(p->p_lock);
2053 		error = kauth_authorize_process(l->l_cred,
2054 		    KAUTH_PROCESS_CANSEE, p,
2055 		    KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2056 		    NULL, NULL);
2057 		mutex_exit(p->p_lock);
2058 		if (error != 0) {
2059 			/*
2060 			 * Don't leak kauth retval if we're silently
2061 			 * skipping this entry.
2062 			 */
2063 			error = 0;
2064 			continue;
2065 		}
2066 
2067 		/*
2068 		 * Grab a hold on the process.
2069 		 */
2070 		if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2071 			continue;
2072 		}
2073 		mutex_exit(&proc_lock);
2074 
2075 		fd = p->p_fd;
2076 		mutex_enter(&fd->fd_lock);
2077 		dt = fd->fd_dt;
2078 		for (i = 0; i < dt->dt_nfiles; i++) {
2079 			fdfile_t *ff;
2080 
2081 			if ((ff = dt->dt_ff[i]) == NULL) {
2082 				continue;
2083 			}
2084 			if ((fp = atomic_load_consume(&ff->ff_file)) == NULL) {
2085 				continue;
2086 			}
2087 
2088 			mutex_enter(&fp->f_lock);
2089 
2090 			if ((fp->f_count == 0) ||
2091 			    (fp->f_marker == sysctl_file_marker)) {
2092 				mutex_exit(&fp->f_lock);
2093 				continue;
2094 			}
2095 
2096 			/* Check that we have enough space. */
2097 			if (buflen < sizeof(struct file)) {
2098 				*oldlenp = where - start;
2099 				mutex_exit(&fp->f_lock);
2100 				error = ENOMEM;
2101 				break;
2102 			}
2103 
2104 			fill_file(&fbuf, fp);
2105 			mutex_exit(&fp->f_lock);
2106 			error = sysctl_copyout(l, &fbuf, where, sizeof(fbuf));
2107 			if (error) {
2108 				break;
2109 			}
2110 			buflen -= sizeof(struct file);
2111 			where += sizeof(struct file);
2112 
2113 			fp->f_marker = sysctl_file_marker;
2114 		}
2115 		mutex_exit(&fd->fd_lock);
2116 
2117 		/*
2118 		 * Release reference to process.
2119 		 */
2120 		mutex_enter(&proc_lock);
2121 		rw_exit(&p->p_reflock);
2122 
2123 		if (error)
2124 			break;
2125 	}
2126 
2127 	sysctl_file_marker++;
2128 	/* Reset all markers if wrapped. */
2129 	if (sysctl_file_marker == 0) {
2130 		sysctl_file_marker_reset();
2131 		sysctl_file_marker++;
2132 	}
2133 
2134 	mutex_exit(&proc_lock);
2135 	mutex_exit(&sysctl_file_marker_lock);
2136 
2137 	*oldlenp = where - start;
2138 	sysctl_relock();
2139 	return error;
2140 }
2141 
2142 /*
2143  * sysctl helper function for kern.file2
2144  */
2145 static int
2146 sysctl_kern_file2(SYSCTLFN_ARGS)
2147 {
2148 	struct proc *p;
2149 	struct file *fp;
2150 	struct filedesc *fd;
2151 	struct kinfo_file kf;
2152 	char *dp;
2153 	u_int i, op;
2154 	size_t len, needed, elem_size, out_size;
2155 	int error, arg, elem_count;
2156 	fdfile_t *ff;
2157 	fdtab_t *dt;
2158 
2159 	if (namelen == 1 && name[0] == CTL_QUERY)
2160 		return sysctl_query(SYSCTLFN_CALL(rnode));
2161 
2162 	if (namelen != 4)
2163 		return EINVAL;
2164 
2165 	error = 0;
2166 	dp = oldp;
2167 	len = (oldp != NULL) ? *oldlenp : 0;
2168 	op = name[0];
2169 	arg = name[1];
2170 	elem_size = name[2];
2171 	elem_count = name[3];
2172 	out_size = MIN(sizeof(kf), elem_size);
2173 	needed = 0;
2174 
2175 	if (elem_size < 1 || elem_count < 0)
2176 		return EINVAL;
2177 
2178 	switch (op) {
2179 	case KERN_FILE_BYFILE:
2180 	case KERN_FILE_BYPID:
2181 		/*
2182 		 * We're traversing the process list in both cases; the BYFILE
2183 		 * case does additional work of keeping track of files already
2184 		 * looked at.
2185 		 */
2186 
2187 		/* doesn't use arg so it must be zero */
2188 		if ((op == KERN_FILE_BYFILE) && (arg != 0))
2189 			return EINVAL;
2190 
2191 		if ((op == KERN_FILE_BYPID) && (arg < -1))
2192 			/* -1 means all processes */
2193 			return EINVAL;
2194 
2195 		sysctl_unlock();
2196 		if (op == KERN_FILE_BYFILE)
2197 			mutex_enter(&sysctl_file_marker_lock);
2198 		mutex_enter(&proc_lock);
2199 		PROCLIST_FOREACH(p, &allproc) {
2200 			if (p->p_stat == SIDL) {
2201 				/* skip embryonic processes */
2202 				continue;
2203 			}
2204 			if (arg > 0 && p->p_pid != arg) {
2205 				/* pick only the one we want */
2206 				/* XXX want 0 to mean "kernel files" */
2207 				continue;
2208 			}
2209 			mutex_enter(p->p_lock);
2210 			error = kauth_authorize_process(l->l_cred,
2211 			    KAUTH_PROCESS_CANSEE, p,
2212 			    KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2213 			    NULL, NULL);
2214 			mutex_exit(p->p_lock);
2215 			if (error != 0) {
2216 				/*
2217 				 * Don't leak kauth retval if we're silently
2218 				 * skipping this entry.
2219 				 */
2220 				error = 0;
2221 				continue;
2222 			}
2223 
2224 			/*
2225 			 * Grab a hold on the process.
2226 			 */
2227 			if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2228 				continue;
2229 			}
2230 			mutex_exit(&proc_lock);
2231 
2232 			fd = p->p_fd;
2233 			mutex_enter(&fd->fd_lock);
2234 			dt = fd->fd_dt;
2235 			for (i = 0; i < dt->dt_nfiles; i++) {
2236 				if ((ff = dt->dt_ff[i]) == NULL) {
2237 					continue;
2238 				}
2239 				if ((fp = atomic_load_consume(&ff->ff_file)) ==
2240 				    NULL) {
2241 					continue;
2242 				}
2243 
2244 				if ((op == KERN_FILE_BYFILE) &&
2245 				    (fp->f_marker == sysctl_file_marker)) {
2246 					continue;
2247 				}
2248 				if (len >= elem_size && elem_count > 0) {
2249 					mutex_enter(&fp->f_lock);
2250 					fill_file2(&kf, fp, ff, i, p->p_pid);
2251 					mutex_exit(&fp->f_lock);
2252 					mutex_exit(&fd->fd_lock);
2253 					error = sysctl_copyout(l,
2254 					    &kf, dp, out_size);
2255 					mutex_enter(&fd->fd_lock);
2256 					if (error)
2257 						break;
2258 					dp += elem_size;
2259 					len -= elem_size;
2260 				}
2261 				if (op == KERN_FILE_BYFILE)
2262 					fp->f_marker = sysctl_file_marker;
2263 				needed += elem_size;
2264 				if (elem_count > 0 && elem_count != INT_MAX)
2265 					elem_count--;
2266 			}
2267 			mutex_exit(&fd->fd_lock);
2268 
2269 			/*
2270 			 * Release reference to process.
2271 			 */
2272 			mutex_enter(&proc_lock);
2273 			rw_exit(&p->p_reflock);
2274 		}
2275 		if (op == KERN_FILE_BYFILE) {
2276 			sysctl_file_marker++;
2277 
2278 			/* Reset all markers if wrapped. */
2279 			if (sysctl_file_marker == 0) {
2280 				sysctl_file_marker_reset();
2281 				sysctl_file_marker++;
2282 			}
2283 		}
2284 		mutex_exit(&proc_lock);
2285 		if (op == KERN_FILE_BYFILE)
2286 			mutex_exit(&sysctl_file_marker_lock);
2287 		sysctl_relock();
2288 		break;
2289 	default:
2290 		return EINVAL;
2291 	}
2292 
2293 	if (oldp == NULL)
2294 		needed += KERN_FILESLOP * elem_size;
2295 	*oldlenp = needed;
2296 
2297 	return error;
2298 }
2299 
2300 static void
2301 fill_file(struct file *fp, const struct file *fpsrc)
2302 {
2303 	const bool allowaddr = get_expose_address(curproc);
2304 
2305 	memset(fp, 0, sizeof(*fp));
2306 
2307 	fp->f_offset = fpsrc->f_offset;
2308 	COND_SET_PTR(fp->f_cred, fpsrc->f_cred, allowaddr);
2309 	COND_SET_CPTR(fp->f_ops, fpsrc->f_ops, allowaddr);
2310 	COND_SET_STRUCT(fp->f_undata, fpsrc->f_undata, allowaddr);
2311 	COND_SET_STRUCT(fp->f_list, fpsrc->f_list, allowaddr);
2312 	fp->f_flag = fpsrc->f_flag;
2313 	fp->f_marker = fpsrc->f_marker;
2314 	fp->f_type = fpsrc->f_type;
2315 	fp->f_advice = fpsrc->f_advice;
2316 	fp->f_count = fpsrc->f_count;
2317 	fp->f_msgcount = fpsrc->f_msgcount;
2318 	fp->f_unpcount = fpsrc->f_unpcount;
2319 	COND_SET_STRUCT(fp->f_unplist, fpsrc->f_unplist, allowaddr);
2320 }
2321 
2322 static void
2323 fill_file2(struct kinfo_file *kp, const file_t *fp, const fdfile_t *ff,
2324 	  int i, pid_t pid)
2325 {
2326 	const bool allowaddr = get_expose_address(curproc);
2327 
2328 	memset(kp, 0, sizeof(*kp));
2329 
2330 	COND_SET_VALUE(kp->ki_fileaddr, PTRTOUINT64(fp), allowaddr);
2331 	kp->ki_flag =		fp->f_flag;
2332 	kp->ki_iflags =		0;
2333 	kp->ki_ftype =		fp->f_type;
2334 	kp->ki_count =		fp->f_count;
2335 	kp->ki_msgcount =	fp->f_msgcount;
2336 	COND_SET_VALUE(kp->ki_fucred, PTRTOUINT64(fp->f_cred), allowaddr);
2337 	kp->ki_fuid =		kauth_cred_geteuid(fp->f_cred);
2338 	kp->ki_fgid =		kauth_cred_getegid(fp->f_cred);
2339 	COND_SET_VALUE(kp->ki_fops, PTRTOUINT64(fp->f_ops), allowaddr);
2340 	kp->ki_foffset =	fp->f_offset;
2341 	COND_SET_VALUE(kp->ki_fdata, PTRTOUINT64(fp->f_data), allowaddr);
2342 
2343 	/* vnode information to glue this file to something */
2344 	if (fp->f_type == DTYPE_VNODE) {
2345 		struct vnode *vp = fp->f_vnode;
2346 
2347 		COND_SET_VALUE(kp->ki_vun, PTRTOUINT64(vp->v_un.vu_socket),
2348 		    allowaddr);
2349 		kp->ki_vsize =	vp->v_size;
2350 		kp->ki_vtype =	vp->v_type;
2351 		kp->ki_vtag =	vp->v_tag;
2352 		COND_SET_VALUE(kp->ki_vdata, PTRTOUINT64(vp->v_data),
2353 		    allowaddr);
2354 	}
2355 
2356 	/* process information when retrieved via KERN_FILE_BYPID */
2357 	if (ff != NULL) {
2358 		kp->ki_pid =		pid;
2359 		kp->ki_fd =		i;
2360 		kp->ki_ofileflags =	ff->ff_exclose;
2361 		kp->ki_usecount =	ff->ff_refcnt;
2362 	}
2363 }
2364