xref: /netbsd-src/sys/kern/kern_descrip.c (revision ccd9df534e375a4366c5b55f23782053c7a98d82)
1 /*	$NetBSD: kern_descrip.c,v 1.262 2023/10/04 22:17:09 ad Exp $	*/
2 
3 /*-
4  * Copyright (c) 2008, 2009, 2023 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 1.262 2023/10/04 22:17:09 ad 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 
110 static int	file_ctor(void *, void *, int);
111 static void	file_dtor(void *, void *);
112 static void	fdfile_ctor(fdfile_t *);
113 static void	fdfile_dtor(fdfile_t *);
114 static int	filedesc_ctor(void *, void *, int);
115 static void	filedesc_dtor(void *, void *);
116 static int	filedescopen(dev_t, int, int, lwp_t *);
117 
118 static int sysctl_kern_file(SYSCTLFN_PROTO);
119 static int sysctl_kern_file2(SYSCTLFN_PROTO);
120 static void fill_file(struct file *, const struct file *);
121 static void fill_file2(struct kinfo_file *, const file_t *, const fdfile_t *,
122 		      int, pid_t);
123 
124 const struct cdevsw filedesc_cdevsw = {
125 	.d_open = filedescopen,
126 	.d_close = noclose,
127 	.d_read = noread,
128 	.d_write = nowrite,
129 	.d_ioctl = noioctl,
130 	.d_stop = nostop,
131 	.d_tty = notty,
132 	.d_poll = nopoll,
133 	.d_mmap = nommap,
134 	.d_kqfilter = nokqfilter,
135 	.d_discard = nodiscard,
136 	.d_flag = D_OTHER | D_MPSAFE
137 };
138 
139 /* For ease of reading. */
140 __strong_alias(fd_putvnode,fd_putfile)
141 __strong_alias(fd_putsock,fd_putfile)
142 
143 /*
144  * Initialize the descriptor system.
145  */
146 void
147 fd_sys_init(void)
148 {
149 	static struct sysctllog *clog;
150 
151 	mutex_init(&filelist_lock, MUTEX_DEFAULT, IPL_NONE);
152 
153 	LIST_INIT(&filehead);
154 
155 	file_cache = pool_cache_init(sizeof(file_t), coherency_unit, 0,
156 	    0, "file", NULL, IPL_NONE, file_ctor, file_dtor, NULL);
157 	KASSERT(file_cache != NULL);
158 
159 	filedesc_cache = pool_cache_init(sizeof(filedesc_t), coherency_unit,
160 	    0, 0, "filedesc", NULL, IPL_NONE, filedesc_ctor, filedesc_dtor,
161 	    NULL);
162 	KASSERT(filedesc_cache != NULL);
163 
164 	sysctl_createv(&clog, 0, NULL, NULL,
165 		       CTLFLAG_PERMANENT,
166 		       CTLTYPE_STRUCT, "file",
167 		       SYSCTL_DESCR("System open file table"),
168 		       sysctl_kern_file, 0, NULL, 0,
169 		       CTL_KERN, KERN_FILE, CTL_EOL);
170 	sysctl_createv(&clog, 0, NULL, NULL,
171 		       CTLFLAG_PERMANENT,
172 		       CTLTYPE_STRUCT, "file2",
173 		       SYSCTL_DESCR("System open file table"),
174 		       sysctl_kern_file2, 0, NULL, 0,
175 		       CTL_KERN, KERN_FILE2, CTL_EOL);
176 }
177 
178 static bool
179 fd_isused(filedesc_t *fdp, unsigned fd)
180 {
181 	u_int off = fd >> NDENTRYSHIFT;
182 
183 	KASSERT(fd < atomic_load_consume(&fdp->fd_dt)->dt_nfiles);
184 
185 	return (fdp->fd_lomap[off] & (1U << (fd & NDENTRYMASK))) != 0;
186 }
187 
188 /*
189  * Verify that the bitmaps match the descriptor table.
190  */
191 static inline void
192 fd_checkmaps(filedesc_t *fdp)
193 {
194 #ifdef DEBUG
195 	fdtab_t *dt;
196 	u_int fd;
197 
198 	KASSERT(fdp->fd_refcnt <= 1 || mutex_owned(&fdp->fd_lock));
199 
200 	dt = fdp->fd_dt;
201 	if (fdp->fd_refcnt == -1) {
202 		/*
203 		 * fd_free tears down the table without maintaining its bitmap.
204 		 */
205 		return;
206 	}
207 	for (fd = 0; fd < dt->dt_nfiles; fd++) {
208 		if (fd < NDFDFILE) {
209 			KASSERT(dt->dt_ff[fd] ==
210 			    (fdfile_t *)fdp->fd_dfdfile[fd]);
211 		}
212 		if (dt->dt_ff[fd] == NULL) {
213 			KASSERT(!fd_isused(fdp, fd));
214 		} else if (dt->dt_ff[fd]->ff_file != NULL) {
215 			KASSERT(fd_isused(fdp, fd));
216 		}
217 	}
218 #endif
219 }
220 
221 static int
222 fd_next_zero(filedesc_t *fdp, uint32_t *bitmap, int want, u_int bits)
223 {
224 	int i, off, maxoff;
225 	uint32_t sub;
226 
227 	KASSERT(mutex_owned(&fdp->fd_lock));
228 
229 	fd_checkmaps(fdp);
230 
231 	if (want > bits)
232 		return -1;
233 
234 	off = want >> NDENTRYSHIFT;
235 	i = want & NDENTRYMASK;
236 	if (i) {
237 		sub = bitmap[off] | ((u_int)~0 >> (NDENTRIES - i));
238 		if (sub != ~0)
239 			goto found;
240 		off++;
241 	}
242 
243 	maxoff = NDLOSLOTS(bits);
244 	while (off < maxoff) {
245 		if ((sub = bitmap[off]) != ~0)
246 			goto found;
247 		off++;
248 	}
249 
250 	return -1;
251 
252  found:
253 	return (off << NDENTRYSHIFT) + ffs(~sub) - 1;
254 }
255 
256 static int
257 fd_last_set(filedesc_t *fd, int last)
258 {
259 	int off, i;
260 	fdfile_t **ff = fd->fd_dt->dt_ff;
261 	uint32_t *bitmap = fd->fd_lomap;
262 
263 	KASSERT(mutex_owned(&fd->fd_lock));
264 
265 	fd_checkmaps(fd);
266 
267 	off = (last - 1) >> NDENTRYSHIFT;
268 
269 	while (off >= 0 && !bitmap[off])
270 		off--;
271 
272 	if (off < 0)
273 		return -1;
274 
275 	i = ((off + 1) << NDENTRYSHIFT) - 1;
276 	if (i >= last)
277 		i = last - 1;
278 
279 	/* XXX should use bitmap */
280 	while (i > 0 && (ff[i] == NULL || !ff[i]->ff_allocated))
281 		i--;
282 
283 	return i;
284 }
285 
286 static inline void
287 fd_used(filedesc_t *fdp, unsigned fd)
288 {
289 	u_int off = fd >> NDENTRYSHIFT;
290 	fdfile_t *ff;
291 
292 	ff = fdp->fd_dt->dt_ff[fd];
293 
294 	KASSERT(mutex_owned(&fdp->fd_lock));
295 	KASSERT((fdp->fd_lomap[off] & (1U << (fd & NDENTRYMASK))) == 0);
296 	KASSERT(ff != NULL);
297 	KASSERT(ff->ff_file == NULL);
298 	KASSERT(!ff->ff_allocated);
299 
300 	ff->ff_allocated = true;
301 	fdp->fd_lomap[off] |= 1U << (fd & NDENTRYMASK);
302 	if (__predict_false(fdp->fd_lomap[off] == ~0)) {
303 		KASSERT((fdp->fd_himap[off >> NDENTRYSHIFT] &
304 		    (1U << (off & NDENTRYMASK))) == 0);
305 		fdp->fd_himap[off >> NDENTRYSHIFT] |= 1U << (off & NDENTRYMASK);
306 	}
307 
308 	if ((int)fd > fdp->fd_lastfile) {
309 		fdp->fd_lastfile = fd;
310 	}
311 
312 	fd_checkmaps(fdp);
313 }
314 
315 static inline void
316 fd_unused(filedesc_t *fdp, unsigned fd)
317 {
318 	u_int off = fd >> NDENTRYSHIFT;
319 	fdfile_t *ff;
320 
321 	ff = fdp->fd_dt->dt_ff[fd];
322 
323 	KASSERT(mutex_owned(&fdp->fd_lock));
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 		    (1U << (off & NDENTRYMASK))) != 0);
335 		fdp->fd_himap[off >> NDENTRYSHIFT] &=
336 		    ~(1U << (off & NDENTRYMASK));
337 	}
338 	KASSERT((fdp->fd_lomap[off] & (1U << (fd & NDENTRYMASK))) != 0);
339 	fdp->fd_lomap[off] &= ~(1U << (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 = atomic_load_consume(&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 		 * In particular, we must coordinate the following four
391 		 * memory operations:
392 		 *
393 		 *	A. fd_close store ff->ff_file = NULL
394 		 *	B. fd_close refcnt = atomic_dec_uint_nv(&ff->ff_refcnt)
395 		 *	C. fd_getfile atomic_inc_uint(&ff->ff_refcnt)
396 		 *	D. fd_getfile load fp = ff->ff_file
397 		 *
398 		 * If the order is D;A;B;C:
399 		 *
400 		 *	1. D: fp = ff->ff_file
401 		 *	2. A: ff->ff_file = NULL
402 		 *	3. B: refcnt = atomic_dec_uint_nv(&ff->ff_refcnt)
403 		 *	4. C: atomic_inc_uint(&ff->ff_refcnt)
404 		 *
405 		 * then fd_close determines that there are no more
406 		 * references and decides to free fp immediately, at
407 		 * the same that fd_getfile ends up with an fp that's
408 		 * about to be freed.  *boom*
409 		 *
410 		 * By making B a release operation in fd_close, and by
411 		 * making C an acquire operation in fd_getfile, since
412 		 * they are atomic operations on the same object, which
413 		 * has a total modification order, we guarantee either:
414 		 *
415 		 *	- B happens before C.  Then since A is
416 		 *	  sequenced before B in fd_close, and C is
417 		 *	  sequenced before D in fd_getfile, we
418 		 *	  guarantee A happens before D, so fd_getfile
419 		 *	  reads a null fp and safely fails.
420 		 *
421 		 *	- C happens before B.  Then fd_getfile may read
422 		 *	  null or nonnull, but either way, fd_close
423 		 *	  will safely wait for references to drain.
424 		 */
425 		atomic_inc_uint(&ff->ff_refcnt);
426 		membar_acquire();
427 	}
428 
429 	/*
430 	 * If the file is not open or is being closed then put the
431 	 * reference back.
432 	 */
433 	fp = atomic_load_consume(&ff->ff_file);
434 	if (__predict_true(fp != NULL)) {
435 		return fp;
436 	}
437 	fd_putfile(fd);
438 	return NULL;
439 }
440 
441 /*
442  * Release a reference to a file descriptor acquired with fd_getfile().
443  */
444 void
445 fd_putfile(unsigned fd)
446 {
447 	filedesc_t *fdp;
448 	fdfile_t *ff;
449 	u_int u, v;
450 
451 	fdp = curlwp->l_fd;
452 	KASSERT(fd < atomic_load_consume(&fdp->fd_dt)->dt_nfiles);
453 	ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
454 
455 	KASSERT(ff != NULL);
456 	KASSERT((ff->ff_refcnt & FR_MASK) > 0);
457 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
458 
459 	if (fdp->fd_refcnt == 1) {
460 		/*
461 		 * Single threaded: don't need to worry about concurrent
462 		 * access (other than earlier calls to kqueue, which may
463 		 * hold a reference to the descriptor).
464 		 */
465 		if (__predict_false((ff->ff_refcnt & FR_CLOSING) != 0)) {
466 			fd_close(fd);
467 			return;
468 		}
469 		ff->ff_refcnt--;
470 		return;
471 	}
472 
473 	/*
474 	 * Ensure that any use of the file is complete and globally
475 	 * visible before dropping the final reference.  If no membar,
476 	 * the current CPU could still access memory associated with
477 	 * the file after it has been freed or recycled by another
478 	 * CPU.
479 	 */
480 	membar_release();
481 
482 	/*
483 	 * Be optimistic and start out with the assumption that no other
484 	 * threads are trying to close the descriptor.  If the CAS fails,
485 	 * we lost a race and/or it's being closed.
486 	 */
487 	for (u = ff->ff_refcnt & FR_MASK;; u = v) {
488 		v = atomic_cas_uint(&ff->ff_refcnt, u, u - 1);
489 		if (__predict_true(u == v)) {
490 			return;
491 		}
492 		if (__predict_false((v & FR_CLOSING) != 0)) {
493 			break;
494 		}
495 	}
496 
497 	/* Another thread is waiting to close the file: join it. */
498 	(void)fd_close(fd);
499 }
500 
501 /*
502  * Convenience wrapper around fd_getfile() that returns reference
503  * to a vnode.
504  */
505 int
506 fd_getvnode(unsigned fd, file_t **fpp)
507 {
508 	vnode_t *vp;
509 	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_VNODE)) {
516 		fd_putfile(fd);
517 		return EINVAL;
518 	}
519 	vp = fp->f_vnode;
520 	if (__predict_false(vp->v_type == VBAD)) {
521 		/* XXX Is this case really necessary? */
522 		fd_putfile(fd);
523 		return EBADF;
524 	}
525 	*fpp = fp;
526 	return 0;
527 }
528 
529 /*
530  * Convenience wrapper around fd_getfile() that returns reference
531  * to a socket.
532  */
533 int
534 fd_getsock1(unsigned fd, struct socket **sop, file_t **fp)
535 {
536 	*fp = fd_getfile(fd);
537 	if (__predict_false(*fp == NULL)) {
538 		return EBADF;
539 	}
540 	if (__predict_false((*fp)->f_type != DTYPE_SOCKET)) {
541 		fd_putfile(fd);
542 		return ENOTSOCK;
543 	}
544 	*sop = (*fp)->f_socket;
545 	return 0;
546 }
547 
548 int
549 fd_getsock(unsigned fd, struct socket **sop)
550 {
551 	file_t *fp;
552 	return fd_getsock1(fd, sop, &fp);
553 }
554 
555 /*
556  * Look up the file structure corresponding to a file descriptor
557  * and return it with a reference held on the file, not the
558  * descriptor.
559  *
560  * This is heavyweight and only used when accessing descriptors
561  * from a foreign process.  The caller must ensure that `p' does
562  * not exit or fork across this call.
563  *
564  * To release the file (not descriptor) reference, use closef().
565  */
566 file_t *
567 fd_getfile2(proc_t *p, unsigned fd)
568 {
569 	filedesc_t *fdp;
570 	fdfile_t *ff;
571 	file_t *fp;
572 	fdtab_t *dt;
573 
574 	fdp = p->p_fd;
575 	mutex_enter(&fdp->fd_lock);
576 	dt = fdp->fd_dt;
577 	if (fd >= dt->dt_nfiles) {
578 		mutex_exit(&fdp->fd_lock);
579 		return NULL;
580 	}
581 	if ((ff = dt->dt_ff[fd]) == NULL) {
582 		mutex_exit(&fdp->fd_lock);
583 		return NULL;
584 	}
585 	if ((fp = atomic_load_consume(&ff->ff_file)) == NULL) {
586 		mutex_exit(&fdp->fd_lock);
587 		return NULL;
588 	}
589 	mutex_enter(&fp->f_lock);
590 	fp->f_count++;
591 	mutex_exit(&fp->f_lock);
592 	mutex_exit(&fdp->fd_lock);
593 
594 	return fp;
595 }
596 
597 /*
598  * Internal form of close.  Must be called with a reference to the
599  * descriptor, and will drop the reference.  When all descriptor
600  * references are dropped, releases the descriptor slot and a single
601  * reference to the file structure.
602  */
603 int
604 fd_close(unsigned fd)
605 {
606 	struct flock lf;
607 	filedesc_t *fdp;
608 	fdfile_t *ff;
609 	file_t *fp;
610 	proc_t *p;
611 	lwp_t *l;
612 	u_int refcnt;
613 
614 	l = curlwp;
615 	p = l->l_proc;
616 	fdp = l->l_fd;
617 	ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
618 
619 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
620 
621 	mutex_enter(&fdp->fd_lock);
622 	KASSERT((ff->ff_refcnt & FR_MASK) > 0);
623 	fp = atomic_load_consume(&ff->ff_file);
624 	if (__predict_false(fp == NULL)) {
625 		/*
626 		 * Another user of the file is already closing, and is
627 		 * waiting for other users of the file to drain.  Release
628 		 * our reference, and wake up the closer.
629 		 */
630 		membar_release();
631 		atomic_dec_uint(&ff->ff_refcnt);
632 		cv_broadcast(&ff->ff_closing);
633 		mutex_exit(&fdp->fd_lock);
634 
635 		/*
636 		 * An application error, so pretend that the descriptor
637 		 * was already closed.  We can't safely wait for it to
638 		 * be closed without potentially deadlocking.
639 		 */
640 		return (EBADF);
641 	}
642 	KASSERT((ff->ff_refcnt & FR_CLOSING) == 0);
643 
644 	/*
645 	 * There may be multiple users of this file within the process.
646 	 * Notify existing and new users that the file is closing.  This
647 	 * will prevent them from adding additional uses to this file
648 	 * while we are closing it.
649 	 */
650 	atomic_store_relaxed(&ff->ff_file, NULL);
651 	ff->ff_exclose = false;
652 
653 	/*
654 	 * We expect the caller to hold a descriptor reference - drop it.
655 	 * The reference count may increase beyond zero at this point due
656 	 * to an erroneous descriptor reference by an application, but
657 	 * fd_getfile() will notice that the file is being closed and drop
658 	 * the reference again.
659 	 */
660 	if (fdp->fd_refcnt == 1) {
661 		/* Single threaded. */
662 		refcnt = --(ff->ff_refcnt);
663 	} else {
664 		/* Multi threaded. */
665 		membar_release();
666 		refcnt = atomic_dec_uint_nv(&ff->ff_refcnt);
667 		membar_acquire();
668 	}
669 	if (__predict_false(refcnt != 0)) {
670 		/*
671 		 * Wait for other references to drain.  This is typically
672 		 * an application error - the descriptor is being closed
673 		 * while still in use.
674 		 * (Or just a threaded application trying to unblock its
675 		 * thread that sleeps in (say) accept()).
676 		 */
677 		atomic_or_uint(&ff->ff_refcnt, FR_CLOSING);
678 
679 		/*
680 		 * Remove any knotes attached to the file.  A knote
681 		 * attached to the descriptor can hold references on it.
682 		 */
683 		mutex_exit(&fdp->fd_lock);
684 		if (!SLIST_EMPTY(&ff->ff_knlist)) {
685 			knote_fdclose(fd);
686 		}
687 
688 		/*
689 		 * Since the file system code doesn't know which fd
690 		 * each request came from (think dup()), we have to
691 		 * ask it to return ERESTART for any long-term blocks.
692 		 * The re-entry through read/write/etc will detect the
693 		 * closed fd and return EBAFD.
694 		 * Blocked partial writes may return a short length.
695 		 */
696 		(*fp->f_ops->fo_restart)(fp);
697 		mutex_enter(&fdp->fd_lock);
698 
699 		/*
700 		 * We need to see the count drop to zero at least once,
701 		 * in order to ensure that all pre-existing references
702 		 * have been drained.  New references past this point are
703 		 * of no interest.
704 		 * XXX (dsl) this may need to call fo_restart() after a
705 		 * timeout to guarantee that all the system calls exit.
706 		 */
707 		while ((ff->ff_refcnt & FR_MASK) != 0) {
708 			cv_wait(&ff->ff_closing, &fdp->fd_lock);
709 		}
710 		atomic_and_uint(&ff->ff_refcnt, ~FR_CLOSING);
711 	} else {
712 		/* If no references, there must be no knotes. */
713 		KASSERT(SLIST_EMPTY(&ff->ff_knlist));
714 	}
715 
716 	/*
717 	 * POSIX record locking dictates that any close releases ALL
718 	 * locks owned by this process.  This is handled by setting
719 	 * a flag in the unlock to free ONLY locks obeying POSIX
720 	 * semantics, and not to free BSD-style file locks.
721 	 * If the descriptor was in a message, POSIX-style locks
722 	 * aren't passed with the descriptor.
723 	 */
724 	if (__predict_false((p->p_flag & PK_ADVLOCK) != 0) &&
725 	    fp->f_ops->fo_advlock != NULL) {
726 		lf.l_whence = SEEK_SET;
727 		lf.l_start = 0;
728 		lf.l_len = 0;
729 		lf.l_type = F_UNLCK;
730 		mutex_exit(&fdp->fd_lock);
731 		(void)(*fp->f_ops->fo_advlock)(fp, p, F_UNLCK, &lf, F_POSIX);
732 		mutex_enter(&fdp->fd_lock);
733 	}
734 
735 	/* Free descriptor slot. */
736 	fd_unused(fdp, fd);
737 	mutex_exit(&fdp->fd_lock);
738 
739 	/* Now drop reference to the file itself. */
740 	return closef(fp);
741 }
742 
743 /*
744  * Duplicate a file descriptor.
745  */
746 int
747 fd_dup(file_t *fp, int minfd, int *newp, bool exclose)
748 {
749 	proc_t *p = curproc;
750 	fdtab_t *dt;
751 	int error;
752 
753 	while ((error = fd_alloc(p, minfd, newp)) != 0) {
754 		if (error != ENOSPC) {
755 			return error;
756 		}
757 		fd_tryexpand(p);
758 	}
759 
760 	dt = atomic_load_consume(&curlwp->l_fd->fd_dt);
761 	dt->dt_ff[*newp]->ff_exclose = exclose;
762 	fd_affix(p, fp, *newp);
763 	return 0;
764 }
765 
766 /*
767  * dup2 operation.
768  */
769 int
770 fd_dup2(file_t *fp, unsigned newfd, int flags)
771 {
772 	filedesc_t *fdp = curlwp->l_fd;
773 	fdfile_t *ff;
774 	fdtab_t *dt;
775 
776 	if (flags & ~(O_CLOEXEC|O_NONBLOCK|O_NOSIGPIPE))
777 		return EINVAL;
778 	/*
779 	 * Ensure there are enough slots in the descriptor table,
780 	 * and allocate an fdfile_t up front in case we need it.
781 	 */
782 	while (newfd >= atomic_load_consume(&fdp->fd_dt)->dt_nfiles) {
783 		fd_tryexpand(curproc);
784 	}
785 	ff = kmem_alloc(sizeof(*ff), KM_SLEEP);
786 	fdfile_ctor(ff);
787 
788 	/*
789 	 * If there is already a file open, close it.  If the file is
790 	 * half open, wait for it to be constructed before closing it.
791 	 * XXX Potential for deadlock here?
792 	 */
793 	mutex_enter(&fdp->fd_lock);
794 	while (fd_isused(fdp, newfd)) {
795 		mutex_exit(&fdp->fd_lock);
796 		if (fd_getfile(newfd) != NULL) {
797 			(void)fd_close(newfd);
798 		} else {
799 			/*
800 			 * Crummy, but unlikely to happen.
801 			 * Can occur if we interrupt another
802 			 * thread while it is opening a file.
803 			 */
804 			kpause("dup2", false, 1, NULL);
805 		}
806 		mutex_enter(&fdp->fd_lock);
807 	}
808 	dt = fdp->fd_dt;
809 	if (dt->dt_ff[newfd] == NULL) {
810 		KASSERT(newfd >= NDFDFILE);
811 		dt->dt_ff[newfd] = ff;
812 		ff = NULL;
813 	}
814 	fd_used(fdp, newfd);
815 	mutex_exit(&fdp->fd_lock);
816 
817 	dt->dt_ff[newfd]->ff_exclose = (flags & O_CLOEXEC) != 0;
818 	fp->f_flag |= flags & (FNONBLOCK|FNOSIGPIPE);
819 	/* Slot is now allocated.  Insert copy of the file. */
820 	fd_affix(curproc, fp, newfd);
821 	if (ff != NULL) {
822 		cv_destroy(&ff->ff_closing);
823 		kmem_free(ff, sizeof(*ff));
824 	}
825 	return 0;
826 }
827 
828 /*
829  * Drop reference to a file structure.
830  */
831 int
832 closef(file_t *fp)
833 {
834 	struct flock lf;
835 	int error;
836 
837 	/*
838 	 * Drop reference.  If referenced elsewhere it's still open
839 	 * and we have nothing more to do.
840 	 */
841 	mutex_enter(&fp->f_lock);
842 	KASSERT(fp->f_count > 0);
843 	if (--fp->f_count > 0) {
844 		mutex_exit(&fp->f_lock);
845 		return 0;
846 	}
847 	KASSERT(fp->f_count == 0);
848 	mutex_exit(&fp->f_lock);
849 
850 	/* We held the last reference - release locks, close and free. */
851 	if (fp->f_ops->fo_advlock == NULL) {
852 		KASSERT((fp->f_flag & FHASLOCK) == 0);
853 	} else if (fp->f_flag & FHASLOCK) {
854 		lf.l_whence = SEEK_SET;
855 		lf.l_start = 0;
856 		lf.l_len = 0;
857 		lf.l_type = F_UNLCK;
858 		(void)(*fp->f_ops->fo_advlock)(fp, fp, F_UNLCK, &lf, F_FLOCK);
859 	}
860 	if (fp->f_ops != NULL) {
861 		error = (*fp->f_ops->fo_close)(fp);
862 	} else {
863 		error = 0;
864 	}
865 	KASSERT(fp->f_count == 0);
866 	KASSERT(fp->f_cred != NULL);
867 	pool_cache_put(file_cache, fp);
868 
869 	return error;
870 }
871 
872 /*
873  * Allocate a file descriptor for the process.
874  *
875  * Future idea for experimentation: replace all of this with radixtree.
876  */
877 int
878 fd_alloc(proc_t *p, int want, int *result)
879 {
880 	filedesc_t *fdp = p->p_fd;
881 	int i, lim, last, error, hi;
882 	u_int off;
883 	fdtab_t *dt;
884 
885 	KASSERT(p == curproc || p == &proc0);
886 
887 	/*
888 	 * Search for a free descriptor starting at the higher
889 	 * of want or fd_freefile.
890 	 */
891 	mutex_enter(&fdp->fd_lock);
892 	fd_checkmaps(fdp);
893 	dt = fdp->fd_dt;
894 	KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
895 	lim = uimin((int)p->p_rlimit[RLIMIT_NOFILE].rlim_cur, maxfiles);
896 	last = uimin(dt->dt_nfiles, lim);
897 
898 	for (;;) {
899 		if ((i = want) < fdp->fd_freefile)
900 			i = fdp->fd_freefile;
901 		off = i >> NDENTRYSHIFT;
902 		hi = fd_next_zero(fdp, fdp->fd_himap, off,
903 		    (last + NDENTRIES - 1) >> NDENTRYSHIFT);
904 		if (hi == -1)
905 			break;
906 		i = fd_next_zero(fdp, &fdp->fd_lomap[hi],
907 		    hi > off ? 0 : i & NDENTRYMASK, NDENTRIES);
908 		if (i == -1) {
909 			/*
910 			 * Free file descriptor in this block was
911 			 * below want, try again with higher want.
912 			 */
913 			want = (hi + 1) << NDENTRYSHIFT;
914 			continue;
915 		}
916 		i += (hi << NDENTRYSHIFT);
917 		if (i >= last) {
918 			break;
919 		}
920 		if (dt->dt_ff[i] == NULL) {
921 			KASSERT(i >= NDFDFILE);
922 			dt->dt_ff[i] = kmem_alloc(sizeof(fdfile_t), KM_SLEEP);
923 			fdfile_ctor(dt->dt_ff[i]);
924 		}
925 		KASSERT(dt->dt_ff[i]->ff_file == NULL);
926 		fd_used(fdp, i);
927 		if (want <= fdp->fd_freefile) {
928 			fdp->fd_freefile = i;
929 		}
930 		*result = i;
931 		KASSERT(i >= NDFDFILE ||
932 		    dt->dt_ff[i] == (fdfile_t *)fdp->fd_dfdfile[i]);
933 		fd_checkmaps(fdp);
934 		mutex_exit(&fdp->fd_lock);
935 		return 0;
936 	}
937 
938 	/* No space in current array.  Let the caller expand and retry. */
939 	error = (dt->dt_nfiles >= lim) ? EMFILE : ENOSPC;
940 	mutex_exit(&fdp->fd_lock);
941 	return error;
942 }
943 
944 /*
945  * Allocate memory for a descriptor table.
946  */
947 static fdtab_t *
948 fd_dtab_alloc(int n)
949 {
950 	fdtab_t *dt;
951 	size_t sz;
952 
953 	KASSERT(n > NDFILE);
954 
955 	sz = sizeof(*dt) + (n - NDFILE) * sizeof(dt->dt_ff[0]);
956 	dt = kmem_alloc(sz, KM_SLEEP);
957 #ifdef DIAGNOSTIC
958 	memset(dt, 0xff, sz);
959 #endif
960 	dt->dt_nfiles = n;
961 	dt->dt_link = NULL;
962 	return dt;
963 }
964 
965 /*
966  * Free a descriptor table, and all tables linked for deferred free.
967  */
968 static void
969 fd_dtab_free(fdtab_t *dt)
970 {
971 	fdtab_t *next;
972 	size_t sz;
973 
974 	do {
975 		next = dt->dt_link;
976 		KASSERT(dt->dt_nfiles > NDFILE);
977 		sz = sizeof(*dt) +
978 		    (dt->dt_nfiles - NDFILE) * sizeof(dt->dt_ff[0]);
979 #ifdef DIAGNOSTIC
980 		memset(dt, 0xff, sz);
981 #endif
982 		kmem_free(dt, sz);
983 		dt = next;
984 	} while (dt != NULL);
985 }
986 
987 /*
988  * Allocate descriptor bitmap.
989  */
990 static void
991 fd_map_alloc(int n, uint32_t **lo, uint32_t **hi)
992 {
993 	uint8_t *ptr;
994 	size_t szlo, szhi;
995 
996 	KASSERT(n > NDENTRIES);
997 
998 	szlo = NDLOSLOTS(n) * sizeof(uint32_t);
999 	szhi = NDHISLOTS(n) * sizeof(uint32_t);
1000 	ptr = kmem_alloc(szlo + szhi, KM_SLEEP);
1001 	*lo = (uint32_t *)ptr;
1002 	*hi = (uint32_t *)(ptr + szlo);
1003 }
1004 
1005 /*
1006  * Free descriptor bitmap.
1007  */
1008 static void
1009 fd_map_free(int n, uint32_t *lo, uint32_t *hi)
1010 {
1011 	size_t szlo, szhi;
1012 
1013 	KASSERT(n > NDENTRIES);
1014 
1015 	szlo = NDLOSLOTS(n) * sizeof(uint32_t);
1016 	szhi = NDHISLOTS(n) * sizeof(uint32_t);
1017 	KASSERT(hi == (uint32_t *)((uint8_t *)lo + szlo));
1018 	kmem_free(lo, szlo + szhi);
1019 }
1020 
1021 /*
1022  * Expand a process' descriptor table.
1023  */
1024 void
1025 fd_tryexpand(proc_t *p)
1026 {
1027 	filedesc_t *fdp;
1028 	int i, numfiles, oldnfiles;
1029 	fdtab_t *newdt, *dt;
1030 	uint32_t *newhimap, *newlomap;
1031 
1032 	KASSERT(p == curproc || p == &proc0);
1033 
1034 	fdp = p->p_fd;
1035 	newhimap = NULL;
1036 	newlomap = NULL;
1037 	oldnfiles = atomic_load_consume(&fdp->fd_dt)->dt_nfiles;
1038 
1039 	if (oldnfiles < NDEXTENT)
1040 		numfiles = NDEXTENT;
1041 	else
1042 		numfiles = 2 * oldnfiles;
1043 
1044 	newdt = fd_dtab_alloc(numfiles);
1045 	if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1046 		fd_map_alloc(numfiles, &newlomap, &newhimap);
1047 	}
1048 
1049 	mutex_enter(&fdp->fd_lock);
1050 	dt = fdp->fd_dt;
1051 	KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1052 	if (dt->dt_nfiles != oldnfiles) {
1053 		/* fdp changed; caller must retry */
1054 		mutex_exit(&fdp->fd_lock);
1055 		fd_dtab_free(newdt);
1056 		if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1057 			fd_map_free(numfiles, newlomap, newhimap);
1058 		}
1059 		return;
1060 	}
1061 
1062 	/* Copy the existing descriptor table and zero the new portion. */
1063 	i = sizeof(fdfile_t *) * oldnfiles;
1064 	memcpy(newdt->dt_ff, dt->dt_ff, i);
1065 	memset((uint8_t *)newdt->dt_ff + i, 0,
1066 	    numfiles * sizeof(fdfile_t *) - i);
1067 
1068 	/*
1069 	 * Link old descriptor array into list to be discarded.  We defer
1070 	 * freeing until the last reference to the descriptor table goes
1071 	 * away (usually process exit).  This allows us to do lockless
1072 	 * lookups in fd_getfile().
1073 	 */
1074 	if (oldnfiles > NDFILE) {
1075 		if (fdp->fd_refcnt > 1) {
1076 			newdt->dt_link = dt;
1077 		} else {
1078 			fd_dtab_free(dt);
1079 		}
1080 	}
1081 
1082 	if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1083 		i = NDHISLOTS(oldnfiles) * sizeof(uint32_t);
1084 		memcpy(newhimap, fdp->fd_himap, i);
1085 		memset((uint8_t *)newhimap + i, 0,
1086 		    NDHISLOTS(numfiles) * sizeof(uint32_t) - i);
1087 
1088 		i = NDLOSLOTS(oldnfiles) * sizeof(uint32_t);
1089 		memcpy(newlomap, fdp->fd_lomap, i);
1090 		memset((uint8_t *)newlomap + i, 0,
1091 		    NDLOSLOTS(numfiles) * sizeof(uint32_t) - i);
1092 
1093 		if (NDHISLOTS(oldnfiles) > NDHISLOTS(NDFILE)) {
1094 			fd_map_free(oldnfiles, fdp->fd_lomap, fdp->fd_himap);
1095 		}
1096 		fdp->fd_himap = newhimap;
1097 		fdp->fd_lomap = newlomap;
1098 	}
1099 
1100 	/*
1101 	 * All other modifications must become globally visible before
1102 	 * the change to fd_dt.  See fd_getfile().
1103 	 */
1104 	atomic_store_release(&fdp->fd_dt, newdt);
1105 	KASSERT(newdt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1106 	fd_checkmaps(fdp);
1107 	mutex_exit(&fdp->fd_lock);
1108 }
1109 
1110 /*
1111  * Create a new open file structure and allocate a file descriptor
1112  * for the current process.
1113  */
1114 int
1115 fd_allocfile(file_t **resultfp, int *resultfd)
1116 {
1117 	proc_t *p = curproc;
1118 	kauth_cred_t cred;
1119 	file_t *fp;
1120 	int error;
1121 
1122 	while ((error = fd_alloc(p, 0, resultfd)) != 0) {
1123 		if (error != ENOSPC) {
1124 			return error;
1125 		}
1126 		fd_tryexpand(p);
1127 	}
1128 
1129 	fp = pool_cache_get(file_cache, PR_WAITOK);
1130 	if (fp == NULL) {
1131 		fd_abort(p, NULL, *resultfd);
1132 		return ENFILE;
1133 	}
1134 	KASSERT(fp->f_count == 0);
1135 	KASSERT(fp->f_msgcount == 0);
1136 	KASSERT(fp->f_unpcount == 0);
1137 
1138 	/* Replace cached credentials if not what we need. */
1139 	cred = curlwp->l_cred;
1140 	if (__predict_false(cred != fp->f_cred)) {
1141 		kauth_cred_free(fp->f_cred);
1142 		fp->f_cred = kauth_cred_hold(cred);
1143 	}
1144 
1145 	/*
1146 	 * Don't allow recycled files to be scanned.
1147 	 * See uipc_usrreq.c.
1148 	 */
1149 	if (__predict_false((fp->f_flag & FSCAN) != 0)) {
1150 		mutex_enter(&fp->f_lock);
1151 		atomic_and_uint(&fp->f_flag, ~FSCAN);
1152 		mutex_exit(&fp->f_lock);
1153 	}
1154 
1155 	fp->f_advice = 0;
1156 	fp->f_offset = 0;
1157 	*resultfp = fp;
1158 
1159 	return 0;
1160 }
1161 
1162 /*
1163  * Successful creation of a new descriptor: make visible to the process.
1164  */
1165 void
1166 fd_affix(proc_t *p, file_t *fp, unsigned fd)
1167 {
1168 	fdfile_t *ff;
1169 	filedesc_t *fdp;
1170 	fdtab_t *dt;
1171 
1172 	KASSERT(p == curproc || p == &proc0);
1173 
1174 	/* Add a reference to the file structure. */
1175 	mutex_enter(&fp->f_lock);
1176 	fp->f_count++;
1177 	mutex_exit(&fp->f_lock);
1178 
1179 	/*
1180 	 * Insert the new file into the descriptor slot.
1181 	 */
1182 	fdp = p->p_fd;
1183 	dt = atomic_load_consume(&fdp->fd_dt);
1184 	ff = dt->dt_ff[fd];
1185 
1186 	KASSERT(ff != NULL);
1187 	KASSERT(ff->ff_file == NULL);
1188 	KASSERT(ff->ff_allocated);
1189 	KASSERT(fd_isused(fdp, fd));
1190 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1191 
1192 	/* No need to lock in order to make file initially visible. */
1193 	atomic_store_release(&ff->ff_file, fp);
1194 }
1195 
1196 /*
1197  * Abort creation of a new descriptor: free descriptor slot and file.
1198  */
1199 void
1200 fd_abort(proc_t *p, file_t *fp, unsigned fd)
1201 {
1202 	filedesc_t *fdp;
1203 	fdfile_t *ff;
1204 
1205 	KASSERT(p == curproc || p == &proc0);
1206 
1207 	fdp = p->p_fd;
1208 	ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
1209 	ff->ff_exclose = false;
1210 
1211 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1212 
1213 	mutex_enter(&fdp->fd_lock);
1214 	KASSERT(fd_isused(fdp, fd));
1215 	fd_unused(fdp, fd);
1216 	mutex_exit(&fdp->fd_lock);
1217 
1218 	if (fp != NULL) {
1219 		KASSERT(fp->f_count == 0);
1220 		KASSERT(fp->f_cred != NULL);
1221 		pool_cache_put(file_cache, fp);
1222 	}
1223 }
1224 
1225 static int
1226 file_ctor(void *arg, void *obj, int flags)
1227 {
1228 	/*
1229 	 * It's easy to exhaust the open file limit on a system with many
1230 	 * CPUs due to caching.  Allow a bit of leeway to reduce the element
1231 	 * of surprise.
1232 	 */
1233 	u_int slop = PCG_NOBJECTS_NORMAL * (ncpu - 1);
1234 	file_t *fp = obj;
1235 
1236 	memset(fp, 0, sizeof(*fp));
1237 
1238 	mutex_enter(&filelist_lock);
1239 	if (__predict_false(nfiles >= slop + maxfiles)) {
1240 		mutex_exit(&filelist_lock);
1241 		tablefull("file", "increase kern.maxfiles or MAXFILES");
1242 		return ENFILE;
1243 	}
1244 	nfiles++;
1245 	LIST_INSERT_HEAD(&filehead, fp, f_list);
1246 	mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1247 	fp->f_cred = kauth_cred_hold(curlwp->l_cred);
1248 	mutex_exit(&filelist_lock);
1249 
1250 	return 0;
1251 }
1252 
1253 static void
1254 file_dtor(void *arg, void *obj)
1255 {
1256 	file_t *fp = obj;
1257 
1258 	mutex_enter(&filelist_lock);
1259 	nfiles--;
1260 	LIST_REMOVE(fp, f_list);
1261 	mutex_exit(&filelist_lock);
1262 
1263 	KASSERT(fp->f_count == 0);
1264 	kauth_cred_free(fp->f_cred);
1265 	mutex_destroy(&fp->f_lock);
1266 }
1267 
1268 static void
1269 fdfile_ctor(fdfile_t *ff)
1270 {
1271 
1272 	memset(ff, 0, sizeof(*ff));
1273 	cv_init(&ff->ff_closing, "fdclose");
1274 }
1275 
1276 static void
1277 fdfile_dtor(fdfile_t *ff)
1278 {
1279 
1280 	cv_destroy(&ff->ff_closing);
1281 }
1282 
1283 file_t *
1284 fgetdummy(void)
1285 {
1286 	file_t *fp;
1287 
1288 	fp = kmem_zalloc(sizeof(*fp), KM_SLEEP);
1289 	mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1290 	return fp;
1291 }
1292 
1293 void
1294 fputdummy(file_t *fp)
1295 {
1296 
1297 	mutex_destroy(&fp->f_lock);
1298 	kmem_free(fp, sizeof(*fp));
1299 }
1300 
1301 /*
1302  * Create an initial filedesc structure.
1303  */
1304 filedesc_t *
1305 fd_init(filedesc_t *fdp)
1306 {
1307 #ifdef DIAGNOSTIC
1308 	unsigned fd;
1309 #endif
1310 
1311 	if (__predict_true(fdp == NULL)) {
1312 		fdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1313 	} else {
1314 		KASSERT(fdp == &filedesc0);
1315 		filedesc_ctor(NULL, fdp, PR_WAITOK);
1316 	}
1317 
1318 #ifdef DIAGNOSTIC
1319 	KASSERT(fdp->fd_lastfile == -1);
1320 	KASSERT(fdp->fd_lastkqfile == -1);
1321 	KASSERT(fdp->fd_knhash == NULL);
1322 	KASSERT(fdp->fd_freefile == 0);
1323 	KASSERT(fdp->fd_exclose == false);
1324 	KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1325 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1326 	for (fd = 0; fd < NDFDFILE; fd++) {
1327 		KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] ==
1328 		    (fdfile_t *)fdp->fd_dfdfile[fd]);
1329 	}
1330 	for (fd = NDFDFILE; fd < NDFILE; fd++) {
1331 		KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] == NULL);
1332 	}
1333 	KASSERT(fdp->fd_himap == fdp->fd_dhimap);
1334 	KASSERT(fdp->fd_lomap == fdp->fd_dlomap);
1335 #endif	/* DIAGNOSTIC */
1336 
1337 	fdp->fd_refcnt = 1;
1338 	fd_checkmaps(fdp);
1339 
1340 	return fdp;
1341 }
1342 
1343 /*
1344  * Initialize a file descriptor table.
1345  */
1346 static int
1347 filedesc_ctor(void *arg, void *obj, int flag)
1348 {
1349 	filedesc_t *fdp = obj;
1350 	fdfile_t **ffp;
1351 	int i;
1352 
1353 	memset(fdp, 0, sizeof(*fdp));
1354 	mutex_init(&fdp->fd_lock, MUTEX_DEFAULT, IPL_NONE);
1355 	fdp->fd_lastfile = -1;
1356 	fdp->fd_lastkqfile = -1;
1357 	fdp->fd_dt = &fdp->fd_dtbuiltin;
1358 	fdp->fd_dtbuiltin.dt_nfiles = NDFILE;
1359 	fdp->fd_himap = fdp->fd_dhimap;
1360 	fdp->fd_lomap = fdp->fd_dlomap;
1361 
1362 	CTASSERT(sizeof(fdp->fd_dfdfile[0]) >= sizeof(fdfile_t));
1363 	for (i = 0, ffp = fdp->fd_dt->dt_ff; i < NDFDFILE; i++, ffp++) {
1364 		fdfile_ctor(*ffp = (fdfile_t *)fdp->fd_dfdfile[i]);
1365 	}
1366 
1367 	return 0;
1368 }
1369 
1370 static void
1371 filedesc_dtor(void *arg, void *obj)
1372 {
1373 	filedesc_t *fdp = obj;
1374 	int i;
1375 
1376 	for (i = 0; i < NDFDFILE; i++) {
1377 		fdfile_dtor((fdfile_t *)fdp->fd_dfdfile[i]);
1378 	}
1379 
1380 	mutex_destroy(&fdp->fd_lock);
1381 }
1382 
1383 /*
1384  * Make p share curproc's filedesc structure.
1385  */
1386 void
1387 fd_share(struct proc *p)
1388 {
1389 	filedesc_t *fdp;
1390 
1391 	fdp = curlwp->l_fd;
1392 	p->p_fd = fdp;
1393 	atomic_inc_uint(&fdp->fd_refcnt);
1394 }
1395 
1396 /*
1397  * Acquire a hold on a filedesc structure.
1398  */
1399 void
1400 fd_hold(lwp_t *l)
1401 {
1402 	filedesc_t *fdp = l->l_fd;
1403 
1404 	atomic_inc_uint(&fdp->fd_refcnt);
1405 }
1406 
1407 /*
1408  * Copy a filedesc structure.
1409  */
1410 filedesc_t *
1411 fd_copy(void)
1412 {
1413 	filedesc_t *newfdp, *fdp;
1414 	fdfile_t *ff, **ffp, **nffp, *ff2;
1415 	int i, j, numfiles, lastfile, newlast;
1416 	file_t *fp;
1417 	fdtab_t *newdt;
1418 
1419 	fdp = curproc->p_fd;
1420 	newfdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1421 	newfdp->fd_refcnt = 1;
1422 
1423 #ifdef DIAGNOSTIC
1424 	KASSERT(newfdp->fd_lastfile == -1);
1425 	KASSERT(newfdp->fd_lastkqfile == -1);
1426 	KASSERT(newfdp->fd_knhash == NULL);
1427 	KASSERT(newfdp->fd_freefile == 0);
1428 	KASSERT(newfdp->fd_exclose == false);
1429 	KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1430 	KASSERT(newfdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1431 	for (i = 0; i < NDFDFILE; i++) {
1432 		KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] ==
1433 		    (fdfile_t *)&newfdp->fd_dfdfile[i]);
1434 	}
1435 	for (i = NDFDFILE; i < NDFILE; i++) {
1436 		KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] == NULL);
1437 	}
1438 #endif	/* DIAGNOSTIC */
1439 
1440 	mutex_enter(&fdp->fd_lock);
1441 	fd_checkmaps(fdp);
1442 	numfiles = fdp->fd_dt->dt_nfiles;
1443 	lastfile = fdp->fd_lastfile;
1444 
1445 	/*
1446 	 * If the number of open files fits in the internal arrays
1447 	 * of the open file structure, use them, otherwise allocate
1448 	 * additional memory for the number of descriptors currently
1449 	 * in use.
1450 	 */
1451 	if (lastfile < NDFILE) {
1452 		i = NDFILE;
1453 		newdt = newfdp->fd_dt;
1454 		KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1455 	} else {
1456 		/*
1457 		 * Compute the smallest multiple of NDEXTENT needed
1458 		 * for the file descriptors currently in use,
1459 		 * allowing the table to shrink.
1460 		 */
1461 		i = numfiles;
1462 		while (i >= 2 * NDEXTENT && i > lastfile * 2) {
1463 			i /= 2;
1464 		}
1465 		KASSERT(i > NDFILE);
1466 		newdt = fd_dtab_alloc(i);
1467 		newfdp->fd_dt = newdt;
1468 		memcpy(newdt->dt_ff, newfdp->fd_dtbuiltin.dt_ff,
1469 		    NDFDFILE * sizeof(fdfile_t **));
1470 		memset(newdt->dt_ff + NDFDFILE, 0,
1471 		    (i - NDFDFILE) * sizeof(fdfile_t **));
1472 	}
1473 	if (NDHISLOTS(i) <= NDHISLOTS(NDFILE)) {
1474 		newfdp->fd_himap = newfdp->fd_dhimap;
1475 		newfdp->fd_lomap = newfdp->fd_dlomap;
1476 	} else {
1477 		fd_map_alloc(i, &newfdp->fd_lomap, &newfdp->fd_himap);
1478 		KASSERT(i >= NDENTRIES * NDENTRIES);
1479 		memset(newfdp->fd_himap, 0, NDHISLOTS(i)*sizeof(uint32_t));
1480 		memset(newfdp->fd_lomap, 0, NDLOSLOTS(i)*sizeof(uint32_t));
1481 	}
1482 	newfdp->fd_freefile = fdp->fd_freefile;
1483 	newfdp->fd_exclose = fdp->fd_exclose;
1484 
1485 	ffp = fdp->fd_dt->dt_ff;
1486 	nffp = newdt->dt_ff;
1487 	newlast = -1;
1488 	for (i = 0; i <= lastfile; i++, ffp++, nffp++) {
1489 		KASSERT(i >= NDFDFILE ||
1490 		    *nffp == (fdfile_t *)newfdp->fd_dfdfile[i]);
1491 		ff = *ffp;
1492 		if (ff == NULL ||
1493 		    (fp = atomic_load_consume(&ff->ff_file)) == NULL) {
1494 			/* Descriptor unused, or descriptor half open. */
1495 			KASSERT(!fd_isused(newfdp, i));
1496 			continue;
1497 		}
1498 		if (__predict_false(fp->f_type == DTYPE_KQUEUE)) {
1499 			/* kqueue descriptors cannot be copied. */
1500 			if (i < newfdp->fd_freefile) {
1501 				newfdp->fd_freefile = i;
1502 			}
1503 			continue;
1504 		}
1505 		/* It's active: add a reference to the file. */
1506 		mutex_enter(&fp->f_lock);
1507 		fp->f_count++;
1508 		mutex_exit(&fp->f_lock);
1509 
1510 		/* Allocate an fdfile_t to represent it. */
1511 		if (i >= NDFDFILE) {
1512 			ff2 = kmem_alloc(sizeof(*ff2), KM_SLEEP);
1513 			fdfile_ctor(ff2);
1514 			*nffp = ff2;
1515 		} else {
1516 			ff2 = newdt->dt_ff[i];
1517 		}
1518 		ff2->ff_file = fp;
1519 		ff2->ff_exclose = ff->ff_exclose;
1520 		ff2->ff_allocated = true;
1521 
1522 		/* Fix up bitmaps. */
1523 		j = i >> NDENTRYSHIFT;
1524 		KASSERT((newfdp->fd_lomap[j] & (1U << (i & NDENTRYMASK))) == 0);
1525 		newfdp->fd_lomap[j] |= 1U << (i & NDENTRYMASK);
1526 		if (__predict_false(newfdp->fd_lomap[j] == ~0)) {
1527 			KASSERT((newfdp->fd_himap[j >> NDENTRYSHIFT] &
1528 			    (1U << (j & NDENTRYMASK))) == 0);
1529 			newfdp->fd_himap[j >> NDENTRYSHIFT] |=
1530 			    1U << (j & NDENTRYMASK);
1531 		}
1532 		newlast = i;
1533 	}
1534 	KASSERT(newdt->dt_ff[0] == (fdfile_t *)newfdp->fd_dfdfile[0]);
1535 	newfdp->fd_lastfile = newlast;
1536 	fd_checkmaps(newfdp);
1537 	mutex_exit(&fdp->fd_lock);
1538 
1539 	return newfdp;
1540 }
1541 
1542 /*
1543  * Release a filedesc structure.
1544  */
1545 void
1546 fd_free(void)
1547 {
1548 	fdfile_t *ff;
1549 	file_t *fp;
1550 	int fd, nf;
1551 	fdtab_t *dt;
1552 	lwp_t * const l = curlwp;
1553 	filedesc_t * const fdp = l->l_fd;
1554 	const bool noadvlock = (l->l_proc->p_flag & PK_ADVLOCK) == 0;
1555 
1556 	KASSERT(atomic_load_consume(&fdp->fd_dt)->dt_ff[0] ==
1557 	    (fdfile_t *)fdp->fd_dfdfile[0]);
1558 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1559 	KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1560 
1561 	membar_release();
1562 	if (atomic_dec_uint_nv(&fdp->fd_refcnt) > 0)
1563 		return;
1564 	membar_acquire();
1565 
1566 	/*
1567 	 * Close any files that the process holds open.
1568 	 */
1569 	dt = fdp->fd_dt;
1570 	fd_checkmaps(fdp);
1571 #ifdef DEBUG
1572 	fdp->fd_refcnt = -1; /* see fd_checkmaps */
1573 #endif
1574 	for (fd = 0, nf = dt->dt_nfiles; fd < nf; fd++) {
1575 		ff = dt->dt_ff[fd];
1576 		KASSERT(fd >= NDFDFILE ||
1577 		    ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1578 		if (ff == NULL)
1579 			continue;
1580 		if ((fp = atomic_load_consume(&ff->ff_file)) != NULL) {
1581 			/*
1582 			 * Must use fd_close() here if there is
1583 			 * a reference from kqueue or we might have posix
1584 			 * advisory locks.
1585 			 */
1586 			if (__predict_true(ff->ff_refcnt == 0) &&
1587 			    (noadvlock || fp->f_type != DTYPE_VNODE)) {
1588 				ff->ff_file = NULL;
1589 				ff->ff_exclose = false;
1590 				ff->ff_allocated = false;
1591 				closef(fp);
1592 			} else {
1593 				ff->ff_refcnt++;
1594 				fd_close(fd);
1595 			}
1596 		}
1597 		KASSERT(ff->ff_refcnt == 0);
1598 		KASSERT(ff->ff_file == NULL);
1599 		KASSERT(!ff->ff_exclose);
1600 		KASSERT(!ff->ff_allocated);
1601 		if (fd >= NDFDFILE) {
1602 			cv_destroy(&ff->ff_closing);
1603 			kmem_free(ff, sizeof(*ff));
1604 			dt->dt_ff[fd] = NULL;
1605 		}
1606 	}
1607 
1608 	/*
1609 	 * Clean out the descriptor table for the next user and return
1610 	 * to the cache.
1611 	 */
1612 	if (__predict_false(dt != &fdp->fd_dtbuiltin)) {
1613 		fd_dtab_free(fdp->fd_dt);
1614 		/* Otherwise, done above. */
1615 		memset(&fdp->fd_dtbuiltin.dt_ff[NDFDFILE], 0,
1616 		    (NDFILE - NDFDFILE) * sizeof(fdp->fd_dtbuiltin.dt_ff[0]));
1617 		fdp->fd_dt = &fdp->fd_dtbuiltin;
1618 	}
1619 	if (__predict_false(NDHISLOTS(nf) > NDHISLOTS(NDFILE))) {
1620 		KASSERT(fdp->fd_himap != fdp->fd_dhimap);
1621 		KASSERT(fdp->fd_lomap != fdp->fd_dlomap);
1622 		fd_map_free(nf, fdp->fd_lomap, fdp->fd_himap);
1623 	}
1624 	if (__predict_false(fdp->fd_knhash != NULL)) {
1625 		hashdone(fdp->fd_knhash, HASH_LIST, fdp->fd_knhashmask);
1626 		fdp->fd_knhash = NULL;
1627 		fdp->fd_knhashmask = 0;
1628 	} else {
1629 		KASSERT(fdp->fd_knhashmask == 0);
1630 	}
1631 	fdp->fd_dt = &fdp->fd_dtbuiltin;
1632 	fdp->fd_lastkqfile = -1;
1633 	fdp->fd_lastfile = -1;
1634 	fdp->fd_freefile = 0;
1635 	fdp->fd_exclose = false;
1636 	memset(&fdp->fd_startzero, 0, sizeof(*fdp) -
1637 	    offsetof(filedesc_t, fd_startzero));
1638 	fdp->fd_himap = fdp->fd_dhimap;
1639 	fdp->fd_lomap = fdp->fd_dlomap;
1640 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1641 	KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1642 	KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1643 #ifdef DEBUG
1644 	fdp->fd_refcnt = 0; /* see fd_checkmaps */
1645 #endif
1646 	fd_checkmaps(fdp);
1647 	pool_cache_put(filedesc_cache, fdp);
1648 }
1649 
1650 /*
1651  * File Descriptor pseudo-device driver (/dev/fd/).
1652  *
1653  * Opening minor device N dup()s the file (if any) connected to file
1654  * descriptor N belonging to the calling process.  Note that this driver
1655  * consists of only the ``open()'' routine, because all subsequent
1656  * references to this file will be direct to the other driver.
1657  */
1658 static int
1659 filedescopen(dev_t dev, int mode, int type, lwp_t *l)
1660 {
1661 
1662 	/*
1663 	 * XXX Kludge: set dupfd to contain the value of the
1664 	 * the file descriptor being sought for duplication. The error
1665 	 * return ensures that the vnode for this device will be released
1666 	 * by vn_open. Open will detect this special error and take the
1667 	 * actions in fd_dupopen below. Other callers of vn_open or VOP_OPEN
1668 	 * will simply report the error.
1669 	 */
1670 	l->l_dupfd = minor(dev);	/* XXX */
1671 	return EDUPFD;
1672 }
1673 
1674 /*
1675  * Duplicate the specified descriptor to a free descriptor.
1676  *
1677  * old is the original fd.
1678  * moveit is true if we should move rather than duplicate.
1679  * flags are the open flags (converted from O_* to F*).
1680  * newp returns the new fd on success.
1681  *
1682  * These two cases are produced by the EDUPFD and EMOVEFD magic
1683  * errnos, but in the interest of removing that regrettable interface,
1684  * vn_open has been changed to intercept them. Now vn_open returns
1685  * either a vnode or a filehandle, and the filehandle is accompanied
1686  * by a boolean that says whether we should dup (moveit == false) or
1687  * move (moveit == true) the fd.
1688  *
1689  * The dup case is used by /dev/stderr, /proc/self/fd, and such. The
1690  * move case is used by cloner devices that allocate a fd of their
1691  * own (a layering violation that should go away eventually) that
1692  * then needs to be put in the place open() expects it.
1693  */
1694 int
1695 fd_dupopen(int old, bool moveit, int flags, int *newp)
1696 {
1697 	filedesc_t *fdp;
1698 	fdfile_t *ff;
1699 	file_t *fp;
1700 	fdtab_t *dt;
1701 	int error;
1702 
1703 	if ((fp = fd_getfile(old)) == NULL) {
1704 		return EBADF;
1705 	}
1706 	fdp = curlwp->l_fd;
1707 	dt = atomic_load_consume(&fdp->fd_dt);
1708 	ff = dt->dt_ff[old];
1709 
1710 	/*
1711 	 * There are two cases of interest here.
1712 	 *
1713 	 * 1. moveit == false (used to be the EDUPFD magic errno):
1714 	 *    simply dup (old) to file descriptor (new) and return.
1715 	 *
1716 	 * 2. moveit == true (used to be the EMOVEFD magic errno):
1717 	 *    steal away the file structure from (old) and store it in
1718 	 *    (new).  (old) is effectively closed by this operation.
1719 	 */
1720 	if (moveit == false) {
1721 		/*
1722 		 * Check that the mode the file is being opened for is a
1723 		 * subset of the mode of the existing descriptor.
1724 		 */
1725 		if (((flags & (FREAD|FWRITE)) | fp->f_flag) != fp->f_flag) {
1726 			error = EACCES;
1727 			goto out;
1728 		}
1729 
1730 		/* Copy it. */
1731 		error = fd_dup(fp, 0, newp, ff->ff_exclose);
1732 	} else {
1733 		/* Copy it. */
1734 		error = fd_dup(fp, 0, newp, ff->ff_exclose);
1735 		if (error != 0) {
1736 			goto out;
1737 		}
1738 
1739 		/* Steal away the file pointer from 'old'. */
1740 		(void)fd_close(old);
1741 		return 0;
1742 	}
1743 
1744 out:
1745 	fd_putfile(old);
1746 	return error;
1747 }
1748 
1749 /*
1750  * Close open files on exec.
1751  */
1752 void
1753 fd_closeexec(void)
1754 {
1755 	proc_t *p;
1756 	filedesc_t *fdp;
1757 	fdfile_t *ff;
1758 	lwp_t *l;
1759 	fdtab_t *dt;
1760 	int fd;
1761 
1762 	l = curlwp;
1763 	p = l->l_proc;
1764 	fdp = p->p_fd;
1765 
1766 	if (fdp->fd_refcnt > 1) {
1767 		fdp = fd_copy();
1768 		fd_free();
1769 		p->p_fd = fdp;
1770 		l->l_fd = fdp;
1771 	}
1772 	if (!fdp->fd_exclose) {
1773 		return;
1774 	}
1775 	fdp->fd_exclose = false;
1776 	dt = atomic_load_consume(&fdp->fd_dt);
1777 
1778 	for (fd = 0; fd <= fdp->fd_lastfile; fd++) {
1779 		if ((ff = dt->dt_ff[fd]) == NULL) {
1780 			KASSERT(fd >= NDFDFILE);
1781 			continue;
1782 		}
1783 		KASSERT(fd >= NDFDFILE ||
1784 		    ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1785 		if (ff->ff_file == NULL)
1786 			continue;
1787 		if (ff->ff_exclose) {
1788 			/*
1789 			 * We need a reference to close the file.
1790 			 * No other threads can see the fdfile_t at
1791 			 * this point, so don't bother locking.
1792 			 */
1793 			KASSERT((ff->ff_refcnt & FR_CLOSING) == 0);
1794 			ff->ff_refcnt++;
1795 			fd_close(fd);
1796 		}
1797 	}
1798 }
1799 
1800 /*
1801  * Sets descriptor owner. If the owner is a process, 'pgid'
1802  * is set to positive value, process ID. If the owner is process group,
1803  * 'pgid' is set to -pg_id.
1804  */
1805 int
1806 fsetown(pid_t *pgid, u_long cmd, const void *data)
1807 {
1808 	pid_t id = *(const pid_t *)data;
1809 	int error;
1810 
1811 	if (id == INT_MIN)
1812 		return EINVAL;
1813 
1814 	switch (cmd) {
1815 	case TIOCSPGRP:
1816 		if (id < 0)
1817 			return EINVAL;
1818 		id = -id;
1819 		break;
1820 	default:
1821 		break;
1822 	}
1823 	if (id > 0) {
1824 		mutex_enter(&proc_lock);
1825 		error = proc_find(id) ? 0 : ESRCH;
1826 		mutex_exit(&proc_lock);
1827 	} else if (id < 0) {
1828 		error = pgid_in_session(curproc, -id);
1829 	} else {
1830 		error = 0;
1831 	}
1832 	if (!error) {
1833 		*pgid = id;
1834 	}
1835 	return error;
1836 }
1837 
1838 void
1839 fd_set_exclose(struct lwp *l, int fd, bool exclose)
1840 {
1841 	filedesc_t *fdp = l->l_fd;
1842 	fdfile_t *ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
1843 
1844 	ff->ff_exclose = exclose;
1845 	if (exclose)
1846 		fdp->fd_exclose = true;
1847 }
1848 
1849 /*
1850  * Return descriptor owner information. If the value is positive,
1851  * it's process ID. If it's negative, it's process group ID and
1852  * needs the sign removed before use.
1853  */
1854 int
1855 fgetown(pid_t pgid, u_long cmd, void *data)
1856 {
1857 
1858 	switch (cmd) {
1859 	case TIOCGPGRP:
1860 		*(int *)data = -pgid;
1861 		break;
1862 	default:
1863 		*(int *)data = pgid;
1864 		break;
1865 	}
1866 	return 0;
1867 }
1868 
1869 /*
1870  * Send signal to descriptor owner, either process or process group.
1871  */
1872 void
1873 fownsignal(pid_t pgid, int signo, int code, int band, void *fdescdata)
1874 {
1875 	ksiginfo_t ksi;
1876 
1877 	KASSERT(!cpu_intr_p());
1878 
1879 	if (pgid == 0) {
1880 		return;
1881 	}
1882 
1883 	KSI_INIT(&ksi);
1884 	ksi.ksi_signo = signo;
1885 	ksi.ksi_code = code;
1886 	ksi.ksi_band = band;
1887 
1888 	mutex_enter(&proc_lock);
1889 	if (pgid > 0) {
1890 		struct proc *p1;
1891 
1892 		p1 = proc_find(pgid);
1893 		if (p1 != NULL) {
1894 			kpsignal(p1, &ksi, fdescdata);
1895 		}
1896 	} else {
1897 		struct pgrp *pgrp;
1898 
1899 		KASSERT(pgid < 0);
1900 		pgrp = pgrp_find(-pgid);
1901 		if (pgrp != NULL) {
1902 			kpgsignal(pgrp, &ksi, fdescdata, 0);
1903 		}
1904 	}
1905 	mutex_exit(&proc_lock);
1906 }
1907 
1908 int
1909 fd_clone(file_t *fp, unsigned fd, int flag, const struct fileops *fops,
1910 	 void *data)
1911 {
1912 	fdfile_t *ff;
1913 	filedesc_t *fdp;
1914 
1915 	fp->f_flag = flag & FMASK;
1916 	fdp = curproc->p_fd;
1917 	ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
1918 	KASSERT(ff != NULL);
1919 	ff->ff_exclose = (flag & O_CLOEXEC) != 0;
1920 	fp->f_type = DTYPE_MISC;
1921 	fp->f_ops = fops;
1922 	fp->f_data = data;
1923 	curlwp->l_dupfd = fd;
1924 	fd_affix(curproc, fp, fd);
1925 
1926 	return EMOVEFD;
1927 }
1928 
1929 int
1930 fnullop_fcntl(file_t *fp, u_int cmd, void *data)
1931 {
1932 
1933 	if (cmd == F_SETFL)
1934 		return 0;
1935 
1936 	return EOPNOTSUPP;
1937 }
1938 
1939 int
1940 fnullop_poll(file_t *fp, int which)
1941 {
1942 
1943 	return 0;
1944 }
1945 
1946 int
1947 fnullop_kqfilter(file_t *fp, struct knote *kn)
1948 {
1949 
1950 	return EOPNOTSUPP;
1951 }
1952 
1953 void
1954 fnullop_restart(file_t *fp)
1955 {
1956 
1957 }
1958 
1959 int
1960 fbadop_read(file_t *fp, off_t *offset, struct uio *uio,
1961 	    kauth_cred_t cred, int flags)
1962 {
1963 
1964 	return EOPNOTSUPP;
1965 }
1966 
1967 int
1968 fbadop_write(file_t *fp, off_t *offset, struct uio *uio,
1969 	     kauth_cred_t cred, int flags)
1970 {
1971 
1972 	return EOPNOTSUPP;
1973 }
1974 
1975 int
1976 fbadop_ioctl(file_t *fp, u_long com, void *data)
1977 {
1978 
1979 	return EOPNOTSUPP;
1980 }
1981 
1982 int
1983 fbadop_stat(file_t *fp, struct stat *sb)
1984 {
1985 
1986 	return EOPNOTSUPP;
1987 }
1988 
1989 int
1990 fbadop_close(file_t *fp)
1991 {
1992 
1993 	return EOPNOTSUPP;
1994 }
1995 
1996 /*
1997  * sysctl routines pertaining to file descriptors
1998  */
1999 
2000 /* Initialized in sysctl_init() for now... */
2001 extern kmutex_t sysctl_file_marker_lock;
2002 static u_int sysctl_file_marker = 1;
2003 
2004 /*
2005  * Expects to be called with proc_lock and sysctl_file_marker_lock locked.
2006  */
2007 static void
2008 sysctl_file_marker_reset(void)
2009 {
2010 	struct proc *p;
2011 
2012 	PROCLIST_FOREACH(p, &allproc) {
2013 		struct filedesc *fd = p->p_fd;
2014 		fdtab_t *dt;
2015 		u_int i;
2016 
2017 		mutex_enter(&fd->fd_lock);
2018 		dt = fd->fd_dt;
2019 		for (i = 0; i < dt->dt_nfiles; i++) {
2020 			struct file *fp;
2021 			fdfile_t *ff;
2022 
2023 			if ((ff = dt->dt_ff[i]) == NULL) {
2024 				continue;
2025 			}
2026 			if ((fp = atomic_load_consume(&ff->ff_file)) == NULL) {
2027 				continue;
2028 			}
2029 			fp->f_marker = 0;
2030 		}
2031 		mutex_exit(&fd->fd_lock);
2032 	}
2033 }
2034 
2035 /*
2036  * sysctl helper routine for kern.file pseudo-subtree.
2037  */
2038 static int
2039 sysctl_kern_file(SYSCTLFN_ARGS)
2040 {
2041 	const bool allowaddr = get_expose_address(curproc);
2042 	struct filelist flist;
2043 	int error;
2044 	size_t buflen;
2045 	struct file *fp, fbuf;
2046 	char *start, *where;
2047 	struct proc *p;
2048 
2049 	start = where = oldp;
2050 	buflen = *oldlenp;
2051 
2052 	if (where == NULL) {
2053 		/*
2054 		 * overestimate by 10 files
2055 		 */
2056 		*oldlenp = sizeof(filehead) + (nfiles + 10) *
2057 		    sizeof(struct file);
2058 		return 0;
2059 	}
2060 
2061 	/*
2062 	 * first sysctl_copyout filehead
2063 	 */
2064 	if (buflen < sizeof(filehead)) {
2065 		*oldlenp = 0;
2066 		return 0;
2067 	}
2068 	sysctl_unlock();
2069 	if (allowaddr) {
2070 		memcpy(&flist, &filehead, sizeof(flist));
2071 	} else {
2072 		memset(&flist, 0, sizeof(flist));
2073 	}
2074 	error = sysctl_copyout(l, &flist, where, sizeof(flist));
2075 	if (error) {
2076 		sysctl_relock();
2077 		return error;
2078 	}
2079 	buflen -= sizeof(flist);
2080 	where += sizeof(flist);
2081 
2082 	/*
2083 	 * followed by an array of file structures
2084 	 */
2085 	mutex_enter(&sysctl_file_marker_lock);
2086 	mutex_enter(&proc_lock);
2087 	PROCLIST_FOREACH(p, &allproc) {
2088 		struct filedesc *fd;
2089 		fdtab_t *dt;
2090 		u_int i;
2091 
2092 		if (p->p_stat == SIDL) {
2093 			/* skip embryonic processes */
2094 			continue;
2095 		}
2096 		mutex_enter(p->p_lock);
2097 		error = kauth_authorize_process(l->l_cred,
2098 		    KAUTH_PROCESS_CANSEE, p,
2099 		    KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2100 		    NULL, NULL);
2101 		mutex_exit(p->p_lock);
2102 		if (error != 0) {
2103 			/*
2104 			 * Don't leak kauth retval if we're silently
2105 			 * skipping this entry.
2106 			 */
2107 			error = 0;
2108 			continue;
2109 		}
2110 
2111 		/*
2112 		 * Grab a hold on the process.
2113 		 */
2114 		if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2115 			continue;
2116 		}
2117 		mutex_exit(&proc_lock);
2118 
2119 		fd = p->p_fd;
2120 		mutex_enter(&fd->fd_lock);
2121 		dt = fd->fd_dt;
2122 		for (i = 0; i < dt->dt_nfiles; i++) {
2123 			fdfile_t *ff;
2124 
2125 			if ((ff = dt->dt_ff[i]) == NULL) {
2126 				continue;
2127 			}
2128 			if ((fp = atomic_load_consume(&ff->ff_file)) == NULL) {
2129 				continue;
2130 			}
2131 
2132 			mutex_enter(&fp->f_lock);
2133 
2134 			if ((fp->f_count == 0) ||
2135 			    (fp->f_marker == sysctl_file_marker)) {
2136 				mutex_exit(&fp->f_lock);
2137 				continue;
2138 			}
2139 
2140 			/* Check that we have enough space. */
2141 			if (buflen < sizeof(struct file)) {
2142 				*oldlenp = where - start;
2143 				mutex_exit(&fp->f_lock);
2144 				error = ENOMEM;
2145 				break;
2146 			}
2147 
2148 			fill_file(&fbuf, fp);
2149 			mutex_exit(&fp->f_lock);
2150 			error = sysctl_copyout(l, &fbuf, where, sizeof(fbuf));
2151 			if (error) {
2152 				break;
2153 			}
2154 			buflen -= sizeof(struct file);
2155 			where += sizeof(struct file);
2156 
2157 			fp->f_marker = sysctl_file_marker;
2158 		}
2159 		mutex_exit(&fd->fd_lock);
2160 
2161 		/*
2162 		 * Release reference to process.
2163 		 */
2164 		mutex_enter(&proc_lock);
2165 		rw_exit(&p->p_reflock);
2166 
2167 		if (error)
2168 			break;
2169 	}
2170 
2171 	sysctl_file_marker++;
2172 	/* Reset all markers if wrapped. */
2173 	if (sysctl_file_marker == 0) {
2174 		sysctl_file_marker_reset();
2175 		sysctl_file_marker++;
2176 	}
2177 
2178 	mutex_exit(&proc_lock);
2179 	mutex_exit(&sysctl_file_marker_lock);
2180 
2181 	*oldlenp = where - start;
2182 	sysctl_relock();
2183 	return error;
2184 }
2185 
2186 /*
2187  * sysctl helper function for kern.file2
2188  */
2189 static int
2190 sysctl_kern_file2(SYSCTLFN_ARGS)
2191 {
2192 	struct proc *p;
2193 	struct file *fp;
2194 	struct filedesc *fd;
2195 	struct kinfo_file kf;
2196 	char *dp;
2197 	u_int i, op;
2198 	size_t len, needed, elem_size, out_size;
2199 	int error, arg, elem_count;
2200 	fdfile_t *ff;
2201 	fdtab_t *dt;
2202 
2203 	if (namelen == 1 && name[0] == CTL_QUERY)
2204 		return sysctl_query(SYSCTLFN_CALL(rnode));
2205 
2206 	if (namelen != 4)
2207 		return EINVAL;
2208 
2209 	error = 0;
2210 	dp = oldp;
2211 	len = (oldp != NULL) ? *oldlenp : 0;
2212 	op = name[0];
2213 	arg = name[1];
2214 	elem_size = name[2];
2215 	elem_count = name[3];
2216 	out_size = MIN(sizeof(kf), elem_size);
2217 	needed = 0;
2218 
2219 	if (elem_size < 1 || elem_count < 0)
2220 		return EINVAL;
2221 
2222 	switch (op) {
2223 	case KERN_FILE_BYFILE:
2224 	case KERN_FILE_BYPID:
2225 		/*
2226 		 * We're traversing the process list in both cases; the BYFILE
2227 		 * case does additional work of keeping track of files already
2228 		 * looked at.
2229 		 */
2230 
2231 		/* doesn't use arg so it must be zero */
2232 		if ((op == KERN_FILE_BYFILE) && (arg != 0))
2233 			return EINVAL;
2234 
2235 		if ((op == KERN_FILE_BYPID) && (arg < -1))
2236 			/* -1 means all processes */
2237 			return EINVAL;
2238 
2239 		sysctl_unlock();
2240 		if (op == KERN_FILE_BYFILE)
2241 			mutex_enter(&sysctl_file_marker_lock);
2242 		mutex_enter(&proc_lock);
2243 		PROCLIST_FOREACH(p, &allproc) {
2244 			if (p->p_stat == SIDL) {
2245 				/* skip embryonic processes */
2246 				continue;
2247 			}
2248 			if (arg > 0 && p->p_pid != arg) {
2249 				/* pick only the one we want */
2250 				/* XXX want 0 to mean "kernel files" */
2251 				continue;
2252 			}
2253 			mutex_enter(p->p_lock);
2254 			error = kauth_authorize_process(l->l_cred,
2255 			    KAUTH_PROCESS_CANSEE, p,
2256 			    KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2257 			    NULL, NULL);
2258 			mutex_exit(p->p_lock);
2259 			if (error != 0) {
2260 				/*
2261 				 * Don't leak kauth retval if we're silently
2262 				 * skipping this entry.
2263 				 */
2264 				error = 0;
2265 				continue;
2266 			}
2267 
2268 			/*
2269 			 * Grab a hold on the process.
2270 			 */
2271 			if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2272 				continue;
2273 			}
2274 			mutex_exit(&proc_lock);
2275 
2276 			fd = p->p_fd;
2277 			mutex_enter(&fd->fd_lock);
2278 			dt = fd->fd_dt;
2279 			for (i = 0; i < dt->dt_nfiles; i++) {
2280 				if ((ff = dt->dt_ff[i]) == NULL) {
2281 					continue;
2282 				}
2283 				if ((fp = atomic_load_consume(&ff->ff_file)) ==
2284 				    NULL) {
2285 					continue;
2286 				}
2287 
2288 				if ((op == KERN_FILE_BYFILE) &&
2289 				    (fp->f_marker == sysctl_file_marker)) {
2290 					continue;
2291 				}
2292 				if (len >= elem_size && elem_count > 0) {
2293 					mutex_enter(&fp->f_lock);
2294 					fill_file2(&kf, fp, ff, i, p->p_pid);
2295 					mutex_exit(&fp->f_lock);
2296 					mutex_exit(&fd->fd_lock);
2297 					error = sysctl_copyout(l,
2298 					    &kf, dp, out_size);
2299 					mutex_enter(&fd->fd_lock);
2300 					if (error)
2301 						break;
2302 					dp += elem_size;
2303 					len -= elem_size;
2304 				}
2305 				if (op == KERN_FILE_BYFILE)
2306 					fp->f_marker = sysctl_file_marker;
2307 				needed += elem_size;
2308 				if (elem_count > 0 && elem_count != INT_MAX)
2309 					elem_count--;
2310 			}
2311 			mutex_exit(&fd->fd_lock);
2312 
2313 			/*
2314 			 * Release reference to process.
2315 			 */
2316 			mutex_enter(&proc_lock);
2317 			rw_exit(&p->p_reflock);
2318 		}
2319 		if (op == KERN_FILE_BYFILE) {
2320 			sysctl_file_marker++;
2321 
2322 			/* Reset all markers if wrapped. */
2323 			if (sysctl_file_marker == 0) {
2324 				sysctl_file_marker_reset();
2325 				sysctl_file_marker++;
2326 			}
2327 		}
2328 		mutex_exit(&proc_lock);
2329 		if (op == KERN_FILE_BYFILE)
2330 			mutex_exit(&sysctl_file_marker_lock);
2331 		sysctl_relock();
2332 		break;
2333 	default:
2334 		return EINVAL;
2335 	}
2336 
2337 	if (oldp == NULL)
2338 		needed += KERN_FILESLOP * elem_size;
2339 	*oldlenp = needed;
2340 
2341 	return error;
2342 }
2343 
2344 static void
2345 fill_file(struct file *fp, const struct file *fpsrc)
2346 {
2347 	const bool allowaddr = get_expose_address(curproc);
2348 
2349 	memset(fp, 0, sizeof(*fp));
2350 
2351 	fp->f_offset = fpsrc->f_offset;
2352 	COND_SET_PTR(fp->f_cred, fpsrc->f_cred, allowaddr);
2353 	COND_SET_CPTR(fp->f_ops, fpsrc->f_ops, allowaddr);
2354 	COND_SET_STRUCT(fp->f_undata, fpsrc->f_undata, allowaddr);
2355 	COND_SET_STRUCT(fp->f_list, fpsrc->f_list, allowaddr);
2356 	fp->f_flag = fpsrc->f_flag;
2357 	fp->f_marker = fpsrc->f_marker;
2358 	fp->f_type = fpsrc->f_type;
2359 	fp->f_advice = fpsrc->f_advice;
2360 	fp->f_count = fpsrc->f_count;
2361 	fp->f_msgcount = fpsrc->f_msgcount;
2362 	fp->f_unpcount = fpsrc->f_unpcount;
2363 	COND_SET_STRUCT(fp->f_unplist, fpsrc->f_unplist, allowaddr);
2364 }
2365 
2366 static void
2367 fill_file2(struct kinfo_file *kp, const file_t *fp, const fdfile_t *ff,
2368 	  int i, pid_t pid)
2369 {
2370 	const bool allowaddr = get_expose_address(curproc);
2371 
2372 	memset(kp, 0, sizeof(*kp));
2373 
2374 	COND_SET_VALUE(kp->ki_fileaddr, PTRTOUINT64(fp), allowaddr);
2375 	kp->ki_flag =		fp->f_flag;
2376 	kp->ki_iflags =		0;
2377 	kp->ki_ftype =		fp->f_type;
2378 	kp->ki_count =		fp->f_count;
2379 	kp->ki_msgcount =	fp->f_msgcount;
2380 	COND_SET_VALUE(kp->ki_fucred, PTRTOUINT64(fp->f_cred), allowaddr);
2381 	kp->ki_fuid =		kauth_cred_geteuid(fp->f_cred);
2382 	kp->ki_fgid =		kauth_cred_getegid(fp->f_cred);
2383 	COND_SET_VALUE(kp->ki_fops, PTRTOUINT64(fp->f_ops), allowaddr);
2384 	kp->ki_foffset =	fp->f_offset;
2385 	COND_SET_VALUE(kp->ki_fdata, PTRTOUINT64(fp->f_data), allowaddr);
2386 
2387 	/* vnode information to glue this file to something */
2388 	if (fp->f_type == DTYPE_VNODE) {
2389 		struct vnode *vp = fp->f_vnode;
2390 
2391 		COND_SET_VALUE(kp->ki_vun, PTRTOUINT64(vp->v_un.vu_socket),
2392 		    allowaddr);
2393 		kp->ki_vsize =	vp->v_size;
2394 		kp->ki_vtype =	vp->v_type;
2395 		kp->ki_vtag =	vp->v_tag;
2396 		COND_SET_VALUE(kp->ki_vdata, PTRTOUINT64(vp->v_data),
2397 		    allowaddr);
2398 	}
2399 
2400 	/* process information when retrieved via KERN_FILE_BYPID */
2401 	if (ff != NULL) {
2402 		kp->ki_pid =		pid;
2403 		kp->ki_fd =		i;
2404 		kp->ki_ofileflags =	ff->ff_exclose;
2405 		kp->ki_usecount =	ff->ff_refcnt;
2406 	}
2407 }
2408