xref: /netbsd-src/sys/kern/kern_descrip.c (revision 32d1c65c71fbdb65a012e8392a62a757dd6853e9)
1 /*	$NetBSD: kern_descrip.c,v 1.264 2024/11/10 00:11:43 kre 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 1.264 2024/11/10 00:11:43 kre 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 	int error;
751 
752 	while ((error = fd_alloc(p, minfd, newp)) != 0) {
753 		if (error != ENOSPC) {
754 			return error;
755 		}
756 		fd_tryexpand(p);
757 	}
758 
759 	fd_set_exclose(curlwp, *newp, exclose);
760 	fd_affix(p, fp, *newp);
761 	return 0;
762 }
763 
764 /*
765  * dup2 operation.
766  */
767 int
768 fd_dup2(file_t *fp, unsigned newfd, int flags)
769 {
770 	filedesc_t *fdp = curlwp->l_fd;
771 	fdfile_t *ff;
772 	fdtab_t *dt;
773 
774 	if (flags & ~(O_CLOEXEC|O_NONBLOCK|O_NOSIGPIPE))
775 		return EINVAL;
776 	/*
777 	 * Ensure there are enough slots in the descriptor table,
778 	 * and allocate an fdfile_t up front in case we need it.
779 	 */
780 	while (newfd >= atomic_load_consume(&fdp->fd_dt)->dt_nfiles) {
781 		fd_tryexpand(curproc);
782 	}
783 	ff = kmem_alloc(sizeof(*ff), KM_SLEEP);
784 	fdfile_ctor(ff);
785 
786 	/*
787 	 * If there is already a file open, close it.  If the file is
788 	 * half open, wait for it to be constructed before closing it.
789 	 * XXX Potential for deadlock here?
790 	 */
791 	mutex_enter(&fdp->fd_lock);
792 	while (fd_isused(fdp, newfd)) {
793 		mutex_exit(&fdp->fd_lock);
794 		if (fd_getfile(newfd) != NULL) {
795 			(void)fd_close(newfd);
796 		} else {
797 			/*
798 			 * Crummy, but unlikely to happen.
799 			 * Can occur if we interrupt another
800 			 * thread while it is opening a file.
801 			 */
802 			kpause("dup2", false, 1, NULL);
803 		}
804 		mutex_enter(&fdp->fd_lock);
805 	}
806 	dt = fdp->fd_dt;
807 	if (dt->dt_ff[newfd] == NULL) {
808 		KASSERT(newfd >= NDFDFILE);
809 		dt->dt_ff[newfd] = ff;
810 		ff = NULL;
811 	}
812 	fd_used(fdp, newfd);
813 	mutex_exit(&fdp->fd_lock);
814 
815 	fd_set_exclose(curlwp, newfd, (flags & O_CLOEXEC) != 0);
816 	fp->f_flag |= flags & (FNONBLOCK|FNOSIGPIPE);
817 	/* Slot is now allocated.  Insert copy of the file. */
818 	fd_affix(curproc, fp, newfd);
819 	if (ff != NULL) {
820 		cv_destroy(&ff->ff_closing);
821 		kmem_free(ff, sizeof(*ff));
822 	}
823 	return 0;
824 }
825 
826 /*
827  * Drop reference to a file structure.
828  */
829 int
830 closef(file_t *fp)
831 {
832 	struct flock lf;
833 	int error;
834 
835 	/*
836 	 * Drop reference.  If referenced elsewhere it's still open
837 	 * and we have nothing more to do.
838 	 */
839 	mutex_enter(&fp->f_lock);
840 	KASSERT(fp->f_count > 0);
841 	if (--fp->f_count > 0) {
842 		mutex_exit(&fp->f_lock);
843 		return 0;
844 	}
845 	KASSERT(fp->f_count == 0);
846 	mutex_exit(&fp->f_lock);
847 
848 	/* We held the last reference - release locks, close and free. */
849 	if (fp->f_ops->fo_advlock == NULL) {
850 		KASSERT((fp->f_flag & FHASLOCK) == 0);
851 	} else if (fp->f_flag & FHASLOCK) {
852 		lf.l_whence = SEEK_SET;
853 		lf.l_start = 0;
854 		lf.l_len = 0;
855 		lf.l_type = F_UNLCK;
856 		(void)(*fp->f_ops->fo_advlock)(fp, fp, F_UNLCK, &lf, F_FLOCK);
857 	}
858 	if (fp->f_ops != NULL) {
859 		error = (*fp->f_ops->fo_close)(fp);
860 	} else {
861 		error = 0;
862 	}
863 	KASSERT(fp->f_count == 0);
864 	KASSERT(fp->f_cred != NULL);
865 	pool_cache_put(file_cache, fp);
866 
867 	return error;
868 }
869 
870 /*
871  * Allocate a file descriptor for the process.
872  *
873  * Future idea for experimentation: replace all of this with radixtree.
874  */
875 int
876 fd_alloc(proc_t *p, int want, int *result)
877 {
878 	filedesc_t *fdp = p->p_fd;
879 	int i, lim, last, error, hi;
880 	u_int off;
881 	fdtab_t *dt;
882 
883 	KASSERT(p == curproc || p == &proc0);
884 
885 	/*
886 	 * Search for a free descriptor starting at the higher
887 	 * of want or fd_freefile.
888 	 */
889 	mutex_enter(&fdp->fd_lock);
890 	fd_checkmaps(fdp);
891 	dt = fdp->fd_dt;
892 	KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
893 	lim = uimin((int)p->p_rlimit[RLIMIT_NOFILE].rlim_cur, maxfiles);
894 	last = uimin(dt->dt_nfiles, lim);
895 
896 	for (;;) {
897 		if ((i = want) < fdp->fd_freefile)
898 			i = fdp->fd_freefile;
899 		off = i >> NDENTRYSHIFT;
900 		hi = fd_next_zero(fdp, fdp->fd_himap, off,
901 		    (last + NDENTRIES - 1) >> NDENTRYSHIFT);
902 		if (hi == -1)
903 			break;
904 		i = fd_next_zero(fdp, &fdp->fd_lomap[hi],
905 		    hi > off ? 0 : i & NDENTRYMASK, NDENTRIES);
906 		if (i == -1) {
907 			/*
908 			 * Free file descriptor in this block was
909 			 * below want, try again with higher want.
910 			 */
911 			want = (hi + 1) << NDENTRYSHIFT;
912 			continue;
913 		}
914 		i += (hi << NDENTRYSHIFT);
915 		if (i >= last) {
916 			break;
917 		}
918 		if (dt->dt_ff[i] == NULL) {
919 			KASSERT(i >= NDFDFILE);
920 			dt->dt_ff[i] = kmem_alloc(sizeof(fdfile_t), KM_SLEEP);
921 			fdfile_ctor(dt->dt_ff[i]);
922 		}
923 		KASSERT(dt->dt_ff[i]->ff_file == NULL);
924 		fd_used(fdp, i);
925 		if (want <= fdp->fd_freefile) {
926 			fdp->fd_freefile = i;
927 		}
928 		*result = i;
929 		KASSERT(i >= NDFDFILE ||
930 		    dt->dt_ff[i] == (fdfile_t *)fdp->fd_dfdfile[i]);
931 		fd_checkmaps(fdp);
932 		mutex_exit(&fdp->fd_lock);
933 		return 0;
934 	}
935 
936 	/* No space in current array.  Let the caller expand and retry. */
937 	error = (dt->dt_nfiles >= lim) ? EMFILE : ENOSPC;
938 	mutex_exit(&fdp->fd_lock);
939 	return error;
940 }
941 
942 /*
943  * Allocate memory for a descriptor table.
944  */
945 static fdtab_t *
946 fd_dtab_alloc(int n)
947 {
948 	fdtab_t *dt;
949 	size_t sz;
950 
951 	KASSERT(n > NDFILE);
952 
953 	sz = sizeof(*dt) + (n - NDFILE) * sizeof(dt->dt_ff[0]);
954 	dt = kmem_alloc(sz, KM_SLEEP);
955 #ifdef DIAGNOSTIC
956 	memset(dt, 0xff, sz);
957 #endif
958 	dt->dt_nfiles = n;
959 	dt->dt_link = NULL;
960 	return dt;
961 }
962 
963 /*
964  * Free a descriptor table, and all tables linked for deferred free.
965  */
966 static void
967 fd_dtab_free(fdtab_t *dt)
968 {
969 	fdtab_t *next;
970 	size_t sz;
971 
972 	do {
973 		next = dt->dt_link;
974 		KASSERT(dt->dt_nfiles > NDFILE);
975 		sz = sizeof(*dt) +
976 		    (dt->dt_nfiles - NDFILE) * sizeof(dt->dt_ff[0]);
977 #ifdef DIAGNOSTIC
978 		memset(dt, 0xff, sz);
979 #endif
980 		kmem_free(dt, sz);
981 		dt = next;
982 	} while (dt != NULL);
983 }
984 
985 /*
986  * Allocate descriptor bitmap.
987  */
988 static void
989 fd_map_alloc(int n, uint32_t **lo, uint32_t **hi)
990 {
991 	uint8_t *ptr;
992 	size_t szlo, szhi;
993 
994 	KASSERT(n > NDENTRIES);
995 
996 	szlo = NDLOSLOTS(n) * sizeof(uint32_t);
997 	szhi = NDHISLOTS(n) * sizeof(uint32_t);
998 	ptr = kmem_alloc(szlo + szhi, KM_SLEEP);
999 	*lo = (uint32_t *)ptr;
1000 	*hi = (uint32_t *)(ptr + szlo);
1001 }
1002 
1003 /*
1004  * Free descriptor bitmap.
1005  */
1006 static void
1007 fd_map_free(int n, uint32_t *lo, uint32_t *hi)
1008 {
1009 	size_t szlo, szhi;
1010 
1011 	KASSERT(n > NDENTRIES);
1012 
1013 	szlo = NDLOSLOTS(n) * sizeof(uint32_t);
1014 	szhi = NDHISLOTS(n) * sizeof(uint32_t);
1015 	KASSERT(hi == (uint32_t *)((uint8_t *)lo + szlo));
1016 	kmem_free(lo, szlo + szhi);
1017 }
1018 
1019 /*
1020  * Expand a process' descriptor table.
1021  */
1022 void
1023 fd_tryexpand(proc_t *p)
1024 {
1025 	filedesc_t *fdp;
1026 	int i, numfiles, oldnfiles;
1027 	fdtab_t *newdt, *dt;
1028 	uint32_t *newhimap, *newlomap;
1029 
1030 	KASSERT(p == curproc || p == &proc0);
1031 
1032 	fdp = p->p_fd;
1033 	newhimap = NULL;
1034 	newlomap = NULL;
1035 	oldnfiles = atomic_load_consume(&fdp->fd_dt)->dt_nfiles;
1036 
1037 	if (oldnfiles < NDEXTENT)
1038 		numfiles = NDEXTENT;
1039 	else
1040 		numfiles = 2 * oldnfiles;
1041 
1042 	newdt = fd_dtab_alloc(numfiles);
1043 	if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1044 		fd_map_alloc(numfiles, &newlomap, &newhimap);
1045 	}
1046 
1047 	mutex_enter(&fdp->fd_lock);
1048 	dt = fdp->fd_dt;
1049 	KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1050 	if (dt->dt_nfiles != oldnfiles) {
1051 		/* fdp changed; caller must retry */
1052 		mutex_exit(&fdp->fd_lock);
1053 		fd_dtab_free(newdt);
1054 		if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1055 			fd_map_free(numfiles, newlomap, newhimap);
1056 		}
1057 		return;
1058 	}
1059 
1060 	/* Copy the existing descriptor table and zero the new portion. */
1061 	i = sizeof(fdfile_t *) * oldnfiles;
1062 	memcpy(newdt->dt_ff, dt->dt_ff, i);
1063 	memset((uint8_t *)newdt->dt_ff + i, 0,
1064 	    numfiles * sizeof(fdfile_t *) - i);
1065 
1066 	/*
1067 	 * Link old descriptor array into list to be discarded.  We defer
1068 	 * freeing until the last reference to the descriptor table goes
1069 	 * away (usually process exit).  This allows us to do lockless
1070 	 * lookups in fd_getfile().
1071 	 */
1072 	if (oldnfiles > NDFILE) {
1073 		if (fdp->fd_refcnt > 1) {
1074 			newdt->dt_link = dt;
1075 		} else {
1076 			fd_dtab_free(dt);
1077 		}
1078 	}
1079 
1080 	if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1081 		i = NDHISLOTS(oldnfiles) * sizeof(uint32_t);
1082 		memcpy(newhimap, fdp->fd_himap, i);
1083 		memset((uint8_t *)newhimap + i, 0,
1084 		    NDHISLOTS(numfiles) * sizeof(uint32_t) - i);
1085 
1086 		i = NDLOSLOTS(oldnfiles) * sizeof(uint32_t);
1087 		memcpy(newlomap, fdp->fd_lomap, i);
1088 		memset((uint8_t *)newlomap + i, 0,
1089 		    NDLOSLOTS(numfiles) * sizeof(uint32_t) - i);
1090 
1091 		if (NDHISLOTS(oldnfiles) > NDHISLOTS(NDFILE)) {
1092 			fd_map_free(oldnfiles, fdp->fd_lomap, fdp->fd_himap);
1093 		}
1094 		fdp->fd_himap = newhimap;
1095 		fdp->fd_lomap = newlomap;
1096 	}
1097 
1098 	/*
1099 	 * All other modifications must become globally visible before
1100 	 * the change to fd_dt.  See fd_getfile().
1101 	 */
1102 	atomic_store_release(&fdp->fd_dt, newdt);
1103 	KASSERT(newdt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1104 	fd_checkmaps(fdp);
1105 	mutex_exit(&fdp->fd_lock);
1106 }
1107 
1108 /*
1109  * Create a new open file structure and allocate a file descriptor
1110  * for the current process.
1111  */
1112 int
1113 fd_allocfile(file_t **resultfp, int *resultfd)
1114 {
1115 	proc_t *p = curproc;
1116 	kauth_cred_t cred;
1117 	file_t *fp;
1118 	int error;
1119 
1120 	while ((error = fd_alloc(p, 0, resultfd)) != 0) {
1121 		if (error != ENOSPC) {
1122 			return error;
1123 		}
1124 		fd_tryexpand(p);
1125 	}
1126 
1127 	fp = pool_cache_get(file_cache, PR_WAITOK);
1128 	if (fp == NULL) {
1129 		fd_abort(p, NULL, *resultfd);
1130 		return ENFILE;
1131 	}
1132 	KASSERT(fp->f_count == 0);
1133 	KASSERT(fp->f_msgcount == 0);
1134 	KASSERT(fp->f_unpcount == 0);
1135 
1136 	/* Replace cached credentials if not what we need. */
1137 	cred = curlwp->l_cred;
1138 	if (__predict_false(cred != fp->f_cred)) {
1139 		kauth_cred_free(fp->f_cred);
1140 		fp->f_cred = kauth_cred_hold(cred);
1141 	}
1142 
1143 	/*
1144 	 * Don't allow recycled files to be scanned.
1145 	 * See uipc_usrreq.c.
1146 	 */
1147 	if (__predict_false((fp->f_flag & FSCAN) != 0)) {
1148 		mutex_enter(&fp->f_lock);
1149 		atomic_and_uint(&fp->f_flag, ~FSCAN);
1150 		mutex_exit(&fp->f_lock);
1151 	}
1152 
1153 	fp->f_advice = 0;
1154 	fp->f_offset = 0;
1155 	*resultfp = fp;
1156 
1157 	return 0;
1158 }
1159 
1160 /*
1161  * Successful creation of a new descriptor: make visible to the process.
1162  */
1163 void
1164 fd_affix(proc_t *p, file_t *fp, unsigned fd)
1165 {
1166 	fdfile_t *ff;
1167 	filedesc_t *fdp;
1168 	fdtab_t *dt;
1169 
1170 	KASSERT(p == curproc || p == &proc0);
1171 
1172 	/* Add a reference to the file structure. */
1173 	mutex_enter(&fp->f_lock);
1174 	fp->f_count++;
1175 	mutex_exit(&fp->f_lock);
1176 
1177 	/*
1178 	 * Insert the new file into the descriptor slot.
1179 	 */
1180 	fdp = p->p_fd;
1181 	dt = atomic_load_consume(&fdp->fd_dt);
1182 	ff = dt->dt_ff[fd];
1183 
1184 	KASSERT(ff != NULL);
1185 	KASSERT(ff->ff_file == NULL);
1186 	KASSERT(ff->ff_allocated);
1187 	KASSERT(fd_isused(fdp, fd));
1188 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1189 
1190 	/* No need to lock in order to make file initially visible. */
1191 	atomic_store_release(&ff->ff_file, fp);
1192 }
1193 
1194 /*
1195  * Abort creation of a new descriptor: free descriptor slot and file.
1196  */
1197 void
1198 fd_abort(proc_t *p, file_t *fp, unsigned fd)
1199 {
1200 	filedesc_t *fdp;
1201 	fdfile_t *ff;
1202 
1203 	KASSERT(p == curproc || p == &proc0);
1204 
1205 	fdp = p->p_fd;
1206 	ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
1207 	ff->ff_exclose = false;
1208 
1209 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1210 
1211 	mutex_enter(&fdp->fd_lock);
1212 	KASSERT(fd_isused(fdp, fd));
1213 	fd_unused(fdp, fd);
1214 	mutex_exit(&fdp->fd_lock);
1215 
1216 	if (fp != NULL) {
1217 		KASSERT(fp->f_count == 0);
1218 		KASSERT(fp->f_cred != NULL);
1219 		pool_cache_put(file_cache, fp);
1220 	}
1221 }
1222 
1223 static int
1224 file_ctor(void *arg, void *obj, int flags)
1225 {
1226 	/*
1227 	 * It's easy to exhaust the open file limit on a system with many
1228 	 * CPUs due to caching.  Allow a bit of leeway to reduce the element
1229 	 * of surprise.
1230 	 */
1231 	u_int slop = PCG_NOBJECTS_NORMAL * (ncpu - 1);
1232 	file_t *fp = obj;
1233 
1234 	memset(fp, 0, sizeof(*fp));
1235 
1236 	mutex_enter(&filelist_lock);
1237 	if (__predict_false(nfiles >= slop + maxfiles)) {
1238 		mutex_exit(&filelist_lock);
1239 		tablefull("file", "increase kern.maxfiles or MAXFILES");
1240 		return ENFILE;
1241 	}
1242 	nfiles++;
1243 	LIST_INSERT_HEAD(&filehead, fp, f_list);
1244 	mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1245 	fp->f_cred = kauth_cred_hold(curlwp->l_cred);
1246 	mutex_exit(&filelist_lock);
1247 
1248 	return 0;
1249 }
1250 
1251 static void
1252 file_dtor(void *arg, void *obj)
1253 {
1254 	file_t *fp = obj;
1255 
1256 	mutex_enter(&filelist_lock);
1257 	nfiles--;
1258 	LIST_REMOVE(fp, f_list);
1259 	mutex_exit(&filelist_lock);
1260 
1261 	KASSERT(fp->f_count == 0);
1262 	kauth_cred_free(fp->f_cred);
1263 	mutex_destroy(&fp->f_lock);
1264 }
1265 
1266 static void
1267 fdfile_ctor(fdfile_t *ff)
1268 {
1269 
1270 	memset(ff, 0, sizeof(*ff));
1271 	cv_init(&ff->ff_closing, "fdclose");
1272 }
1273 
1274 static void
1275 fdfile_dtor(fdfile_t *ff)
1276 {
1277 
1278 	cv_destroy(&ff->ff_closing);
1279 }
1280 
1281 file_t *
1282 fgetdummy(void)
1283 {
1284 	file_t *fp;
1285 
1286 	fp = kmem_zalloc(sizeof(*fp), KM_SLEEP);
1287 	mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1288 	return fp;
1289 }
1290 
1291 void
1292 fputdummy(file_t *fp)
1293 {
1294 
1295 	mutex_destroy(&fp->f_lock);
1296 	kmem_free(fp, sizeof(*fp));
1297 }
1298 
1299 /*
1300  * Create an initial filedesc structure.
1301  */
1302 filedesc_t *
1303 fd_init(filedesc_t *fdp)
1304 {
1305 #ifdef DIAGNOSTIC
1306 	unsigned fd;
1307 #endif
1308 
1309 	if (__predict_true(fdp == NULL)) {
1310 		fdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1311 	} else {
1312 		KASSERT(fdp == &filedesc0);
1313 		filedesc_ctor(NULL, fdp, PR_WAITOK);
1314 	}
1315 
1316 #ifdef DIAGNOSTIC
1317 	KASSERT(fdp->fd_lastfile == -1);
1318 	KASSERT(fdp->fd_lastkqfile == -1);
1319 	KASSERT(fdp->fd_knhash == NULL);
1320 	KASSERT(fdp->fd_freefile == 0);
1321 	KASSERT(fdp->fd_exclose == false);
1322 	KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1323 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1324 	for (fd = 0; fd < NDFDFILE; fd++) {
1325 		KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] ==
1326 		    (fdfile_t *)fdp->fd_dfdfile[fd]);
1327 	}
1328 	for (fd = NDFDFILE; fd < NDFILE; fd++) {
1329 		KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] == NULL);
1330 	}
1331 	KASSERT(fdp->fd_himap == fdp->fd_dhimap);
1332 	KASSERT(fdp->fd_lomap == fdp->fd_dlomap);
1333 #endif	/* DIAGNOSTIC */
1334 
1335 	fdp->fd_refcnt = 1;
1336 	fd_checkmaps(fdp);
1337 
1338 	return fdp;
1339 }
1340 
1341 /*
1342  * Initialize a file descriptor table.
1343  */
1344 static int
1345 filedesc_ctor(void *arg, void *obj, int flag)
1346 {
1347 	filedesc_t *fdp = obj;
1348 	fdfile_t **ffp;
1349 	int i;
1350 
1351 	memset(fdp, 0, sizeof(*fdp));
1352 	mutex_init(&fdp->fd_lock, MUTEX_DEFAULT, IPL_NONE);
1353 	fdp->fd_lastfile = -1;
1354 	fdp->fd_lastkqfile = -1;
1355 	fdp->fd_dt = &fdp->fd_dtbuiltin;
1356 	fdp->fd_dtbuiltin.dt_nfiles = NDFILE;
1357 	fdp->fd_himap = fdp->fd_dhimap;
1358 	fdp->fd_lomap = fdp->fd_dlomap;
1359 
1360 	CTASSERT(sizeof(fdp->fd_dfdfile[0]) >= sizeof(fdfile_t));
1361 	for (i = 0, ffp = fdp->fd_dt->dt_ff; i < NDFDFILE; i++, ffp++) {
1362 		fdfile_ctor(*ffp = (fdfile_t *)fdp->fd_dfdfile[i]);
1363 	}
1364 
1365 	return 0;
1366 }
1367 
1368 static void
1369 filedesc_dtor(void *arg, void *obj)
1370 {
1371 	filedesc_t *fdp = obj;
1372 	int i;
1373 
1374 	for (i = 0; i < NDFDFILE; i++) {
1375 		fdfile_dtor((fdfile_t *)fdp->fd_dfdfile[i]);
1376 	}
1377 
1378 	mutex_destroy(&fdp->fd_lock);
1379 }
1380 
1381 /*
1382  * Make p share curproc's filedesc structure.
1383  */
1384 void
1385 fd_share(struct proc *p)
1386 {
1387 	filedesc_t *fdp;
1388 
1389 	fdp = curlwp->l_fd;
1390 	p->p_fd = fdp;
1391 	atomic_inc_uint(&fdp->fd_refcnt);
1392 }
1393 
1394 /*
1395  * Acquire a hold on a filedesc structure.
1396  */
1397 void
1398 fd_hold(lwp_t *l)
1399 {
1400 	filedesc_t *fdp = l->l_fd;
1401 
1402 	atomic_inc_uint(&fdp->fd_refcnt);
1403 }
1404 
1405 /*
1406  * Copy a filedesc structure.
1407  */
1408 filedesc_t *
1409 fd_copy(void)
1410 {
1411 	filedesc_t *newfdp, *fdp;
1412 	fdfile_t *ff, **ffp, **nffp, *ff2;
1413 	int i, j, numfiles, lastfile, newlast;
1414 	file_t *fp;
1415 	fdtab_t *newdt;
1416 
1417 	fdp = curproc->p_fd;
1418 	newfdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1419 	newfdp->fd_refcnt = 1;
1420 
1421 #ifdef DIAGNOSTIC
1422 	KASSERT(newfdp->fd_lastfile == -1);
1423 	KASSERT(newfdp->fd_lastkqfile == -1);
1424 	KASSERT(newfdp->fd_knhash == NULL);
1425 	KASSERT(newfdp->fd_freefile == 0);
1426 	KASSERT(newfdp->fd_exclose == false);
1427 	KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1428 	KASSERT(newfdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1429 	for (i = 0; i < NDFDFILE; i++) {
1430 		KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] ==
1431 		    (fdfile_t *)&newfdp->fd_dfdfile[i]);
1432 	}
1433 	for (i = NDFDFILE; i < NDFILE; i++) {
1434 		KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] == NULL);
1435 	}
1436 #endif	/* DIAGNOSTIC */
1437 
1438 	mutex_enter(&fdp->fd_lock);
1439 	fd_checkmaps(fdp);
1440 	numfiles = fdp->fd_dt->dt_nfiles;
1441 	lastfile = fdp->fd_lastfile;
1442 
1443 	/*
1444 	 * If the number of open files fits in the internal arrays
1445 	 * of the open file structure, use them, otherwise allocate
1446 	 * additional memory for the number of descriptors currently
1447 	 * in use.
1448 	 */
1449 	if (lastfile < NDFILE) {
1450 		i = NDFILE;
1451 		newdt = newfdp->fd_dt;
1452 		KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1453 	} else {
1454 		/*
1455 		 * Compute the smallest multiple of NDEXTENT needed
1456 		 * for the file descriptors currently in use,
1457 		 * allowing the table to shrink.
1458 		 */
1459 		i = numfiles;
1460 		while (i >= 2 * NDEXTENT && i > lastfile * 2) {
1461 			i /= 2;
1462 		}
1463 		KASSERT(i > NDFILE);
1464 		newdt = fd_dtab_alloc(i);
1465 		newfdp->fd_dt = newdt;
1466 		memcpy(newdt->dt_ff, newfdp->fd_dtbuiltin.dt_ff,
1467 		    NDFDFILE * sizeof(fdfile_t **));
1468 		memset(newdt->dt_ff + NDFDFILE, 0,
1469 		    (i - NDFDFILE) * sizeof(fdfile_t **));
1470 	}
1471 	if (NDHISLOTS(i) <= NDHISLOTS(NDFILE)) {
1472 		newfdp->fd_himap = newfdp->fd_dhimap;
1473 		newfdp->fd_lomap = newfdp->fd_dlomap;
1474 	} else {
1475 		fd_map_alloc(i, &newfdp->fd_lomap, &newfdp->fd_himap);
1476 		KASSERT(i >= NDENTRIES * NDENTRIES);
1477 		memset(newfdp->fd_himap, 0, NDHISLOTS(i)*sizeof(uint32_t));
1478 		memset(newfdp->fd_lomap, 0, NDLOSLOTS(i)*sizeof(uint32_t));
1479 	}
1480 	newfdp->fd_freefile = fdp->fd_freefile;
1481 	newfdp->fd_exclose = fdp->fd_exclose;
1482 
1483 	ffp = fdp->fd_dt->dt_ff;
1484 	nffp = newdt->dt_ff;
1485 	newlast = -1;
1486 	for (i = 0; i <= lastfile; i++, ffp++, nffp++) {
1487 		KASSERT(i >= NDFDFILE ||
1488 		    *nffp == (fdfile_t *)newfdp->fd_dfdfile[i]);
1489 		ff = *ffp;
1490 		if (ff == NULL ||
1491 		    (fp = atomic_load_consume(&ff->ff_file)) == NULL) {
1492 			/* Descriptor unused, or descriptor half open. */
1493 			KASSERT(!fd_isused(newfdp, i));
1494 			continue;
1495 		}
1496 		if (__predict_false(fp->f_type == DTYPE_KQUEUE)) {
1497 			/* kqueue descriptors cannot be copied. */
1498 			if (i < newfdp->fd_freefile) {
1499 				newfdp->fd_freefile = i;
1500 			}
1501 			continue;
1502 		}
1503 		/* It's active: add a reference to the file. */
1504 		mutex_enter(&fp->f_lock);
1505 		fp->f_count++;
1506 		mutex_exit(&fp->f_lock);
1507 
1508 		/* Allocate an fdfile_t to represent it. */
1509 		if (i >= NDFDFILE) {
1510 			ff2 = kmem_alloc(sizeof(*ff2), KM_SLEEP);
1511 			fdfile_ctor(ff2);
1512 			*nffp = ff2;
1513 		} else {
1514 			ff2 = newdt->dt_ff[i];
1515 		}
1516 		ff2->ff_file = fp;
1517 		ff2->ff_exclose = ff->ff_exclose;
1518 		ff2->ff_allocated = true;
1519 
1520 		/* Fix up bitmaps. */
1521 		j = i >> NDENTRYSHIFT;
1522 		KASSERT((newfdp->fd_lomap[j] & (1U << (i & NDENTRYMASK))) == 0);
1523 		newfdp->fd_lomap[j] |= 1U << (i & NDENTRYMASK);
1524 		if (__predict_false(newfdp->fd_lomap[j] == ~0)) {
1525 			KASSERT((newfdp->fd_himap[j >> NDENTRYSHIFT] &
1526 			    (1U << (j & NDENTRYMASK))) == 0);
1527 			newfdp->fd_himap[j >> NDENTRYSHIFT] |=
1528 			    1U << (j & NDENTRYMASK);
1529 		}
1530 		newlast = i;
1531 	}
1532 	KASSERT(newdt->dt_ff[0] == (fdfile_t *)newfdp->fd_dfdfile[0]);
1533 	newfdp->fd_lastfile = newlast;
1534 	fd_checkmaps(newfdp);
1535 	mutex_exit(&fdp->fd_lock);
1536 
1537 	return newfdp;
1538 }
1539 
1540 /*
1541  * Release a filedesc structure.
1542  */
1543 void
1544 fd_free(void)
1545 {
1546 	fdfile_t *ff;
1547 	file_t *fp;
1548 	int fd, nf;
1549 	fdtab_t *dt;
1550 	lwp_t * const l = curlwp;
1551 	filedesc_t * const fdp = l->l_fd;
1552 	const bool noadvlock = (l->l_proc->p_flag & PK_ADVLOCK) == 0;
1553 
1554 	KASSERT(atomic_load_consume(&fdp->fd_dt)->dt_ff[0] ==
1555 	    (fdfile_t *)fdp->fd_dfdfile[0]);
1556 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1557 	KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1558 
1559 	membar_release();
1560 	if (atomic_dec_uint_nv(&fdp->fd_refcnt) > 0)
1561 		return;
1562 	membar_acquire();
1563 
1564 	/*
1565 	 * Close any files that the process holds open.
1566 	 */
1567 	dt = fdp->fd_dt;
1568 	fd_checkmaps(fdp);
1569 #ifdef DEBUG
1570 	fdp->fd_refcnt = -1; /* see fd_checkmaps */
1571 #endif
1572 	for (fd = 0, nf = dt->dt_nfiles; fd < nf; fd++) {
1573 		ff = dt->dt_ff[fd];
1574 		KASSERT(fd >= NDFDFILE ||
1575 		    ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1576 		if (ff == NULL)
1577 			continue;
1578 		if ((fp = atomic_load_consume(&ff->ff_file)) != NULL) {
1579 			/*
1580 			 * Must use fd_close() here if there is
1581 			 * a reference from kqueue or we might have posix
1582 			 * advisory locks.
1583 			 */
1584 			if (__predict_true(ff->ff_refcnt == 0) &&
1585 			    (noadvlock || fp->f_type != DTYPE_VNODE)) {
1586 				ff->ff_file = NULL;
1587 				ff->ff_exclose = false;
1588 				ff->ff_allocated = false;
1589 				closef(fp);
1590 			} else {
1591 				ff->ff_refcnt++;
1592 				fd_close(fd);
1593 			}
1594 		}
1595 		KASSERT(ff->ff_refcnt == 0);
1596 		KASSERT(ff->ff_file == NULL);
1597 		KASSERT(!ff->ff_exclose);
1598 		KASSERT(!ff->ff_allocated);
1599 		if (fd >= NDFDFILE) {
1600 			cv_destroy(&ff->ff_closing);
1601 			kmem_free(ff, sizeof(*ff));
1602 			dt->dt_ff[fd] = NULL;
1603 		}
1604 	}
1605 
1606 	/*
1607 	 * Clean out the descriptor table for the next user and return
1608 	 * to the cache.
1609 	 */
1610 	if (__predict_false(dt != &fdp->fd_dtbuiltin)) {
1611 		fd_dtab_free(fdp->fd_dt);
1612 		/* Otherwise, done above. */
1613 		memset(&fdp->fd_dtbuiltin.dt_ff[NDFDFILE], 0,
1614 		    (NDFILE - NDFDFILE) * sizeof(fdp->fd_dtbuiltin.dt_ff[0]));
1615 		fdp->fd_dt = &fdp->fd_dtbuiltin;
1616 	}
1617 	if (__predict_false(NDHISLOTS(nf) > NDHISLOTS(NDFILE))) {
1618 		KASSERT(fdp->fd_himap != fdp->fd_dhimap);
1619 		KASSERT(fdp->fd_lomap != fdp->fd_dlomap);
1620 		fd_map_free(nf, fdp->fd_lomap, fdp->fd_himap);
1621 	}
1622 	if (__predict_false(fdp->fd_knhash != NULL)) {
1623 		hashdone(fdp->fd_knhash, HASH_LIST, fdp->fd_knhashmask);
1624 		fdp->fd_knhash = NULL;
1625 		fdp->fd_knhashmask = 0;
1626 	} else {
1627 		KASSERT(fdp->fd_knhashmask == 0);
1628 	}
1629 	fdp->fd_dt = &fdp->fd_dtbuiltin;
1630 	fdp->fd_lastkqfile = -1;
1631 	fdp->fd_lastfile = -1;
1632 	fdp->fd_freefile = 0;
1633 	fdp->fd_exclose = false;
1634 	memset(&fdp->fd_startzero, 0, sizeof(*fdp) -
1635 	    offsetof(filedesc_t, fd_startzero));
1636 	fdp->fd_himap = fdp->fd_dhimap;
1637 	fdp->fd_lomap = fdp->fd_dlomap;
1638 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1639 	KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1640 	KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1641 #ifdef DEBUG
1642 	fdp->fd_refcnt = 0; /* see fd_checkmaps */
1643 #endif
1644 	fd_checkmaps(fdp);
1645 	pool_cache_put(filedesc_cache, fdp);
1646 }
1647 
1648 /*
1649  * File Descriptor pseudo-device driver (/dev/fd/).
1650  *
1651  * Opening minor device N dup()s the file (if any) connected to file
1652  * descriptor N belonging to the calling process.  Note that this driver
1653  * consists of only the ``open()'' routine, because all subsequent
1654  * references to this file will be direct to the other driver.
1655  */
1656 static int
1657 filedescopen(dev_t dev, int mode, int type, lwp_t *l)
1658 {
1659 
1660 	/*
1661 	 * XXX Kludge: set dupfd to contain the value of the
1662 	 * the file descriptor being sought for duplication. The error
1663 	 * return ensures that the vnode for this device will be released
1664 	 * by vn_open. Open will detect this special error and take the
1665 	 * actions in fd_dupopen below. Other callers of vn_open or VOP_OPEN
1666 	 * will simply report the error.
1667 	 */
1668 	l->l_dupfd = minor(dev);	/* XXX */
1669 	return EDUPFD;
1670 }
1671 
1672 /*
1673  * Duplicate the specified descriptor to a free descriptor.
1674  *
1675  * old is the original fd.
1676  * moveit is true if we should move rather than duplicate.
1677  * flags are the open flags (converted from O_* to F*).
1678  * newp returns the new fd on success.
1679  *
1680  * These two cases are produced by the EDUPFD and EMOVEFD magic
1681  * errnos, but in the interest of removing that regrettable interface,
1682  * vn_open has been changed to intercept them. Now vn_open returns
1683  * either a vnode or a filehandle, and the filehandle is accompanied
1684  * by a boolean that says whether we should dup (moveit == false) or
1685  * move (moveit == true) the fd.
1686  *
1687  * The dup case is used by /dev/stderr, /proc/self/fd, and such. The
1688  * move case is used by cloner devices that allocate a fd of their
1689  * own (a layering violation that should go away eventually) that
1690  * then needs to be put in the place open() expects it.
1691  */
1692 int
1693 fd_dupopen(int old, bool moveit, int flags, int *newp)
1694 {
1695 	filedesc_t *fdp;
1696 	fdfile_t *ff;
1697 	file_t *fp;
1698 	fdtab_t *dt;
1699 	int error;
1700 
1701 	if ((fp = fd_getfile(old)) == NULL) {
1702 		return EBADF;
1703 	}
1704 	fdp = curlwp->l_fd;
1705 	dt = atomic_load_consume(&fdp->fd_dt);
1706 	ff = dt->dt_ff[old];
1707 
1708 	/*
1709 	 * There are two cases of interest here.
1710 	 *
1711 	 * 1. moveit == false (used to be the EDUPFD magic errno):
1712 	 *    simply dup (old) to file descriptor (new) and return.
1713 	 *
1714 	 * 2. moveit == true (used to be the EMOVEFD magic errno):
1715 	 *    steal away the file structure from (old) and store it in
1716 	 *    (new).  (old) is effectively closed by this operation.
1717 	 */
1718 	if (moveit == false) {
1719 		/*
1720 		 * Check that the mode the file is being opened for is a
1721 		 * subset of the mode of the existing descriptor.
1722 		 */
1723 		if (((flags & (FREAD|FWRITE)) | fp->f_flag) != fp->f_flag) {
1724 			error = EACCES;
1725 			goto out;
1726 		}
1727 
1728 		/* Copy it. */
1729 		error = fd_dup(fp, 0, newp, ff->ff_exclose);
1730 	} else {
1731 		/* Copy it. */
1732 		error = fd_dup(fp, 0, newp, ff->ff_exclose);
1733 		if (error != 0) {
1734 			goto out;
1735 		}
1736 
1737 		/* Steal away the file pointer from 'old'. */
1738 		(void)fd_close(old);
1739 		return 0;
1740 	}
1741 
1742 out:
1743 	fd_putfile(old);
1744 	return error;
1745 }
1746 
1747 /*
1748  * Close open files on exec.
1749  */
1750 void
1751 fd_closeexec(void)
1752 {
1753 	proc_t *p;
1754 	filedesc_t *fdp;
1755 	fdfile_t *ff;
1756 	lwp_t *l;
1757 	fdtab_t *dt;
1758 	int fd;
1759 
1760 	l = curlwp;
1761 	p = l->l_proc;
1762 	fdp = p->p_fd;
1763 
1764 	if (fdp->fd_refcnt > 1) {
1765 		fdp = fd_copy();
1766 		fd_free();
1767 		p->p_fd = fdp;
1768 		l->l_fd = fdp;
1769 	}
1770 	if (!fdp->fd_exclose) {
1771 		return;
1772 	}
1773 	fdp->fd_exclose = false;
1774 	dt = atomic_load_consume(&fdp->fd_dt);
1775 
1776 	for (fd = 0; fd <= fdp->fd_lastfile; fd++) {
1777 		if ((ff = dt->dt_ff[fd]) == NULL) {
1778 			KASSERT(fd >= NDFDFILE);
1779 			continue;
1780 		}
1781 		KASSERT(fd >= NDFDFILE ||
1782 		    ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1783 		if (ff->ff_file == NULL)
1784 			continue;
1785 		if (ff->ff_exclose) {
1786 			/*
1787 			 * We need a reference to close the file.
1788 			 * No other threads can see the fdfile_t at
1789 			 * this point, so don't bother locking.
1790 			 */
1791 			KASSERT((ff->ff_refcnt & FR_CLOSING) == 0);
1792 			ff->ff_refcnt++;
1793 			fd_close(fd);
1794 		}
1795 	}
1796 }
1797 
1798 /*
1799  * Sets descriptor owner. If the owner is a process, 'pgid'
1800  * is set to positive value, process ID. If the owner is process group,
1801  * 'pgid' is set to -pg_id.
1802  */
1803 int
1804 fsetown(pid_t *pgid, u_long cmd, const void *data)
1805 {
1806 	pid_t id = *(const pid_t *)data;
1807 	int error;
1808 
1809 	if (id <= INT_MIN)
1810 		return EINVAL;
1811 
1812 	switch (cmd) {
1813 	case TIOCSPGRP:
1814 		if (id < 0)
1815 			return EINVAL;
1816 		id = -id;
1817 		break;
1818 	default:
1819 		break;
1820 	}
1821 	if (id > 0) {
1822 		mutex_enter(&proc_lock);
1823 		error = proc_find(id) ? 0 : ESRCH;
1824 		mutex_exit(&proc_lock);
1825 	} else if (id < 0) {
1826 		error = pgid_in_session(curproc, -id);
1827 	} else {
1828 		error = 0;
1829 	}
1830 	if (!error) {
1831 		*pgid = id;
1832 	}
1833 	return error;
1834 }
1835 
1836 void
1837 fd_set_exclose(struct lwp *l, int fd, bool exclose)
1838 {
1839 	filedesc_t *fdp = l->l_fd;
1840 	fdfile_t *ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
1841 
1842 	ff->ff_exclose = exclose;
1843 	if (exclose)
1844 		fdp->fd_exclose = true;
1845 }
1846 
1847 /*
1848  * Return descriptor owner information. If the value is positive,
1849  * it's process ID. If it's negative, it's process group ID and
1850  * needs the sign removed before use.
1851  */
1852 int
1853 fgetown(pid_t pgid, u_long cmd, void *data)
1854 {
1855 
1856 	switch (cmd) {
1857 	case TIOCGPGRP:
1858 		KASSERT(pgid > INT_MIN);
1859 		*(int *)data = -pgid;
1860 		break;
1861 	default:
1862 		*(int *)data = pgid;
1863 		break;
1864 	}
1865 	return 0;
1866 }
1867 
1868 /*
1869  * Send signal to descriptor owner, either process or process group.
1870  */
1871 void
1872 fownsignal(pid_t pgid, int signo, int code, int band, void *fdescdata)
1873 {
1874 	ksiginfo_t ksi;
1875 
1876 	KASSERT(!cpu_intr_p());
1877 
1878 	if (pgid == 0) {
1879 		return;
1880 	}
1881 
1882 	KSI_INIT(&ksi);
1883 	ksi.ksi_signo = signo;
1884 	ksi.ksi_code = code;
1885 	ksi.ksi_band = band;
1886 
1887 	mutex_enter(&proc_lock);
1888 	if (pgid > 0) {
1889 		struct proc *p1;
1890 
1891 		p1 = proc_find(pgid);
1892 		if (p1 != NULL) {
1893 			kpsignal(p1, &ksi, fdescdata);
1894 		}
1895 	} else {
1896 		struct pgrp *pgrp;
1897 
1898 		KASSERT(pgid < 0 && pgid > INT_MIN);
1899 		pgrp = pgrp_find(-pgid);
1900 		if (pgrp != NULL) {
1901 			kpgsignal(pgrp, &ksi, fdescdata, 0);
1902 		}
1903 	}
1904 	mutex_exit(&proc_lock);
1905 }
1906 
1907 int
1908 fd_clone(file_t *fp, unsigned fd, int flag, const struct fileops *fops,
1909 	 void *data)
1910 {
1911 
1912 	fp->f_flag = flag & FMASK;
1913 	fd_set_exclose(curlwp, fd, (flag & O_CLOEXEC) != 0);
1914 	fp->f_type = DTYPE_MISC;
1915 	fp->f_ops = fops;
1916 	fp->f_data = data;
1917 	curlwp->l_dupfd = fd;
1918 	fd_affix(curproc, fp, fd);
1919 
1920 	return EMOVEFD;
1921 }
1922 
1923 int
1924 fnullop_fcntl(file_t *fp, u_int cmd, void *data)
1925 {
1926 
1927 	if (cmd == F_SETFL)
1928 		return 0;
1929 
1930 	return EOPNOTSUPP;
1931 }
1932 
1933 int
1934 fnullop_poll(file_t *fp, int which)
1935 {
1936 
1937 	return 0;
1938 }
1939 
1940 int
1941 fnullop_kqfilter(file_t *fp, struct knote *kn)
1942 {
1943 
1944 	return EOPNOTSUPP;
1945 }
1946 
1947 void
1948 fnullop_restart(file_t *fp)
1949 {
1950 
1951 }
1952 
1953 int
1954 fbadop_read(file_t *fp, off_t *offset, struct uio *uio,
1955 	    kauth_cred_t cred, int flags)
1956 {
1957 
1958 	return EOPNOTSUPP;
1959 }
1960 
1961 int
1962 fbadop_write(file_t *fp, off_t *offset, struct uio *uio,
1963 	     kauth_cred_t cred, int flags)
1964 {
1965 
1966 	return EOPNOTSUPP;
1967 }
1968 
1969 int
1970 fbadop_ioctl(file_t *fp, u_long com, void *data)
1971 {
1972 
1973 	return EOPNOTSUPP;
1974 }
1975 
1976 int
1977 fbadop_stat(file_t *fp, struct stat *sb)
1978 {
1979 
1980 	return EOPNOTSUPP;
1981 }
1982 
1983 int
1984 fbadop_close(file_t *fp)
1985 {
1986 
1987 	return EOPNOTSUPP;
1988 }
1989 
1990 /*
1991  * sysctl routines pertaining to file descriptors
1992  */
1993 
1994 /* Initialized in sysctl_init() for now... */
1995 extern kmutex_t sysctl_file_marker_lock;
1996 static u_int sysctl_file_marker = 1;
1997 
1998 /*
1999  * Expects to be called with proc_lock and sysctl_file_marker_lock locked.
2000  */
2001 static void
2002 sysctl_file_marker_reset(void)
2003 {
2004 	struct proc *p;
2005 
2006 	PROCLIST_FOREACH(p, &allproc) {
2007 		struct filedesc *fd = p->p_fd;
2008 		fdtab_t *dt;
2009 		u_int i;
2010 
2011 		mutex_enter(&fd->fd_lock);
2012 		dt = fd->fd_dt;
2013 		for (i = 0; i < dt->dt_nfiles; i++) {
2014 			struct file *fp;
2015 			fdfile_t *ff;
2016 
2017 			if ((ff = dt->dt_ff[i]) == NULL) {
2018 				continue;
2019 			}
2020 			if ((fp = atomic_load_consume(&ff->ff_file)) == NULL) {
2021 				continue;
2022 			}
2023 			fp->f_marker = 0;
2024 		}
2025 		mutex_exit(&fd->fd_lock);
2026 	}
2027 }
2028 
2029 /*
2030  * sysctl helper routine for kern.file pseudo-subtree.
2031  */
2032 static int
2033 sysctl_kern_file(SYSCTLFN_ARGS)
2034 {
2035 	const bool allowaddr = get_expose_address(curproc);
2036 	struct filelist flist;
2037 	int error;
2038 	size_t buflen;
2039 	struct file *fp, fbuf;
2040 	char *start, *where;
2041 	struct proc *p;
2042 
2043 	start = where = oldp;
2044 	buflen = *oldlenp;
2045 
2046 	if (where == NULL) {
2047 		/*
2048 		 * overestimate by 10 files
2049 		 */
2050 		*oldlenp = sizeof(filehead) + (nfiles + 10) *
2051 		    sizeof(struct file);
2052 		return 0;
2053 	}
2054 
2055 	/*
2056 	 * first sysctl_copyout filehead
2057 	 */
2058 	if (buflen < sizeof(filehead)) {
2059 		*oldlenp = 0;
2060 		return 0;
2061 	}
2062 	sysctl_unlock();
2063 	if (allowaddr) {
2064 		memcpy(&flist, &filehead, sizeof(flist));
2065 	} else {
2066 		memset(&flist, 0, sizeof(flist));
2067 	}
2068 	error = sysctl_copyout(l, &flist, where, sizeof(flist));
2069 	if (error) {
2070 		sysctl_relock();
2071 		return error;
2072 	}
2073 	buflen -= sizeof(flist);
2074 	where += sizeof(flist);
2075 
2076 	/*
2077 	 * followed by an array of file structures
2078 	 */
2079 	mutex_enter(&sysctl_file_marker_lock);
2080 	mutex_enter(&proc_lock);
2081 	PROCLIST_FOREACH(p, &allproc) {
2082 		struct filedesc *fd;
2083 		fdtab_t *dt;
2084 		u_int i;
2085 
2086 		if (p->p_stat == SIDL) {
2087 			/* skip embryonic processes */
2088 			continue;
2089 		}
2090 		mutex_enter(p->p_lock);
2091 		error = kauth_authorize_process(l->l_cred,
2092 		    KAUTH_PROCESS_CANSEE, p,
2093 		    KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2094 		    NULL, NULL);
2095 		mutex_exit(p->p_lock);
2096 		if (error != 0) {
2097 			/*
2098 			 * Don't leak kauth retval if we're silently
2099 			 * skipping this entry.
2100 			 */
2101 			error = 0;
2102 			continue;
2103 		}
2104 
2105 		/*
2106 		 * Grab a hold on the process.
2107 		 */
2108 		if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2109 			continue;
2110 		}
2111 		mutex_exit(&proc_lock);
2112 
2113 		fd = p->p_fd;
2114 		mutex_enter(&fd->fd_lock);
2115 		dt = fd->fd_dt;
2116 		for (i = 0; i < dt->dt_nfiles; i++) {
2117 			fdfile_t *ff;
2118 
2119 			if ((ff = dt->dt_ff[i]) == NULL) {
2120 				continue;
2121 			}
2122 			if ((fp = atomic_load_consume(&ff->ff_file)) == NULL) {
2123 				continue;
2124 			}
2125 
2126 			mutex_enter(&fp->f_lock);
2127 
2128 			if ((fp->f_count == 0) ||
2129 			    (fp->f_marker == sysctl_file_marker)) {
2130 				mutex_exit(&fp->f_lock);
2131 				continue;
2132 			}
2133 
2134 			/* Check that we have enough space. */
2135 			if (buflen < sizeof(struct file)) {
2136 				*oldlenp = where - start;
2137 				mutex_exit(&fp->f_lock);
2138 				error = ENOMEM;
2139 				break;
2140 			}
2141 
2142 			fill_file(&fbuf, fp);
2143 			mutex_exit(&fp->f_lock);
2144 			error = sysctl_copyout(l, &fbuf, where, sizeof(fbuf));
2145 			if (error) {
2146 				break;
2147 			}
2148 			buflen -= sizeof(struct file);
2149 			where += sizeof(struct file);
2150 
2151 			fp->f_marker = sysctl_file_marker;
2152 		}
2153 		mutex_exit(&fd->fd_lock);
2154 
2155 		/*
2156 		 * Release reference to process.
2157 		 */
2158 		mutex_enter(&proc_lock);
2159 		rw_exit(&p->p_reflock);
2160 
2161 		if (error)
2162 			break;
2163 	}
2164 
2165 	sysctl_file_marker++;
2166 	/* Reset all markers if wrapped. */
2167 	if (sysctl_file_marker == 0) {
2168 		sysctl_file_marker_reset();
2169 		sysctl_file_marker++;
2170 	}
2171 
2172 	mutex_exit(&proc_lock);
2173 	mutex_exit(&sysctl_file_marker_lock);
2174 
2175 	*oldlenp = where - start;
2176 	sysctl_relock();
2177 	return error;
2178 }
2179 
2180 /*
2181  * sysctl helper function for kern.file2
2182  */
2183 static int
2184 sysctl_kern_file2(SYSCTLFN_ARGS)
2185 {
2186 	struct proc *p;
2187 	struct file *fp;
2188 	struct filedesc *fd;
2189 	struct kinfo_file kf;
2190 	char *dp;
2191 	u_int i, op;
2192 	size_t len, needed, elem_size, out_size;
2193 	int error, arg, elem_count;
2194 	fdfile_t *ff;
2195 	fdtab_t *dt;
2196 
2197 	if (namelen == 1 && name[0] == CTL_QUERY)
2198 		return sysctl_query(SYSCTLFN_CALL(rnode));
2199 
2200 	if (namelen != 4)
2201 		return EINVAL;
2202 
2203 	error = 0;
2204 	dp = oldp;
2205 	len = (oldp != NULL) ? *oldlenp : 0;
2206 	op = name[0];
2207 	arg = name[1];
2208 	elem_size = name[2];
2209 	elem_count = name[3];
2210 	out_size = MIN(sizeof(kf), elem_size);
2211 	needed = 0;
2212 
2213 	if (elem_size < 1 || elem_count < 0)
2214 		return EINVAL;
2215 
2216 	switch (op) {
2217 	case KERN_FILE_BYFILE:
2218 	case KERN_FILE_BYPID:
2219 		/*
2220 		 * We're traversing the process list in both cases; the BYFILE
2221 		 * case does additional work of keeping track of files already
2222 		 * looked at.
2223 		 */
2224 
2225 		/* doesn't use arg so it must be zero */
2226 		if ((op == KERN_FILE_BYFILE) && (arg != 0))
2227 			return EINVAL;
2228 
2229 		if ((op == KERN_FILE_BYPID) && (arg < -1))
2230 			/* -1 means all processes */
2231 			return EINVAL;
2232 
2233 		sysctl_unlock();
2234 		if (op == KERN_FILE_BYFILE)
2235 			mutex_enter(&sysctl_file_marker_lock);
2236 		mutex_enter(&proc_lock);
2237 		PROCLIST_FOREACH(p, &allproc) {
2238 			if (p->p_stat == SIDL) {
2239 				/* skip embryonic processes */
2240 				continue;
2241 			}
2242 			if (arg > 0 && p->p_pid != arg) {
2243 				/* pick only the one we want */
2244 				/* XXX want 0 to mean "kernel files" */
2245 				continue;
2246 			}
2247 			mutex_enter(p->p_lock);
2248 			error = kauth_authorize_process(l->l_cred,
2249 			    KAUTH_PROCESS_CANSEE, p,
2250 			    KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2251 			    NULL, NULL);
2252 			mutex_exit(p->p_lock);
2253 			if (error != 0) {
2254 				/*
2255 				 * Don't leak kauth retval if we're silently
2256 				 * skipping this entry.
2257 				 */
2258 				error = 0;
2259 				continue;
2260 			}
2261 
2262 			/*
2263 			 * Grab a hold on the process.
2264 			 */
2265 			if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2266 				continue;
2267 			}
2268 			mutex_exit(&proc_lock);
2269 
2270 			fd = p->p_fd;
2271 			mutex_enter(&fd->fd_lock);
2272 			dt = fd->fd_dt;
2273 			for (i = 0; i < dt->dt_nfiles; i++) {
2274 				if ((ff = dt->dt_ff[i]) == NULL) {
2275 					continue;
2276 				}
2277 				if ((fp = atomic_load_consume(&ff->ff_file)) ==
2278 				    NULL) {
2279 					continue;
2280 				}
2281 
2282 				if ((op == KERN_FILE_BYFILE) &&
2283 				    (fp->f_marker == sysctl_file_marker)) {
2284 					continue;
2285 				}
2286 				if (len >= elem_size && elem_count > 0) {
2287 					mutex_enter(&fp->f_lock);
2288 					fill_file2(&kf, fp, ff, i, p->p_pid);
2289 					mutex_exit(&fp->f_lock);
2290 					mutex_exit(&fd->fd_lock);
2291 					error = sysctl_copyout(l,
2292 					    &kf, dp, out_size);
2293 					mutex_enter(&fd->fd_lock);
2294 					if (error)
2295 						break;
2296 					dp += elem_size;
2297 					len -= elem_size;
2298 				}
2299 				if (op == KERN_FILE_BYFILE)
2300 					fp->f_marker = sysctl_file_marker;
2301 				needed += elem_size;
2302 				if (elem_count > 0 && elem_count != INT_MAX)
2303 					elem_count--;
2304 			}
2305 			mutex_exit(&fd->fd_lock);
2306 
2307 			/*
2308 			 * Release reference to process.
2309 			 */
2310 			mutex_enter(&proc_lock);
2311 			rw_exit(&p->p_reflock);
2312 		}
2313 		if (op == KERN_FILE_BYFILE) {
2314 			sysctl_file_marker++;
2315 
2316 			/* Reset all markers if wrapped. */
2317 			if (sysctl_file_marker == 0) {
2318 				sysctl_file_marker_reset();
2319 				sysctl_file_marker++;
2320 			}
2321 		}
2322 		mutex_exit(&proc_lock);
2323 		if (op == KERN_FILE_BYFILE)
2324 			mutex_exit(&sysctl_file_marker_lock);
2325 		sysctl_relock();
2326 		break;
2327 	default:
2328 		return EINVAL;
2329 	}
2330 
2331 	if (oldp == NULL)
2332 		needed += KERN_FILESLOP * elem_size;
2333 	*oldlenp = needed;
2334 
2335 	return error;
2336 }
2337 
2338 static void
2339 fill_file(struct file *fp, const struct file *fpsrc)
2340 {
2341 	const bool allowaddr = get_expose_address(curproc);
2342 
2343 	memset(fp, 0, sizeof(*fp));
2344 
2345 	fp->f_offset = fpsrc->f_offset;
2346 	COND_SET_PTR(fp->f_cred, fpsrc->f_cred, allowaddr);
2347 	COND_SET_CPTR(fp->f_ops, fpsrc->f_ops, allowaddr);
2348 	COND_SET_STRUCT(fp->f_undata, fpsrc->f_undata, allowaddr);
2349 	COND_SET_STRUCT(fp->f_list, fpsrc->f_list, allowaddr);
2350 	fp->f_flag = fpsrc->f_flag;
2351 	fp->f_marker = fpsrc->f_marker;
2352 	fp->f_type = fpsrc->f_type;
2353 	fp->f_advice = fpsrc->f_advice;
2354 	fp->f_count = fpsrc->f_count;
2355 	fp->f_msgcount = fpsrc->f_msgcount;
2356 	fp->f_unpcount = fpsrc->f_unpcount;
2357 	COND_SET_STRUCT(fp->f_unplist, fpsrc->f_unplist, allowaddr);
2358 }
2359 
2360 static void
2361 fill_file2(struct kinfo_file *kp, const file_t *fp, const fdfile_t *ff,
2362 	  int i, pid_t pid)
2363 {
2364 	const bool allowaddr = get_expose_address(curproc);
2365 
2366 	memset(kp, 0, sizeof(*kp));
2367 
2368 	COND_SET_VALUE(kp->ki_fileaddr, PTRTOUINT64(fp), allowaddr);
2369 	kp->ki_flag =		fp->f_flag;
2370 	kp->ki_iflags =		0;
2371 	kp->ki_ftype =		fp->f_type;
2372 	kp->ki_count =		fp->f_count;
2373 	kp->ki_msgcount =	fp->f_msgcount;
2374 	COND_SET_VALUE(kp->ki_fucred, PTRTOUINT64(fp->f_cred), allowaddr);
2375 	kp->ki_fuid =		kauth_cred_geteuid(fp->f_cred);
2376 	kp->ki_fgid =		kauth_cred_getegid(fp->f_cred);
2377 	COND_SET_VALUE(kp->ki_fops, PTRTOUINT64(fp->f_ops), allowaddr);
2378 	kp->ki_foffset =	fp->f_offset;
2379 	COND_SET_VALUE(kp->ki_fdata, PTRTOUINT64(fp->f_data), allowaddr);
2380 
2381 	/* vnode information to glue this file to something */
2382 	if (fp->f_type == DTYPE_VNODE) {
2383 		struct vnode *vp = fp->f_vnode;
2384 
2385 		COND_SET_VALUE(kp->ki_vun, PTRTOUINT64(vp->v_un.vu_socket),
2386 		    allowaddr);
2387 		kp->ki_vsize =	vp->v_size;
2388 		kp->ki_vtype =	vp->v_type;
2389 		kp->ki_vtag =	vp->v_tag;
2390 		COND_SET_VALUE(kp->ki_vdata, PTRTOUINT64(vp->v_data),
2391 		    allowaddr);
2392 	}
2393 
2394 	/* process information when retrieved via KERN_FILE_BYPID */
2395 	if (ff != NULL) {
2396 		kp->ki_pid =		pid;
2397 		kp->ki_fd =		i;
2398 		kp->ki_ofileflags =	ff->ff_exclose;
2399 		kp->ki_usecount =	ff->ff_refcnt;
2400 	}
2401 }
2402