xref: /netbsd-src/sys/kern/kern_descrip.c (revision 46f5119e40af2e51998f686b2fdcc76b5488f7f3)
1 /*	$NetBSD: kern_descrip.c,v 1.214 2011/04/24 20:30:38 rmind Exp $	*/
2 
3 /*-
4  * Copyright (c) 2008, 2009 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Andrew Doran.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 /*
33  * Copyright (c) 1982, 1986, 1989, 1991, 1993
34  *	The Regents of the University of California.  All rights reserved.
35  * (c) UNIX System Laboratories, Inc.
36  * All or some portions of this file are derived from material licensed
37  * to the University of California by American Telephone and Telegraph
38  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
39  * the permission of UNIX System Laboratories, Inc.
40  *
41  * Redistribution and use in source and binary forms, with or without
42  * modification, are permitted provided that the following conditions
43  * are met:
44  * 1. Redistributions of source code must retain the above copyright
45  *    notice, this list of conditions and the following disclaimer.
46  * 2. Redistributions in binary form must reproduce the above copyright
47  *    notice, this list of conditions and the following disclaimer in the
48  *    documentation and/or other materials provided with the distribution.
49  * 3. Neither the name of the University nor the names of its contributors
50  *    may be used to endorse or promote products derived from this software
51  *    without specific prior written permission.
52  *
53  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
54  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
55  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
56  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
57  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
58  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
59  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
60  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
61  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
62  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
63  * SUCH DAMAGE.
64  *
65  *	@(#)kern_descrip.c	8.8 (Berkeley) 2/14/95
66  */
67 
68 /*
69  * File descriptor management.
70  */
71 
72 #include <sys/cdefs.h>
73 __KERNEL_RCSID(0, "$NetBSD: kern_descrip.c,v 1.214 2011/04/24 20:30:38 rmind Exp $");
74 
75 #include <sys/param.h>
76 #include <sys/systm.h>
77 #include <sys/filedesc.h>
78 #include <sys/kernel.h>
79 #include <sys/proc.h>
80 #include <sys/file.h>
81 #include <sys/socket.h>
82 #include <sys/socketvar.h>
83 #include <sys/stat.h>
84 #include <sys/ioctl.h>
85 #include <sys/fcntl.h>
86 #include <sys/pool.h>
87 #include <sys/unistd.h>
88 #include <sys/resourcevar.h>
89 #include <sys/conf.h>
90 #include <sys/event.h>
91 #include <sys/kauth.h>
92 #include <sys/atomic.h>
93 #include <sys/syscallargs.h>
94 #include <sys/cpu.h>
95 #include <sys/kmem.h>
96 #include <sys/vnode.h>
97 #include <sys/sysctl.h>
98 #include <sys/ktrace.h>
99 
100 /*
101  * A list (head) of open files, counter, and lock protecting them.
102  */
103 struct filelist		filehead	__cacheline_aligned;
104 static u_int		nfiles		__cacheline_aligned;
105 kmutex_t		filelist_lock	__cacheline_aligned;
106 
107 static pool_cache_t	filedesc_cache	__read_mostly;
108 static pool_cache_t	file_cache	__read_mostly;
109 static pool_cache_t	fdfile_cache	__read_mostly;
110 
111 static int	file_ctor(void *, void *, int);
112 static void	file_dtor(void *, void *);
113 static int	fdfile_ctor(void *, void *, int);
114 static void	fdfile_dtor(void *, void *);
115 static int	filedesc_ctor(void *, void *, int);
116 static void	filedesc_dtor(void *, void *);
117 static int	filedescopen(dev_t, int, int, lwp_t *);
118 
119 static int sysctl_kern_file(SYSCTLFN_PROTO);
120 static int sysctl_kern_file2(SYSCTLFN_PROTO);
121 static void fill_file(struct kinfo_file *, const file_t *, const fdfile_t *,
122 		      int, pid_t);
123 
124 const struct cdevsw filedesc_cdevsw = {
125 	filedescopen, noclose, noread, nowrite, noioctl,
126 	nostop, notty, nopoll, nommap, nokqfilter, D_OTHER | D_MPSAFE,
127 };
128 
129 /* For ease of reading. */
130 __strong_alias(fd_putvnode,fd_putfile)
131 __strong_alias(fd_putsock,fd_putfile)
132 
133 /*
134  * Initialize the descriptor system.
135  */
136 void
137 fd_sys_init(void)
138 {
139 	static struct sysctllog *clog;
140 
141 	mutex_init(&filelist_lock, MUTEX_DEFAULT, IPL_NONE);
142 
143 	file_cache = pool_cache_init(sizeof(file_t), coherency_unit, 0,
144 	    0, "file", NULL, IPL_NONE, file_ctor, file_dtor, NULL);
145 	KASSERT(file_cache != NULL);
146 
147 	fdfile_cache = pool_cache_init(sizeof(fdfile_t), coherency_unit, 0,
148 	    PR_LARGECACHE, "fdfile", NULL, IPL_NONE, fdfile_ctor, fdfile_dtor,
149 	    NULL);
150 	KASSERT(fdfile_cache != NULL);
151 
152 	filedesc_cache = pool_cache_init(sizeof(filedesc_t), coherency_unit,
153 	    0, 0, "filedesc", NULL, IPL_NONE, filedesc_ctor, filedesc_dtor,
154 	    NULL);
155 	KASSERT(filedesc_cache != NULL);
156 
157 	sysctl_createv(&clog, 0, NULL, NULL,
158 		       CTLFLAG_PERMANENT, CTLTYPE_NODE, "kern", NULL,
159 		       NULL, 0, NULL, 0, CTL_KERN, CTL_EOL);
160 	sysctl_createv(&clog, 0, NULL, NULL,
161 		       CTLFLAG_PERMANENT,
162 		       CTLTYPE_STRUCT, "file",
163 		       SYSCTL_DESCR("System open file table"),
164 		       sysctl_kern_file, 0, NULL, 0,
165 		       CTL_KERN, KERN_FILE, CTL_EOL);
166 	sysctl_createv(&clog, 0, NULL, NULL,
167 		       CTLFLAG_PERMANENT,
168 		       CTLTYPE_STRUCT, "file2",
169 		       SYSCTL_DESCR("System open file table"),
170 		       sysctl_kern_file2, 0, NULL, 0,
171 		       CTL_KERN, KERN_FILE2, CTL_EOL);
172 }
173 
174 static bool
175 fd_isused(filedesc_t *fdp, unsigned fd)
176 {
177 	u_int off = fd >> NDENTRYSHIFT;
178 
179 	KASSERT(fd < fdp->fd_dt->dt_nfiles);
180 
181 	return (fdp->fd_lomap[off] & (1 << (fd & NDENTRYMASK))) != 0;
182 }
183 
184 /*
185  * Verify that the bitmaps match the descriptor table.
186  */
187 static inline void
188 fd_checkmaps(filedesc_t *fdp)
189 {
190 #ifdef DEBUG
191 	fdtab_t *dt;
192 	u_int fd;
193 
194 	dt = fdp->fd_dt;
195 	if (fdp->fd_refcnt == -1) {
196 		/*
197 		 * fd_free tears down the table without maintaining its bitmap.
198 		 */
199 		return;
200 	}
201 	for (fd = 0; fd < dt->dt_nfiles; fd++) {
202 		if (fd < NDFDFILE) {
203 			KASSERT(dt->dt_ff[fd] ==
204 			    (fdfile_t *)fdp->fd_dfdfile[fd]);
205 		}
206 		if (dt->dt_ff[fd] == NULL) {
207 			KASSERT(!fd_isused(fdp, fd));
208 		} else if (dt->dt_ff[fd]->ff_file != NULL) {
209 			KASSERT(fd_isused(fdp, fd));
210 		}
211 	}
212 #endif
213 }
214 
215 static int
216 fd_next_zero(filedesc_t *fdp, uint32_t *bitmap, int want, u_int bits)
217 {
218 	int i, off, maxoff;
219 	uint32_t sub;
220 
221 	KASSERT(mutex_owned(&fdp->fd_lock));
222 
223 	fd_checkmaps(fdp);
224 
225 	if (want > bits)
226 		return -1;
227 
228 	off = want >> NDENTRYSHIFT;
229 	i = want & NDENTRYMASK;
230 	if (i) {
231 		sub = bitmap[off] | ((u_int)~0 >> (NDENTRIES - i));
232 		if (sub != ~0)
233 			goto found;
234 		off++;
235 	}
236 
237 	maxoff = NDLOSLOTS(bits);
238 	while (off < maxoff) {
239 		if ((sub = bitmap[off]) != ~0)
240 			goto found;
241 		off++;
242 	}
243 
244 	return -1;
245 
246  found:
247 	return (off << NDENTRYSHIFT) + ffs(~sub) - 1;
248 }
249 
250 static int
251 fd_last_set(filedesc_t *fd, int last)
252 {
253 	int off, i;
254 	fdfile_t **ff = fd->fd_dt->dt_ff;
255 	uint32_t *bitmap = fd->fd_lomap;
256 
257 	KASSERT(mutex_owned(&fd->fd_lock));
258 
259 	fd_checkmaps(fd);
260 
261 	off = (last - 1) >> NDENTRYSHIFT;
262 
263 	while (off >= 0 && !bitmap[off])
264 		off--;
265 
266 	if (off < 0)
267 		return -1;
268 
269 	i = ((off + 1) << NDENTRYSHIFT) - 1;
270 	if (i >= last)
271 		i = last - 1;
272 
273 	/* XXX should use bitmap */
274 	while (i > 0 && (ff[i] == NULL || !ff[i]->ff_allocated))
275 		i--;
276 
277 	return i;
278 }
279 
280 static inline void
281 fd_used(filedesc_t *fdp, unsigned fd)
282 {
283 	u_int off = fd >> NDENTRYSHIFT;
284 	fdfile_t *ff;
285 
286 	ff = fdp->fd_dt->dt_ff[fd];
287 
288 	KASSERT(mutex_owned(&fdp->fd_lock));
289 	KASSERT((fdp->fd_lomap[off] & (1 << (fd & NDENTRYMASK))) == 0);
290 	KASSERT(ff != NULL);
291 	KASSERT(ff->ff_file == NULL);
292 	KASSERT(!ff->ff_allocated);
293 
294 	ff->ff_allocated = 1;
295 	fdp->fd_lomap[off] |= 1 << (fd & NDENTRYMASK);
296 	if (__predict_false(fdp->fd_lomap[off] == ~0)) {
297 		KASSERT((fdp->fd_himap[off >> NDENTRYSHIFT] &
298 		    (1 << (off & NDENTRYMASK))) == 0);
299 		fdp->fd_himap[off >> NDENTRYSHIFT] |= 1 << (off & NDENTRYMASK);
300 	}
301 
302 	if ((int)fd > fdp->fd_lastfile) {
303 		fdp->fd_lastfile = fd;
304 	}
305 
306 	fd_checkmaps(fdp);
307 }
308 
309 static inline void
310 fd_unused(filedesc_t *fdp, unsigned fd)
311 {
312 	u_int off = fd >> NDENTRYSHIFT;
313 	fdfile_t *ff;
314 
315 	ff = fdp->fd_dt->dt_ff[fd];
316 
317 	/*
318 	 * Don't assert the lock is held here, as we may be copying
319 	 * the table during exec() and it is not needed there.
320 	 * procfs and sysctl are locked out by proc::p_reflock.
321 	 *
322 	 * KASSERT(mutex_owned(&fdp->fd_lock));
323 	 */
324 	KASSERT(ff != NULL);
325 	KASSERT(ff->ff_file == NULL);
326 	KASSERT(ff->ff_allocated);
327 
328 	if (fd < fdp->fd_freefile) {
329 		fdp->fd_freefile = fd;
330 	}
331 
332 	if (fdp->fd_lomap[off] == ~0) {
333 		KASSERT((fdp->fd_himap[off >> NDENTRYSHIFT] &
334 		    (1 << (off & NDENTRYMASK))) != 0);
335 		fdp->fd_himap[off >> NDENTRYSHIFT] &=
336 		    ~(1 << (off & NDENTRYMASK));
337 	}
338 	KASSERT((fdp->fd_lomap[off] & (1 << (fd & NDENTRYMASK))) != 0);
339 	fdp->fd_lomap[off] &= ~(1 << (fd & NDENTRYMASK));
340 	ff->ff_allocated = 0;
341 
342 	KASSERT(fd <= fdp->fd_lastfile);
343 	if (fd == fdp->fd_lastfile) {
344 		fdp->fd_lastfile = fd_last_set(fdp, fd);
345 	}
346 	fd_checkmaps(fdp);
347 }
348 
349 /*
350  * Look up the file structure corresponding to a file descriptor
351  * and return the file, holding a reference on the descriptor.
352  */
353 file_t *
354 fd_getfile(unsigned fd)
355 {
356 	filedesc_t *fdp;
357 	fdfile_t *ff;
358 	file_t *fp;
359 	fdtab_t *dt;
360 
361 	/*
362 	 * Look up the fdfile structure representing this descriptor.
363 	 * We are doing this unlocked.  See fd_tryexpand().
364 	 */
365 	fdp = curlwp->l_fd;
366 	dt = fdp->fd_dt;
367 	if (__predict_false(fd >= dt->dt_nfiles)) {
368 		return NULL;
369 	}
370 	ff = dt->dt_ff[fd];
371 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
372 	if (__predict_false(ff == NULL)) {
373 		return NULL;
374 	}
375 
376 	/* Now get a reference to the descriptor. */
377 	if (fdp->fd_refcnt == 1) {
378 		/*
379 		 * Single threaded: don't need to worry about concurrent
380 		 * access (other than earlier calls to kqueue, which may
381 		 * hold a reference to the descriptor).
382 		 */
383 		ff->ff_refcnt++;
384 	} else {
385 		/*
386 		 * Multi threaded: issue a memory barrier to ensure that we
387 		 * acquire the file pointer _after_ adding a reference.  If
388 		 * no memory barrier, we could fetch a stale pointer.
389 		 */
390 		atomic_inc_uint(&ff->ff_refcnt);
391 #ifndef __HAVE_ATOMIC_AS_MEMBAR
392 		membar_enter();
393 #endif
394 	}
395 
396 	/*
397 	 * If the file is not open or is being closed then put the
398 	 * reference back.
399 	 */
400 	fp = ff->ff_file;
401 	if (__predict_true(fp != NULL)) {
402 		return fp;
403 	}
404 	fd_putfile(fd);
405 	return NULL;
406 }
407 
408 /*
409  * Release a reference to a file descriptor acquired with fd_getfile().
410  */
411 void
412 fd_putfile(unsigned fd)
413 {
414 	filedesc_t *fdp;
415 	fdfile_t *ff;
416 	u_int u, v;
417 
418 	fdp = curlwp->l_fd;
419 	ff = fdp->fd_dt->dt_ff[fd];
420 
421 	KASSERT(fd < fdp->fd_dt->dt_nfiles);
422 	KASSERT(ff != NULL);
423 	KASSERT((ff->ff_refcnt & FR_MASK) > 0);
424 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
425 
426 	if (fdp->fd_refcnt == 1) {
427 		/*
428 		 * Single threaded: don't need to worry about concurrent
429 		 * access (other than earlier calls to kqueue, which may
430 		 * hold a reference to the descriptor).
431 		 */
432 		if (__predict_false((ff->ff_refcnt & FR_CLOSING) != 0)) {
433 			fd_close(fd);
434 			return;
435 		}
436 		ff->ff_refcnt--;
437 		return;
438 	}
439 
440 	/*
441 	 * Ensure that any use of the file is complete and globally
442 	 * visible before dropping the final reference.  If no membar,
443 	 * the current CPU could still access memory associated with
444 	 * the file after it has been freed or recycled by another
445 	 * CPU.
446 	 */
447 #ifndef __HAVE_ATOMIC_AS_MEMBAR
448 	membar_exit();
449 #endif
450 
451 	/*
452 	 * Be optimistic and start out with the assumption that no other
453 	 * threads are trying to close the descriptor.  If the CAS fails,
454 	 * we lost a race and/or it's being closed.
455 	 */
456 	for (u = ff->ff_refcnt & FR_MASK;; u = v) {
457 		v = atomic_cas_uint(&ff->ff_refcnt, u, u - 1);
458 		if (__predict_true(u == v)) {
459 			return;
460 		}
461 		if (__predict_false((v & FR_CLOSING) != 0)) {
462 			break;
463 		}
464 	}
465 
466 	/* Another thread is waiting to close the file: join it. */
467 	(void)fd_close(fd);
468 }
469 
470 /*
471  * Convenience wrapper around fd_getfile() that returns reference
472  * to a vnode.
473  */
474 int
475 fd_getvnode(unsigned fd, file_t **fpp)
476 {
477 	vnode_t *vp;
478 	file_t *fp;
479 
480 	fp = fd_getfile(fd);
481 	if (__predict_false(fp == NULL)) {
482 		return EBADF;
483 	}
484 	if (__predict_false(fp->f_type != DTYPE_VNODE)) {
485 		fd_putfile(fd);
486 		return EINVAL;
487 	}
488 	vp = fp->f_data;
489 	if (__predict_false(vp->v_type == VBAD)) {
490 		/* XXX Is this case really necessary? */
491 		fd_putfile(fd);
492 		return EBADF;
493 	}
494 	*fpp = fp;
495 	return 0;
496 }
497 
498 /*
499  * Convenience wrapper around fd_getfile() that returns reference
500  * to a socket.
501  */
502 int
503 fd_getsock(unsigned fd, struct socket **sop)
504 {
505 	file_t *fp;
506 
507 	fp = fd_getfile(fd);
508 	if (__predict_false(fp == NULL)) {
509 		return EBADF;
510 	}
511 	if (__predict_false(fp->f_type != DTYPE_SOCKET)) {
512 		fd_putfile(fd);
513 		return ENOTSOCK;
514 	}
515 	*sop = fp->f_data;
516 	return 0;
517 }
518 
519 /*
520  * Look up the file structure corresponding to a file descriptor
521  * and return it with a reference held on the file, not the
522  * descriptor.
523  *
524  * This is heavyweight and only used when accessing descriptors
525  * from a foreign process.  The caller must ensure that `p' does
526  * not exit or fork across this call.
527  *
528  * To release the file (not descriptor) reference, use closef().
529  */
530 file_t *
531 fd_getfile2(proc_t *p, unsigned fd)
532 {
533 	filedesc_t *fdp;
534 	fdfile_t *ff;
535 	file_t *fp;
536 	fdtab_t *dt;
537 
538 	fdp = p->p_fd;
539 	mutex_enter(&fdp->fd_lock);
540 	dt = fdp->fd_dt;
541 	if (fd >= dt->dt_nfiles) {
542 		mutex_exit(&fdp->fd_lock);
543 		return NULL;
544 	}
545 	if ((ff = dt->dt_ff[fd]) == NULL) {
546 		mutex_exit(&fdp->fd_lock);
547 		return NULL;
548 	}
549 	if ((fp = ff->ff_file) == NULL) {
550 		mutex_exit(&fdp->fd_lock);
551 		return NULL;
552 	}
553 	mutex_enter(&fp->f_lock);
554 	fp->f_count++;
555 	mutex_exit(&fp->f_lock);
556 	mutex_exit(&fdp->fd_lock);
557 
558 	return fp;
559 }
560 
561 /*
562  * Internal form of close.  Must be called with a reference to the
563  * descriptor, and will drop the reference.  When all descriptor
564  * references are dropped, releases the descriptor slot and a single
565  * reference to the file structure.
566  */
567 int
568 fd_close(unsigned fd)
569 {
570 	struct flock lf;
571 	filedesc_t *fdp;
572 	fdfile_t *ff;
573 	file_t *fp;
574 	proc_t *p;
575 	lwp_t *l;
576 	u_int refcnt;
577 
578 	l = curlwp;
579 	p = l->l_proc;
580 	fdp = l->l_fd;
581 	ff = fdp->fd_dt->dt_ff[fd];
582 
583 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
584 
585 	mutex_enter(&fdp->fd_lock);
586 	KASSERT((ff->ff_refcnt & FR_MASK) > 0);
587 	if (__predict_false(ff->ff_file == NULL)) {
588 		/*
589 		 * Another user of the file is already closing, and is
590 		 * waiting for other users of the file to drain.  Release
591 		 * our reference, and wake up the closer.
592 		 */
593 		atomic_dec_uint(&ff->ff_refcnt);
594 		cv_broadcast(&ff->ff_closing);
595 		mutex_exit(&fdp->fd_lock);
596 
597 		/*
598 		 * An application error, so pretend that the descriptor
599 		 * was already closed.  We can't safely wait for it to
600 		 * be closed without potentially deadlocking.
601 		 */
602 		return (EBADF);
603 	}
604 	KASSERT((ff->ff_refcnt & FR_CLOSING) == 0);
605 
606 	/*
607 	 * There may be multiple users of this file within the process.
608 	 * Notify existing and new users that the file is closing.  This
609 	 * will prevent them from adding additional uses to this file
610 	 * while we are closing it.
611 	 */
612 	fp = ff->ff_file;
613 	ff->ff_file = NULL;
614 	ff->ff_exclose = false;
615 
616 	/*
617 	 * We expect the caller to hold a descriptor reference - drop it.
618 	 * The reference count may increase beyond zero at this point due
619 	 * to an erroneous descriptor reference by an application, but
620 	 * fd_getfile() will notice that the file is being closed and drop
621 	 * the reference again.
622 	 */
623 	if (fdp->fd_refcnt == 1) {
624 		/* Single threaded. */
625 		refcnt = --(ff->ff_refcnt);
626 	} else {
627 		/* Multi threaded. */
628 #ifndef __HAVE_ATOMIC_AS_MEMBAR
629 		membar_producer();
630 #endif
631 		refcnt = atomic_dec_uint_nv(&ff->ff_refcnt);
632 	}
633 	if (__predict_false(refcnt != 0)) {
634 		/*
635 		 * Wait for other references to drain.  This is typically
636 		 * an application error - the descriptor is being closed
637 		 * while still in use.
638 		 * (Or just a threaded application trying to unblock its
639 		 * thread that sleeps in (say) accept()).
640 		 */
641 		atomic_or_uint(&ff->ff_refcnt, FR_CLOSING);
642 
643 		/*
644 		 * Remove any knotes attached to the file.  A knote
645 		 * attached to the descriptor can hold references on it.
646 		 */
647 		mutex_exit(&fdp->fd_lock);
648 		if (!SLIST_EMPTY(&ff->ff_knlist)) {
649 			knote_fdclose(fd);
650 		}
651 
652 		/*
653 		 * Since the file system code doesn't know which fd
654 		 * each request came from (think dup()), we have to
655 		 * ask it to return ERESTART for any long-term blocks.
656 		 * The re-entry through read/write/etc will detect the
657 		 * closed fd and return EBAFD.
658 		 * Blocked partial writes may return a short length.
659 		 */
660 		(*fp->f_ops->fo_restart)(fp);
661 		mutex_enter(&fdp->fd_lock);
662 
663 		/*
664 		 * We need to see the count drop to zero at least once,
665 		 * in order to ensure that all pre-existing references
666 		 * have been drained.  New references past this point are
667 		 * of no interest.
668 		 * XXX (dsl) this may need to call fo_restart() after a
669 		 * timeout to guarantee that all the system calls exit.
670 		 */
671 		while ((ff->ff_refcnt & FR_MASK) != 0) {
672 			cv_wait(&ff->ff_closing, &fdp->fd_lock);
673 		}
674 		atomic_and_uint(&ff->ff_refcnt, ~FR_CLOSING);
675 	} else {
676 		/* If no references, there must be no knotes. */
677 		KASSERT(SLIST_EMPTY(&ff->ff_knlist));
678 	}
679 
680 	/*
681 	 * POSIX record locking dictates that any close releases ALL
682 	 * locks owned by this process.  This is handled by setting
683 	 * a flag in the unlock to free ONLY locks obeying POSIX
684 	 * semantics, and not to free BSD-style file locks.
685 	 * If the descriptor was in a message, POSIX-style locks
686 	 * aren't passed with the descriptor.
687 	 */
688 	if (__predict_false((p->p_flag & PK_ADVLOCK) != 0 &&
689 	    fp->f_type == DTYPE_VNODE)) {
690 		lf.l_whence = SEEK_SET;
691 		lf.l_start = 0;
692 		lf.l_len = 0;
693 		lf.l_type = F_UNLCK;
694 		mutex_exit(&fdp->fd_lock);
695 		(void)VOP_ADVLOCK(fp->f_data, p, F_UNLCK, &lf, F_POSIX);
696 		mutex_enter(&fdp->fd_lock);
697 	}
698 
699 	/* Free descriptor slot. */
700 	fd_unused(fdp, fd);
701 	mutex_exit(&fdp->fd_lock);
702 
703 	/* Now drop reference to the file itself. */
704 	return closef(fp);
705 }
706 
707 /*
708  * Duplicate a file descriptor.
709  */
710 int
711 fd_dup(file_t *fp, int minfd, int *newp, bool exclose)
712 {
713 	proc_t *p = curproc;
714 	int error;
715 
716 	while ((error = fd_alloc(p, minfd, newp)) != 0) {
717 		if (error != ENOSPC) {
718 			return error;
719 		}
720 		fd_tryexpand(p);
721 	}
722 
723 	curlwp->l_fd->fd_dt->dt_ff[*newp]->ff_exclose = exclose;
724 	fd_affix(p, fp, *newp);
725 	return 0;
726 }
727 
728 /*
729  * dup2 operation.
730  */
731 int
732 fd_dup2(file_t *fp, unsigned new)
733 {
734 	filedesc_t *fdp = curlwp->l_fd;
735 	fdfile_t *ff;
736 	fdtab_t *dt;
737 
738 	/*
739 	 * Ensure there are enough slots in the descriptor table,
740 	 * and allocate an fdfile_t up front in case we need it.
741 	 */
742 	while (new >= fdp->fd_dt->dt_nfiles) {
743 		fd_tryexpand(curproc);
744 	}
745 	ff = pool_cache_get(fdfile_cache, PR_WAITOK);
746 
747 	/*
748 	 * If there is already a file open, close it.  If the file is
749 	 * half open, wait for it to be constructed before closing it.
750 	 * XXX Potential for deadlock here?
751 	 */
752 	mutex_enter(&fdp->fd_lock);
753 	while (fd_isused(fdp, new)) {
754 		mutex_exit(&fdp->fd_lock);
755 		if (fd_getfile(new) != NULL) {
756 			(void)fd_close(new);
757 		} else {
758 			/*
759 			 * Crummy, but unlikely to happen.
760 			 * Can occur if we interrupt another
761 			 * thread while it is opening a file.
762 			 */
763 			kpause("dup2", false, 1, NULL);
764 		}
765 		mutex_enter(&fdp->fd_lock);
766 	}
767 	dt = fdp->fd_dt;
768 	if (dt->dt_ff[new] == NULL) {
769 		KASSERT(new >= NDFDFILE);
770 		dt->dt_ff[new] = ff;
771 		ff = NULL;
772 	}
773 	fd_used(fdp, new);
774 	mutex_exit(&fdp->fd_lock);
775 
776 	/* Slot is now allocated.  Insert copy of the file. */
777 	fd_affix(curproc, fp, new);
778 	if (ff != NULL) {
779 		pool_cache_put(fdfile_cache, ff);
780 	}
781 	return 0;
782 }
783 
784 /*
785  * Drop reference to a file structure.
786  */
787 int
788 closef(file_t *fp)
789 {
790 	struct flock lf;
791 	int error;
792 
793 	/*
794 	 * Drop reference.  If referenced elsewhere it's still open
795 	 * and we have nothing more to do.
796 	 */
797 	mutex_enter(&fp->f_lock);
798 	KASSERT(fp->f_count > 0);
799 	if (--fp->f_count > 0) {
800 		mutex_exit(&fp->f_lock);
801 		return 0;
802 	}
803 	KASSERT(fp->f_count == 0);
804 	mutex_exit(&fp->f_lock);
805 
806 	/* We held the last reference - release locks, close and free. */
807 	if ((fp->f_flag & FHASLOCK) && fp->f_type == DTYPE_VNODE) {
808 		lf.l_whence = SEEK_SET;
809 		lf.l_start = 0;
810 		lf.l_len = 0;
811 		lf.l_type = F_UNLCK;
812 		(void)VOP_ADVLOCK(fp->f_data, fp, F_UNLCK, &lf, F_FLOCK);
813 	}
814 	if (fp->f_ops != NULL) {
815 		error = (*fp->f_ops->fo_close)(fp);
816 	} else {
817 		error = 0;
818 	}
819 	KASSERT(fp->f_count == 0);
820 	KASSERT(fp->f_cred != NULL);
821 	pool_cache_put(file_cache, fp);
822 
823 	return error;
824 }
825 
826 /*
827  * Allocate a file descriptor for the process.
828  */
829 int
830 fd_alloc(proc_t *p, int want, int *result)
831 {
832 	filedesc_t *fdp = p->p_fd;
833 	int i, lim, last, error;
834 	u_int off, new;
835 	fdtab_t *dt;
836 
837 	KASSERT(p == curproc || p == &proc0);
838 
839 	/*
840 	 * Search for a free descriptor starting at the higher
841 	 * of want or fd_freefile.
842 	 */
843 	mutex_enter(&fdp->fd_lock);
844 	fd_checkmaps(fdp);
845 	dt = fdp->fd_dt;
846 	KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
847 	lim = min((int)p->p_rlimit[RLIMIT_NOFILE].rlim_cur, maxfiles);
848 	last = min(dt->dt_nfiles, lim);
849 	for (;;) {
850 		if ((i = want) < fdp->fd_freefile)
851 			i = fdp->fd_freefile;
852 		off = i >> NDENTRYSHIFT;
853 		new = fd_next_zero(fdp, fdp->fd_himap, off,
854 		    (last + NDENTRIES - 1) >> NDENTRYSHIFT);
855 		if (new == -1)
856 			break;
857 		i = fd_next_zero(fdp, &fdp->fd_lomap[new],
858 		    new > off ? 0 : i & NDENTRYMASK, NDENTRIES);
859 		if (i == -1) {
860 			/*
861 			 * Free file descriptor in this block was
862 			 * below want, try again with higher want.
863 			 */
864 			want = (new + 1) << NDENTRYSHIFT;
865 			continue;
866 		}
867 		i += (new << NDENTRYSHIFT);
868 		if (i >= last) {
869 			break;
870 		}
871 		if (dt->dt_ff[i] == NULL) {
872 			KASSERT(i >= NDFDFILE);
873 			dt->dt_ff[i] = pool_cache_get(fdfile_cache, PR_WAITOK);
874 		}
875 		KASSERT(dt->dt_ff[i]->ff_file == NULL);
876 		fd_used(fdp, i);
877 		if (want <= fdp->fd_freefile) {
878 			fdp->fd_freefile = i;
879 		}
880 		*result = i;
881 		KASSERT(i >= NDFDFILE ||
882 		    dt->dt_ff[i] == (fdfile_t *)fdp->fd_dfdfile[i]);
883 		fd_checkmaps(fdp);
884 		mutex_exit(&fdp->fd_lock);
885 		return 0;
886 	}
887 
888 	/* No space in current array.  Let the caller expand and retry. */
889 	error = (dt->dt_nfiles >= lim) ? EMFILE : ENOSPC;
890 	mutex_exit(&fdp->fd_lock);
891 	return error;
892 }
893 
894 /*
895  * Allocate memory for a descriptor table.
896  */
897 static fdtab_t *
898 fd_dtab_alloc(int n)
899 {
900 	fdtab_t *dt;
901 	size_t sz;
902 
903 	KASSERT(n > NDFILE);
904 
905 	sz = sizeof(*dt) + (n - NDFILE) * sizeof(dt->dt_ff[0]);
906 	dt = kmem_alloc(sz, KM_SLEEP);
907 #ifdef DIAGNOSTIC
908 	memset(dt, 0xff, sz);
909 #endif
910 	dt->dt_nfiles = n;
911 	dt->dt_link = NULL;
912 	return dt;
913 }
914 
915 /*
916  * Free a descriptor table, and all tables linked for deferred free.
917  */
918 static void
919 fd_dtab_free(fdtab_t *dt)
920 {
921 	fdtab_t *next;
922 	size_t sz;
923 
924 	do {
925 		next = dt->dt_link;
926 		KASSERT(dt->dt_nfiles > NDFILE);
927 		sz = sizeof(*dt) +
928 		    (dt->dt_nfiles - NDFILE) * sizeof(dt->dt_ff[0]);
929 #ifdef DIAGNOSTIC
930 		memset(dt, 0xff, sz);
931 #endif
932 		kmem_free(dt, sz);
933 		dt = next;
934 	} while (dt != NULL);
935 }
936 
937 /*
938  * Allocate descriptor bitmap.
939  */
940 static void
941 fd_map_alloc(int n, uint32_t **lo, uint32_t **hi)
942 {
943 	uint8_t *ptr;
944 	size_t szlo, szhi;
945 
946 	KASSERT(n > NDENTRIES);
947 
948 	szlo = NDLOSLOTS(n) * sizeof(uint32_t);
949 	szhi = NDHISLOTS(n) * sizeof(uint32_t);
950 	ptr = kmem_alloc(szlo + szhi, KM_SLEEP);
951 	*lo = (uint32_t *)ptr;
952 	*hi = (uint32_t *)(ptr + szlo);
953 }
954 
955 /*
956  * Free descriptor bitmap.
957  */
958 static void
959 fd_map_free(int n, uint32_t *lo, uint32_t *hi)
960 {
961 	size_t szlo, szhi;
962 
963 	KASSERT(n > NDENTRIES);
964 
965 	szlo = NDLOSLOTS(n) * sizeof(uint32_t);
966 	szhi = NDHISLOTS(n) * sizeof(uint32_t);
967 	KASSERT(hi == (uint32_t *)((uint8_t *)lo + szlo));
968 	kmem_free(lo, szlo + szhi);
969 }
970 
971 /*
972  * Expand a process' descriptor table.
973  */
974 void
975 fd_tryexpand(proc_t *p)
976 {
977 	filedesc_t *fdp;
978 	int i, numfiles, oldnfiles;
979 	fdtab_t *newdt, *dt;
980 	uint32_t *newhimap, *newlomap;
981 
982 	KASSERT(p == curproc || p == &proc0);
983 
984 	fdp = p->p_fd;
985 	newhimap = NULL;
986 	newlomap = NULL;
987 	oldnfiles = fdp->fd_dt->dt_nfiles;
988 
989 	if (oldnfiles < NDEXTENT)
990 		numfiles = NDEXTENT;
991 	else
992 		numfiles = 2 * oldnfiles;
993 
994 	newdt = fd_dtab_alloc(numfiles);
995 	if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
996 		fd_map_alloc(numfiles, &newlomap, &newhimap);
997 	}
998 
999 	mutex_enter(&fdp->fd_lock);
1000 	dt = fdp->fd_dt;
1001 	KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1002 	if (dt->dt_nfiles != oldnfiles) {
1003 		/* fdp changed; caller must retry */
1004 		mutex_exit(&fdp->fd_lock);
1005 		fd_dtab_free(newdt);
1006 		if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1007 			fd_map_free(numfiles, newlomap, newhimap);
1008 		}
1009 		return;
1010 	}
1011 
1012 	/* Copy the existing descriptor table and zero the new portion. */
1013 	i = sizeof(fdfile_t *) * oldnfiles;
1014 	memcpy(newdt->dt_ff, dt->dt_ff, i);
1015 	memset((uint8_t *)newdt->dt_ff + i, 0,
1016 	    numfiles * sizeof(fdfile_t *) - i);
1017 
1018 	/*
1019 	 * Link old descriptor array into list to be discarded.  We defer
1020 	 * freeing until the last reference to the descriptor table goes
1021 	 * away (usually process exit).  This allows us to do lockless
1022 	 * lookups in fd_getfile().
1023 	 */
1024 	if (oldnfiles > NDFILE) {
1025 		if (fdp->fd_refcnt > 1) {
1026 			newdt->dt_link = dt;
1027 		} else {
1028 			fd_dtab_free(dt);
1029 		}
1030 	}
1031 
1032 	if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1033 		i = NDHISLOTS(oldnfiles) * sizeof(uint32_t);
1034 		memcpy(newhimap, fdp->fd_himap, i);
1035 		memset((uint8_t *)newhimap + i, 0,
1036 		    NDHISLOTS(numfiles) * sizeof(uint32_t) - i);
1037 
1038 		i = NDLOSLOTS(oldnfiles) * sizeof(uint32_t);
1039 		memcpy(newlomap, fdp->fd_lomap, i);
1040 		memset((uint8_t *)newlomap + i, 0,
1041 		    NDLOSLOTS(numfiles) * sizeof(uint32_t) - i);
1042 
1043 		if (NDHISLOTS(oldnfiles) > NDHISLOTS(NDFILE)) {
1044 			fd_map_free(oldnfiles, fdp->fd_lomap, fdp->fd_himap);
1045 		}
1046 		fdp->fd_himap = newhimap;
1047 		fdp->fd_lomap = newlomap;
1048 	}
1049 
1050 	/*
1051 	 * All other modifications must become globally visible before
1052 	 * the change to fd_dt.  See fd_getfile().
1053 	 */
1054 	membar_producer();
1055 	fdp->fd_dt = newdt;
1056 	KASSERT(newdt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1057 	fd_checkmaps(fdp);
1058 	mutex_exit(&fdp->fd_lock);
1059 }
1060 
1061 /*
1062  * Create a new open file structure and allocate a file descriptor
1063  * for the current process.
1064  */
1065 int
1066 fd_allocfile(file_t **resultfp, int *resultfd)
1067 {
1068 	proc_t *p = curproc;
1069 	kauth_cred_t cred;
1070 	file_t *fp;
1071 	int error;
1072 
1073 	while ((error = fd_alloc(p, 0, resultfd)) != 0) {
1074 		if (error != ENOSPC) {
1075 			return error;
1076 		}
1077 		fd_tryexpand(p);
1078 	}
1079 
1080 	fp = pool_cache_get(file_cache, PR_WAITOK);
1081 	if (fp == NULL) {
1082 		return ENFILE;
1083 	}
1084 	KASSERT(fp->f_count == 0);
1085 	KASSERT(fp->f_msgcount == 0);
1086 	KASSERT(fp->f_unpcount == 0);
1087 
1088 	/* Replace cached credentials if not what we need. */
1089 	cred = curlwp->l_cred;
1090 	if (__predict_false(cred != fp->f_cred)) {
1091 		kauth_cred_free(fp->f_cred);
1092 		kauth_cred_hold(cred);
1093 		fp->f_cred = cred;
1094 	}
1095 
1096 	/*
1097 	 * Don't allow recycled files to be scanned.
1098 	 * See uipc_usrreq.c.
1099 	 */
1100 	if (__predict_false((fp->f_flag & FSCAN) != 0)) {
1101 		mutex_enter(&fp->f_lock);
1102 		atomic_and_uint(&fp->f_flag, ~FSCAN);
1103 		mutex_exit(&fp->f_lock);
1104 	}
1105 
1106 	fp->f_advice = 0;
1107 	fp->f_offset = 0;
1108 	*resultfp = fp;
1109 
1110 	return 0;
1111 }
1112 
1113 /*
1114  * Successful creation of a new descriptor: make visible to the process.
1115  */
1116 void
1117 fd_affix(proc_t *p, file_t *fp, unsigned fd)
1118 {
1119 	fdfile_t *ff;
1120 	filedesc_t *fdp;
1121 
1122 	KASSERT(p == curproc || p == &proc0);
1123 
1124 	/* Add a reference to the file structure. */
1125 	mutex_enter(&fp->f_lock);
1126 	fp->f_count++;
1127 	mutex_exit(&fp->f_lock);
1128 
1129 	/*
1130 	 * Insert the new file into the descriptor slot.
1131 	 *
1132 	 * The memory barriers provided by lock activity in this routine
1133 	 * ensure that any updates to the file structure become globally
1134 	 * visible before the file becomes visible to other LWPs in the
1135 	 * current process.
1136 	 */
1137 	fdp = p->p_fd;
1138 	ff = fdp->fd_dt->dt_ff[fd];
1139 
1140 	KASSERT(ff != NULL);
1141 	KASSERT(ff->ff_file == NULL);
1142 	KASSERT(ff->ff_allocated);
1143 	KASSERT(fd_isused(fdp, fd));
1144 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1145 
1146 	/* No need to lock in order to make file initially visible. */
1147 	ff->ff_file = fp;
1148 }
1149 
1150 /*
1151  * Abort creation of a new descriptor: free descriptor slot and file.
1152  */
1153 void
1154 fd_abort(proc_t *p, file_t *fp, unsigned fd)
1155 {
1156 	filedesc_t *fdp;
1157 	fdfile_t *ff;
1158 
1159 	KASSERT(p == curproc || p == &proc0);
1160 
1161 	fdp = p->p_fd;
1162 	ff = fdp->fd_dt->dt_ff[fd];
1163 
1164 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1165 
1166 	mutex_enter(&fdp->fd_lock);
1167 	KASSERT(fd_isused(fdp, fd));
1168 	fd_unused(fdp, fd);
1169 	mutex_exit(&fdp->fd_lock);
1170 
1171 	if (fp != NULL) {
1172 		KASSERT(fp->f_count == 0);
1173 		KASSERT(fp->f_cred != NULL);
1174 		pool_cache_put(file_cache, fp);
1175 	}
1176 }
1177 
1178 static int
1179 file_ctor(void *arg, void *obj, int flags)
1180 {
1181 	file_t *fp = obj;
1182 
1183 	memset(fp, 0, sizeof(*fp));
1184 
1185 	mutex_enter(&filelist_lock);
1186 	if (__predict_false(nfiles >= maxfiles)) {
1187 		mutex_exit(&filelist_lock);
1188 		tablefull("file", "increase kern.maxfiles or MAXFILES");
1189 		return ENFILE;
1190 	}
1191 	nfiles++;
1192 	LIST_INSERT_HEAD(&filehead, fp, f_list);
1193 	mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1194 	fp->f_cred = curlwp->l_cred;
1195 	kauth_cred_hold(fp->f_cred);
1196 	mutex_exit(&filelist_lock);
1197 
1198 	return 0;
1199 }
1200 
1201 static void
1202 file_dtor(void *arg, void *obj)
1203 {
1204 	file_t *fp = obj;
1205 
1206 	mutex_enter(&filelist_lock);
1207 	nfiles--;
1208 	LIST_REMOVE(fp, f_list);
1209 	mutex_exit(&filelist_lock);
1210 
1211 	kauth_cred_free(fp->f_cred);
1212 	mutex_destroy(&fp->f_lock);
1213 }
1214 
1215 static int
1216 fdfile_ctor(void *arg, void *obj, int flags)
1217 {
1218 	fdfile_t *ff = obj;
1219 
1220 	memset(ff, 0, sizeof(*ff));
1221 	cv_init(&ff->ff_closing, "fdclose");
1222 
1223 	return 0;
1224 }
1225 
1226 static void
1227 fdfile_dtor(void *arg, void *obj)
1228 {
1229 	fdfile_t *ff = obj;
1230 
1231 	cv_destroy(&ff->ff_closing);
1232 }
1233 
1234 file_t *
1235 fgetdummy(void)
1236 {
1237 	file_t *fp;
1238 
1239 	fp = kmem_zalloc(sizeof(*fp), KM_SLEEP);
1240 	if (fp != NULL) {
1241 		mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1242 	}
1243 	return fp;
1244 }
1245 
1246 void
1247 fputdummy(file_t *fp)
1248 {
1249 
1250 	mutex_destroy(&fp->f_lock);
1251 	kmem_free(fp, sizeof(*fp));
1252 }
1253 
1254 /*
1255  * Create an initial filedesc structure.
1256  */
1257 filedesc_t *
1258 fd_init(filedesc_t *fdp)
1259 {
1260 #ifdef DIAGNOSTIC
1261 	unsigned fd;
1262 #endif
1263 
1264 	if (__predict_true(fdp == NULL)) {
1265 		fdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1266 	} else {
1267 		KASSERT(fdp == &filedesc0);
1268 		filedesc_ctor(NULL, fdp, PR_WAITOK);
1269 	}
1270 
1271 #ifdef DIAGNOSTIC
1272 	KASSERT(fdp->fd_lastfile == -1);
1273 	KASSERT(fdp->fd_lastkqfile == -1);
1274 	KASSERT(fdp->fd_knhash == NULL);
1275 	KASSERT(fdp->fd_freefile == 0);
1276 	KASSERT(fdp->fd_exclose == false);
1277 	KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1278 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1279 	for (fd = 0; fd < NDFDFILE; fd++) {
1280 		KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] ==
1281 		    (fdfile_t *)fdp->fd_dfdfile[fd]);
1282 	}
1283 	for (fd = NDFDFILE; fd < NDFILE; fd++) {
1284 		KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] == NULL);
1285 	}
1286 	KASSERT(fdp->fd_himap == fdp->fd_dhimap);
1287 	KASSERT(fdp->fd_lomap == fdp->fd_dlomap);
1288 #endif	/* DIAGNOSTIC */
1289 
1290 	fdp->fd_refcnt = 1;
1291 	fd_checkmaps(fdp);
1292 
1293 	return fdp;
1294 }
1295 
1296 /*
1297  * Initialize a file descriptor table.
1298  */
1299 static int
1300 filedesc_ctor(void *arg, void *obj, int flag)
1301 {
1302 	filedesc_t *fdp = obj;
1303 	fdfile_t **ffp;
1304 	int i;
1305 
1306 	memset(fdp, 0, sizeof(*fdp));
1307 	mutex_init(&fdp->fd_lock, MUTEX_DEFAULT, IPL_NONE);
1308 	fdp->fd_lastfile = -1;
1309 	fdp->fd_lastkqfile = -1;
1310 	fdp->fd_dt = &fdp->fd_dtbuiltin;
1311 	fdp->fd_dtbuiltin.dt_nfiles = NDFILE;
1312 	fdp->fd_himap = fdp->fd_dhimap;
1313 	fdp->fd_lomap = fdp->fd_dlomap;
1314 
1315 	CTASSERT(sizeof(fdp->fd_dfdfile[0]) >= sizeof(fdfile_t));
1316 	for (i = 0, ffp = fdp->fd_dt->dt_ff; i < NDFDFILE; i++, ffp++) {
1317 		*ffp = (fdfile_t *)fdp->fd_dfdfile[i];
1318 		(void)fdfile_ctor(NULL, fdp->fd_dfdfile[i], PR_WAITOK);
1319 	}
1320 
1321 	return 0;
1322 }
1323 
1324 static void
1325 filedesc_dtor(void *arg, void *obj)
1326 {
1327 	filedesc_t *fdp = obj;
1328 	int i;
1329 
1330 	for (i = 0; i < NDFDFILE; i++) {
1331 		fdfile_dtor(NULL, fdp->fd_dfdfile[i]);
1332 	}
1333 
1334 	mutex_destroy(&fdp->fd_lock);
1335 }
1336 
1337 /*
1338  * Make p share curproc's filedesc structure.
1339  */
1340 void
1341 fd_share(struct proc *p)
1342 {
1343 	filedesc_t *fdp;
1344 
1345 	fdp = curlwp->l_fd;
1346 	p->p_fd = fdp;
1347 	atomic_inc_uint(&fdp->fd_refcnt);
1348 }
1349 
1350 /*
1351  * Acquire a hold on a filedesc structure.
1352  */
1353 void
1354 fd_hold(lwp_t *l)
1355 {
1356 	filedesc_t *fdp = l->l_fd;
1357 
1358 	atomic_inc_uint(&fdp->fd_refcnt);
1359 }
1360 
1361 /*
1362  * Copy a filedesc structure.
1363  */
1364 filedesc_t *
1365 fd_copy(void)
1366 {
1367 	filedesc_t *newfdp, *fdp;
1368 	fdfile_t *ff, **ffp, **nffp, *ff2;
1369 	int i, j, numfiles, lastfile, newlast;
1370 	file_t *fp;
1371 	fdtab_t *newdt;
1372 
1373 	fdp = curproc->p_fd;
1374 	newfdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1375 	newfdp->fd_refcnt = 1;
1376 
1377 #ifdef DIAGNOSTIC
1378 	KASSERT(newfdp->fd_lastfile == -1);
1379 	KASSERT(newfdp->fd_lastkqfile == -1);
1380 	KASSERT(newfdp->fd_knhash == NULL);
1381 	KASSERT(newfdp->fd_freefile == 0);
1382 	KASSERT(newfdp->fd_exclose == false);
1383 	KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1384 	KASSERT(newfdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1385 	for (i = 0; i < NDFDFILE; i++) {
1386 		KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] ==
1387 		    (fdfile_t *)&newfdp->fd_dfdfile[i]);
1388 	}
1389 	for (i = NDFDFILE; i < NDFILE; i++) {
1390 		KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] == NULL);
1391 	}
1392 #endif	/* DIAGNOSTIC */
1393 
1394 	mutex_enter(&fdp->fd_lock);
1395 	fd_checkmaps(fdp);
1396 	numfiles = fdp->fd_dt->dt_nfiles;
1397 	lastfile = fdp->fd_lastfile;
1398 
1399 	/*
1400 	 * If the number of open files fits in the internal arrays
1401 	 * of the open file structure, use them, otherwise allocate
1402 	 * additional memory for the number of descriptors currently
1403 	 * in use.
1404 	 */
1405 	if (lastfile < NDFILE) {
1406 		i = NDFILE;
1407 		newdt = newfdp->fd_dt;
1408 		KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1409 	} else {
1410 		/*
1411 		 * Compute the smallest multiple of NDEXTENT needed
1412 		 * for the file descriptors currently in use,
1413 		 * allowing the table to shrink.
1414 		 */
1415 		i = numfiles;
1416 		while (i >= 2 * NDEXTENT && i > lastfile * 2) {
1417 			i /= 2;
1418 		}
1419 		KASSERT(i > NDFILE);
1420 		newdt = fd_dtab_alloc(i);
1421 		newfdp->fd_dt = newdt;
1422 		memcpy(newdt->dt_ff, newfdp->fd_dtbuiltin.dt_ff,
1423 		    NDFDFILE * sizeof(fdfile_t **));
1424 		memset(newdt->dt_ff + NDFDFILE, 0,
1425 		    (i - NDFDFILE) * sizeof(fdfile_t **));
1426 	}
1427 	if (NDHISLOTS(i) <= NDHISLOTS(NDFILE)) {
1428 		newfdp->fd_himap = newfdp->fd_dhimap;
1429 		newfdp->fd_lomap = newfdp->fd_dlomap;
1430 	} else {
1431 		fd_map_alloc(i, &newfdp->fd_lomap, &newfdp->fd_himap);
1432 		KASSERT(i >= NDENTRIES * NDENTRIES);
1433 		memset(newfdp->fd_himap, 0, NDHISLOTS(i)*sizeof(uint32_t));
1434 		memset(newfdp->fd_lomap, 0, NDLOSLOTS(i)*sizeof(uint32_t));
1435 	}
1436 	newfdp->fd_freefile = fdp->fd_freefile;
1437 	newfdp->fd_exclose = fdp->fd_exclose;
1438 
1439 	ffp = fdp->fd_dt->dt_ff;
1440 	nffp = newdt->dt_ff;
1441 	newlast = -1;
1442 	for (i = 0; i <= (int)lastfile; i++, ffp++, nffp++) {
1443 		KASSERT(i >= NDFDFILE ||
1444 		    *nffp == (fdfile_t *)newfdp->fd_dfdfile[i]);
1445 		ff = *ffp;
1446 		if (ff == NULL || (fp = ff->ff_file) == NULL) {
1447 			/* Descriptor unused, or descriptor half open. */
1448 			KASSERT(!fd_isused(newfdp, i));
1449 			continue;
1450 		}
1451 		if (__predict_false(fp->f_type == DTYPE_KQUEUE)) {
1452 			/* kqueue descriptors cannot be copied. */
1453 			if (i < newfdp->fd_freefile) {
1454 				newfdp->fd_freefile = i;
1455 			}
1456 			continue;
1457 		}
1458 		/* It's active: add a reference to the file. */
1459 		mutex_enter(&fp->f_lock);
1460 		fp->f_count++;
1461 		mutex_exit(&fp->f_lock);
1462 
1463 		/* Allocate an fdfile_t to represent it. */
1464 		if (i >= NDFDFILE) {
1465 			ff2 = pool_cache_get(fdfile_cache, PR_WAITOK);
1466 			*nffp = ff2;
1467 		} else {
1468 			ff2 = newdt->dt_ff[i];
1469 		}
1470 		ff2->ff_file = fp;
1471 		ff2->ff_exclose = ff->ff_exclose;
1472 		ff2->ff_allocated = true;
1473 
1474 		/* Fix up bitmaps. */
1475 		j = i >> NDENTRYSHIFT;
1476 		KASSERT((newfdp->fd_lomap[j] & (1 << (i & NDENTRYMASK))) == 0);
1477 		newfdp->fd_lomap[j] |= 1 << (i & NDENTRYMASK);
1478 		if (__predict_false(newfdp->fd_lomap[j] == ~0)) {
1479 			KASSERT((newfdp->fd_himap[j >> NDENTRYSHIFT] &
1480 			    (1 << (j & NDENTRYMASK))) == 0);
1481 			newfdp->fd_himap[j >> NDENTRYSHIFT] |=
1482 			    1 << (j & NDENTRYMASK);
1483 		}
1484 		newlast = i;
1485 	}
1486 	KASSERT(newdt->dt_ff[0] == (fdfile_t *)newfdp->fd_dfdfile[0]);
1487 	newfdp->fd_lastfile = newlast;
1488 	fd_checkmaps(newfdp);
1489 	mutex_exit(&fdp->fd_lock);
1490 
1491 	return newfdp;
1492 }
1493 
1494 /*
1495  * Release a filedesc structure.
1496  */
1497 void
1498 fd_free(void)
1499 {
1500 	fdfile_t *ff;
1501 	file_t *fp;
1502 	int fd, nf;
1503 	fdtab_t *dt;
1504 	lwp_t * const l = curlwp;
1505 	filedesc_t * const fdp = l->l_fd;
1506 	const bool noadvlock = (l->l_proc->p_flag & PK_ADVLOCK) == 0;
1507 
1508 	KASSERT(fdp->fd_dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1509 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1510 	KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1511 
1512 #ifndef __HAVE_ATOMIC_AS_MEMBAR
1513 	membar_exit();
1514 #endif
1515 	if (atomic_dec_uint_nv(&fdp->fd_refcnt) > 0)
1516 		return;
1517 
1518 	/*
1519 	 * Close any files that the process holds open.
1520 	 */
1521 	dt = fdp->fd_dt;
1522 	fd_checkmaps(fdp);
1523 #ifdef DEBUG
1524 	fdp->fd_refcnt = -1; /* see fd_checkmaps */
1525 #endif
1526 	for (fd = 0, nf = dt->dt_nfiles; fd < nf; fd++) {
1527 		ff = dt->dt_ff[fd];
1528 		KASSERT(fd >= NDFDFILE ||
1529 		    ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1530 		if (ff == NULL)
1531 			continue;
1532 		if ((fp = ff->ff_file) != NULL) {
1533 			/*
1534 			 * Must use fd_close() here if there is
1535 			 * a reference from kqueue or we might have posix
1536 			 * advisory locks.
1537 			 */
1538 			if (__predict_true(ff->ff_refcnt == 0) &&
1539 			    (noadvlock || fp->f_type != DTYPE_VNODE)) {
1540 				ff->ff_file = NULL;
1541 				ff->ff_exclose = false;
1542 				ff->ff_allocated = false;
1543 				closef(fp);
1544 			} else {
1545 				ff->ff_refcnt++;
1546 				fd_close(fd);
1547 			}
1548 		}
1549 		KASSERT(ff->ff_refcnt == 0);
1550 		KASSERT(ff->ff_file == NULL);
1551 		KASSERT(!ff->ff_exclose);
1552 		KASSERT(!ff->ff_allocated);
1553 		if (fd >= NDFDFILE) {
1554 			pool_cache_put(fdfile_cache, ff);
1555 			dt->dt_ff[fd] = NULL;
1556 		}
1557 	}
1558 
1559 	/*
1560 	 * Clean out the descriptor table for the next user and return
1561 	 * to the cache.
1562 	 */
1563 	if (__predict_false(dt != &fdp->fd_dtbuiltin)) {
1564 		fd_dtab_free(fdp->fd_dt);
1565 		/* Otherwise, done above. */
1566 		memset(&fdp->fd_dtbuiltin.dt_ff[NDFDFILE], 0,
1567 		    (NDFILE - NDFDFILE) * sizeof(fdp->fd_dtbuiltin.dt_ff[0]));
1568 		fdp->fd_dt = &fdp->fd_dtbuiltin;
1569 	}
1570 	if (__predict_false(NDHISLOTS(nf) > NDHISLOTS(NDFILE))) {
1571 		KASSERT(fdp->fd_himap != fdp->fd_dhimap);
1572 		KASSERT(fdp->fd_lomap != fdp->fd_dlomap);
1573 		fd_map_free(nf, fdp->fd_lomap, fdp->fd_himap);
1574 	}
1575 	if (__predict_false(fdp->fd_knhash != NULL)) {
1576 		hashdone(fdp->fd_knhash, HASH_LIST, fdp->fd_knhashmask);
1577 		fdp->fd_knhash = NULL;
1578 		fdp->fd_knhashmask = 0;
1579 	} else {
1580 		KASSERT(fdp->fd_knhashmask == 0);
1581 	}
1582 	fdp->fd_dt = &fdp->fd_dtbuiltin;
1583 	fdp->fd_lastkqfile = -1;
1584 	fdp->fd_lastfile = -1;
1585 	fdp->fd_freefile = 0;
1586 	fdp->fd_exclose = false;
1587 	memset(&fdp->fd_startzero, 0, sizeof(*fdp) -
1588 	    offsetof(filedesc_t, fd_startzero));
1589 	fdp->fd_himap = fdp->fd_dhimap;
1590 	fdp->fd_lomap = fdp->fd_dlomap;
1591 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1592 	KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1593 	KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1594 #ifdef DEBUG
1595 	fdp->fd_refcnt = 0; /* see fd_checkmaps */
1596 #endif
1597 	fd_checkmaps(fdp);
1598 	pool_cache_put(filedesc_cache, fdp);
1599 }
1600 
1601 /*
1602  * File Descriptor pseudo-device driver (/dev/fd/).
1603  *
1604  * Opening minor device N dup()s the file (if any) connected to file
1605  * descriptor N belonging to the calling process.  Note that this driver
1606  * consists of only the ``open()'' routine, because all subsequent
1607  * references to this file will be direct to the other driver.
1608  */
1609 static int
1610 filedescopen(dev_t dev, int mode, int type, lwp_t *l)
1611 {
1612 
1613 	/*
1614 	 * XXX Kludge: set dupfd to contain the value of the
1615 	 * the file descriptor being sought for duplication. The error
1616 	 * return ensures that the vnode for this device will be released
1617 	 * by vn_open. Open will detect this special error and take the
1618 	 * actions in fd_dupopen below. Other callers of vn_open or VOP_OPEN
1619 	 * will simply report the error.
1620 	 */
1621 	l->l_dupfd = minor(dev);	/* XXX */
1622 	return EDUPFD;
1623 }
1624 
1625 /*
1626  * Duplicate the specified descriptor to a free descriptor.
1627  */
1628 int
1629 fd_dupopen(int old, int *new, int mode, int error)
1630 {
1631 	filedesc_t *fdp;
1632 	fdfile_t *ff;
1633 	file_t *fp;
1634 	fdtab_t *dt;
1635 
1636 	if ((fp = fd_getfile(old)) == NULL) {
1637 		return EBADF;
1638 	}
1639 	fdp = curlwp->l_fd;
1640 	dt = fdp->fd_dt;
1641 	ff = dt->dt_ff[old];
1642 
1643 	/*
1644 	 * There are two cases of interest here.
1645 	 *
1646 	 * For EDUPFD simply dup (old) to file descriptor
1647 	 * (new) and return.
1648 	 *
1649 	 * For EMOVEFD steal away the file structure from (old) and
1650 	 * store it in (new).  (old) is effectively closed by
1651 	 * this operation.
1652 	 *
1653 	 * Any other error code is just returned.
1654 	 */
1655 	switch (error) {
1656 	case EDUPFD:
1657 		/*
1658 		 * Check that the mode the file is being opened for is a
1659 		 * subset of the mode of the existing descriptor.
1660 		 */
1661 		if (((mode & (FREAD|FWRITE)) | fp->f_flag) != fp->f_flag) {
1662 			error = EACCES;
1663 			break;
1664 		}
1665 
1666 		/* Copy it. */
1667 		error = fd_dup(fp, 0, new, ff->ff_exclose);
1668 		break;
1669 
1670 	case EMOVEFD:
1671 		/* Copy it. */
1672 		error = fd_dup(fp, 0, new, ff->ff_exclose);
1673 		if (error != 0) {
1674 			break;
1675 		}
1676 
1677 		/* Steal away the file pointer from 'old'. */
1678 		(void)fd_close(old);
1679 		return 0;
1680 	}
1681 
1682 	fd_putfile(old);
1683 	return error;
1684 }
1685 
1686 /*
1687  * Close open files on exec.
1688  */
1689 void
1690 fd_closeexec(void)
1691 {
1692 	proc_t *p;
1693 	filedesc_t *fdp;
1694 	fdfile_t *ff;
1695 	lwp_t *l;
1696 	fdtab_t *dt;
1697 	int fd;
1698 
1699 	l = curlwp;
1700 	p = l->l_proc;
1701 	fdp = p->p_fd;
1702 
1703 	if (fdp->fd_refcnt > 1) {
1704 		fdp = fd_copy();
1705 		fd_free();
1706 		p->p_fd = fdp;
1707 		l->l_fd = fdp;
1708 	}
1709 	if (!fdp->fd_exclose) {
1710 		return;
1711 	}
1712 	fdp->fd_exclose = false;
1713 	dt = fdp->fd_dt;
1714 
1715 	for (fd = 0; fd <= fdp->fd_lastfile; fd++) {
1716 		if ((ff = dt->dt_ff[fd]) == NULL) {
1717 			KASSERT(fd >= NDFDFILE);
1718 			continue;
1719 		}
1720 		KASSERT(fd >= NDFDFILE ||
1721 		    ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1722 		if (ff->ff_file == NULL)
1723 			continue;
1724 		if (ff->ff_exclose) {
1725 			/*
1726 			 * We need a reference to close the file.
1727 			 * No other threads can see the fdfile_t at
1728 			 * this point, so don't bother locking.
1729 			 */
1730 			KASSERT((ff->ff_refcnt & FR_CLOSING) == 0);
1731 			ff->ff_refcnt++;
1732 			fd_close(fd);
1733 		}
1734 	}
1735 }
1736 
1737 /*
1738  * Sets descriptor owner. If the owner is a process, 'pgid'
1739  * is set to positive value, process ID. If the owner is process group,
1740  * 'pgid' is set to -pg_id.
1741  */
1742 int
1743 fsetown(pid_t *pgid, u_long cmd, const void *data)
1744 {
1745 	pid_t id = *(const pid_t *)data;
1746 	int error;
1747 
1748 	switch (cmd) {
1749 	case TIOCSPGRP:
1750 		if (id < 0)
1751 			return EINVAL;
1752 		id = -id;
1753 		break;
1754 	default:
1755 		break;
1756 	}
1757 	if (id > 0) {
1758 		mutex_enter(proc_lock);
1759 		error = proc_find(id) ? 0 : ESRCH;
1760 		mutex_exit(proc_lock);
1761 	} else if (id < 0) {
1762 		error = pgid_in_session(curproc, -id);
1763 	} else {
1764 		error = 0;
1765 	}
1766 	if (!error) {
1767 		*pgid = id;
1768 	}
1769 	return error;
1770 }
1771 
1772 void
1773 fd_set_exclose(struct lwp *l, int fd, bool exclose)
1774 {
1775 	filedesc_t *fdp = l->l_fd;
1776 	fdfile_t *ff = fdp->fd_dt->dt_ff[fd];
1777 
1778 	ff->ff_exclose = exclose;
1779 	if (exclose)
1780 		fdp->fd_exclose = true;
1781 }
1782 
1783 /*
1784  * Return descriptor owner information. If the value is positive,
1785  * it's process ID. If it's negative, it's process group ID and
1786  * needs the sign removed before use.
1787  */
1788 int
1789 fgetown(pid_t pgid, u_long cmd, void *data)
1790 {
1791 
1792 	switch (cmd) {
1793 	case TIOCGPGRP:
1794 		*(int *)data = -pgid;
1795 		break;
1796 	default:
1797 		*(int *)data = pgid;
1798 		break;
1799 	}
1800 	return 0;
1801 }
1802 
1803 /*
1804  * Send signal to descriptor owner, either process or process group.
1805  */
1806 void
1807 fownsignal(pid_t pgid, int signo, int code, int band, void *fdescdata)
1808 {
1809 	ksiginfo_t ksi;
1810 
1811 	KASSERT(!cpu_intr_p());
1812 
1813 	if (pgid == 0) {
1814 		return;
1815 	}
1816 
1817 	KSI_INIT(&ksi);
1818 	ksi.ksi_signo = signo;
1819 	ksi.ksi_code = code;
1820 	ksi.ksi_band = band;
1821 
1822 	mutex_enter(proc_lock);
1823 	if (pgid > 0) {
1824 		struct proc *p1;
1825 
1826 		p1 = proc_find(pgid);
1827 		if (p1 != NULL) {
1828 			kpsignal(p1, &ksi, fdescdata);
1829 		}
1830 	} else {
1831 		struct pgrp *pgrp;
1832 
1833 		KASSERT(pgid < 0);
1834 		pgrp = pgrp_find(-pgid);
1835 		if (pgrp != NULL) {
1836 			kpgsignal(pgrp, &ksi, fdescdata, 0);
1837 		}
1838 	}
1839 	mutex_exit(proc_lock);
1840 }
1841 
1842 int
1843 fd_clone(file_t *fp, unsigned fd, int flag, const struct fileops *fops,
1844 	 void *data)
1845 {
1846 
1847 	fp->f_flag = flag;
1848 	fp->f_type = DTYPE_MISC;
1849 	fp->f_ops = fops;
1850 	fp->f_data = data;
1851 	curlwp->l_dupfd = fd;
1852 	fd_affix(curproc, fp, fd);
1853 
1854 	return EMOVEFD;
1855 }
1856 
1857 int
1858 fnullop_fcntl(file_t *fp, u_int cmd, void *data)
1859 {
1860 
1861 	if (cmd == F_SETFL)
1862 		return 0;
1863 
1864 	return EOPNOTSUPP;
1865 }
1866 
1867 int
1868 fnullop_poll(file_t *fp, int which)
1869 {
1870 
1871 	return 0;
1872 }
1873 
1874 int
1875 fnullop_kqfilter(file_t *fp, struct knote *kn)
1876 {
1877 
1878 	return 0;
1879 }
1880 
1881 void
1882 fnullop_restart(file_t *fp)
1883 {
1884 
1885 }
1886 
1887 int
1888 fbadop_read(file_t *fp, off_t *offset, struct uio *uio,
1889 	    kauth_cred_t cred, int flags)
1890 {
1891 
1892 	return EOPNOTSUPP;
1893 }
1894 
1895 int
1896 fbadop_write(file_t *fp, off_t *offset, struct uio *uio,
1897 	     kauth_cred_t cred, int flags)
1898 {
1899 
1900 	return EOPNOTSUPP;
1901 }
1902 
1903 int
1904 fbadop_ioctl(file_t *fp, u_long com, void *data)
1905 {
1906 
1907 	return EOPNOTSUPP;
1908 }
1909 
1910 int
1911 fbadop_stat(file_t *fp, struct stat *sb)
1912 {
1913 
1914 	return EOPNOTSUPP;
1915 }
1916 
1917 int
1918 fbadop_close(file_t *fp)
1919 {
1920 
1921 	return EOPNOTSUPP;
1922 }
1923 
1924 /*
1925  * sysctl routines pertaining to file descriptors
1926  */
1927 
1928 /* Initialized in sysctl_init() for now... */
1929 extern kmutex_t sysctl_file_marker_lock;
1930 static u_int sysctl_file_marker = 1;
1931 
1932 /*
1933  * Expects to be called with proc_lock and sysctl_file_marker_lock locked.
1934  */
1935 static void
1936 sysctl_file_marker_reset(void)
1937 {
1938 	struct proc *p;
1939 
1940 	PROCLIST_FOREACH(p, &allproc) {
1941 		struct filedesc *fd = p->p_fd;
1942 		fdtab_t *dt;
1943 		u_int i;
1944 
1945 		mutex_enter(&fd->fd_lock);
1946 		dt = fd->fd_dt;
1947 		for (i = 0; i < dt->dt_nfiles; i++) {
1948 			struct file *fp;
1949 			fdfile_t *ff;
1950 
1951 			if ((ff = dt->dt_ff[i]) == NULL) {
1952 				continue;
1953 			}
1954 			if ((fp = ff->ff_file) == NULL) {
1955 				continue;
1956 			}
1957 			fp->f_marker = 0;
1958 		}
1959 		mutex_exit(&fd->fd_lock);
1960 	}
1961 }
1962 
1963 /*
1964  * sysctl helper routine for kern.file pseudo-subtree.
1965  */
1966 static int
1967 sysctl_kern_file(SYSCTLFN_ARGS)
1968 {
1969 	int error;
1970 	size_t buflen;
1971 	struct file *fp, fbuf;
1972 	char *start, *where;
1973 	struct proc *p;
1974 
1975 	start = where = oldp;
1976 	buflen = *oldlenp;
1977 
1978 	if (where == NULL) {
1979 		/*
1980 		 * overestimate by 10 files
1981 		 */
1982 		*oldlenp = sizeof(filehead) + (nfiles + 10) *
1983 		    sizeof(struct file);
1984 		return 0;
1985 	}
1986 
1987 	/*
1988 	 * first sysctl_copyout filehead
1989 	 */
1990 	if (buflen < sizeof(filehead)) {
1991 		*oldlenp = 0;
1992 		return 0;
1993 	}
1994 	sysctl_unlock();
1995 	error = sysctl_copyout(l, &filehead, where, sizeof(filehead));
1996 	if (error) {
1997 		sysctl_relock();
1998 		return error;
1999 	}
2000 	buflen -= sizeof(filehead);
2001 	where += sizeof(filehead);
2002 
2003 	/*
2004 	 * followed by an array of file structures
2005 	 */
2006 	mutex_enter(&sysctl_file_marker_lock);
2007 	mutex_enter(proc_lock);
2008 	PROCLIST_FOREACH(p, &allproc) {
2009 		struct filedesc *fd;
2010 		fdtab_t *dt;
2011 		u_int i;
2012 
2013 		if (p->p_stat == SIDL) {
2014 			/* skip embryonic processes */
2015 			continue;
2016 		}
2017 		mutex_enter(p->p_lock);
2018 		error = kauth_authorize_process(l->l_cred,
2019 		    KAUTH_PROCESS_CANSEE, p,
2020 		    KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2021 		    NULL, NULL);
2022 		mutex_exit(p->p_lock);
2023 		if (error != 0) {
2024 			/*
2025 			 * Don't leak kauth retval if we're silently
2026 			 * skipping this entry.
2027 			 */
2028 			error = 0;
2029 			continue;
2030 		}
2031 
2032 		/*
2033 		 * Grab a hold on the process.
2034 		 */
2035 		if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2036 			continue;
2037 		}
2038 		mutex_exit(proc_lock);
2039 
2040 		fd = p->p_fd;
2041 		mutex_enter(&fd->fd_lock);
2042 		dt = fd->fd_dt;
2043 		for (i = 0; i < dt->dt_nfiles; i++) {
2044 			fdfile_t *ff;
2045 
2046 			if ((ff = dt->dt_ff[i]) == NULL) {
2047 				continue;
2048 			}
2049 			if ((fp = ff->ff_file) == NULL) {
2050 				continue;
2051 			}
2052 
2053 			mutex_enter(&fp->f_lock);
2054 
2055 			if ((fp->f_count == 0) ||
2056 			    (fp->f_marker == sysctl_file_marker)) {
2057 				mutex_exit(&fp->f_lock);
2058 				continue;
2059 			}
2060 
2061 			/* Check that we have enough space. */
2062 			if (buflen < sizeof(struct file)) {
2063 				*oldlenp = where - start;
2064 				mutex_exit(&fp->f_lock);
2065 				error = ENOMEM;
2066 				break;
2067 			}
2068 
2069 			memcpy(&fbuf, fp, sizeof(fbuf));
2070 			mutex_exit(&fp->f_lock);
2071 			error = sysctl_copyout(l, &fbuf, where, sizeof(fbuf));
2072 			if (error) {
2073 				break;
2074 			}
2075 			buflen -= sizeof(struct file);
2076 			where += sizeof(struct file);
2077 
2078 			fp->f_marker = sysctl_file_marker;
2079 		}
2080 		mutex_exit(&fd->fd_lock);
2081 
2082 		/*
2083 		 * Release reference to process.
2084 		 */
2085 		mutex_enter(proc_lock);
2086 		rw_exit(&p->p_reflock);
2087 
2088 		if (error)
2089 			break;
2090 	}
2091 
2092 	sysctl_file_marker++;
2093 	/* Reset all markers if wrapped. */
2094 	if (sysctl_file_marker == 0) {
2095 		sysctl_file_marker_reset();
2096 		sysctl_file_marker++;
2097 	}
2098 
2099 	mutex_exit(proc_lock);
2100 	mutex_exit(&sysctl_file_marker_lock);
2101 
2102 	*oldlenp = where - start;
2103 	sysctl_relock();
2104 	return error;
2105 }
2106 
2107 /*
2108  * sysctl helper function for kern.file2
2109  */
2110 static int
2111 sysctl_kern_file2(SYSCTLFN_ARGS)
2112 {
2113 	struct proc *p;
2114 	struct file *fp;
2115 	struct filedesc *fd;
2116 	struct kinfo_file kf;
2117 	char *dp;
2118 	u_int i, op;
2119 	size_t len, needed, elem_size, out_size;
2120 	int error, arg, elem_count;
2121 	fdfile_t *ff;
2122 	fdtab_t *dt;
2123 
2124 	if (namelen == 1 && name[0] == CTL_QUERY)
2125 		return sysctl_query(SYSCTLFN_CALL(rnode));
2126 
2127 	if (namelen != 4)
2128 		return EINVAL;
2129 
2130 	error = 0;
2131 	dp = oldp;
2132 	len = (oldp != NULL) ? *oldlenp : 0;
2133 	op = name[0];
2134 	arg = name[1];
2135 	elem_size = name[2];
2136 	elem_count = name[3];
2137 	out_size = MIN(sizeof(kf), elem_size);
2138 	needed = 0;
2139 
2140 	if (elem_size < 1 || elem_count < 0)
2141 		return EINVAL;
2142 
2143 	switch (op) {
2144 	case KERN_FILE_BYFILE:
2145 	case KERN_FILE_BYPID:
2146 		/*
2147 		 * We're traversing the process list in both cases; the BYFILE
2148 		 * case does additional work of keeping track of files already
2149 		 * looked at.
2150 		 */
2151 
2152 		/* doesn't use arg so it must be zero */
2153 		if ((op == KERN_FILE_BYFILE) && (arg != 0))
2154 			return EINVAL;
2155 
2156 		if ((op == KERN_FILE_BYPID) && (arg < -1))
2157 			/* -1 means all processes */
2158 			return EINVAL;
2159 
2160 		sysctl_unlock();
2161 		if (op == KERN_FILE_BYFILE)
2162 			mutex_enter(&sysctl_file_marker_lock);
2163 		mutex_enter(proc_lock);
2164 		PROCLIST_FOREACH(p, &allproc) {
2165 			if (p->p_stat == SIDL) {
2166 				/* skip embryonic processes */
2167 				continue;
2168 			}
2169 			if (arg > 0 && p->p_pid != arg) {
2170 				/* pick only the one we want */
2171 				/* XXX want 0 to mean "kernel files" */
2172 				continue;
2173 			}
2174 			mutex_enter(p->p_lock);
2175 			error = kauth_authorize_process(l->l_cred,
2176 			    KAUTH_PROCESS_CANSEE, p,
2177 			    KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2178 			    NULL, NULL);
2179 			mutex_exit(p->p_lock);
2180 			if (error != 0) {
2181 				/*
2182 				 * Don't leak kauth retval if we're silently
2183 				 * skipping this entry.
2184 				 */
2185 				error = 0;
2186 				continue;
2187 			}
2188 
2189 			/*
2190 			 * Grab a hold on the process.
2191 			 */
2192 			if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2193 				continue;
2194 			}
2195 			mutex_exit(proc_lock);
2196 
2197 			fd = p->p_fd;
2198 			mutex_enter(&fd->fd_lock);
2199 			dt = fd->fd_dt;
2200 			for (i = 0; i < dt->dt_nfiles; i++) {
2201 				if ((ff = dt->dt_ff[i]) == NULL) {
2202 					continue;
2203 				}
2204 				if ((fp = ff->ff_file) == NULL) {
2205 					continue;
2206 				}
2207 
2208 				if ((op == KERN_FILE_BYFILE) &&
2209 				    (fp->f_marker == sysctl_file_marker)) {
2210 					continue;
2211 				}
2212 				if (len >= elem_size && elem_count > 0) {
2213 					mutex_enter(&fp->f_lock);
2214 					fill_file(&kf, fp, ff, i, p->p_pid);
2215 					mutex_exit(&fp->f_lock);
2216 					mutex_exit(&fd->fd_lock);
2217 					error = sysctl_copyout(l,
2218 					    &kf, dp, out_size);
2219 					mutex_enter(&fd->fd_lock);
2220 					if (error)
2221 						break;
2222 					dp += elem_size;
2223 					len -= elem_size;
2224 				}
2225 				if (op == KERN_FILE_BYFILE)
2226 					fp->f_marker = sysctl_file_marker;
2227 				needed += elem_size;
2228 				if (elem_count > 0 && elem_count != INT_MAX)
2229 					elem_count--;
2230 			}
2231 			mutex_exit(&fd->fd_lock);
2232 
2233 			/*
2234 			 * Release reference to process.
2235 			 */
2236 			mutex_enter(proc_lock);
2237 			rw_exit(&p->p_reflock);
2238 		}
2239 		if (op == KERN_FILE_BYFILE) {
2240 			sysctl_file_marker++;
2241 
2242 			/* Reset all markers if wrapped. */
2243 			if (sysctl_file_marker == 0) {
2244 				sysctl_file_marker_reset();
2245 				sysctl_file_marker++;
2246 			}
2247 		}
2248 		mutex_exit(proc_lock);
2249 		if (op == KERN_FILE_BYFILE)
2250 			mutex_exit(&sysctl_file_marker_lock);
2251 		sysctl_relock();
2252 		break;
2253 	default:
2254 		return EINVAL;
2255 	}
2256 
2257 	if (oldp == NULL)
2258 		needed += KERN_FILESLOP * elem_size;
2259 	*oldlenp = needed;
2260 
2261 	return error;
2262 }
2263 
2264 static void
2265 fill_file(struct kinfo_file *kp, const file_t *fp, const fdfile_t *ff,
2266 	  int i, pid_t pid)
2267 {
2268 
2269 	memset(kp, 0, sizeof(*kp));
2270 
2271 	kp->ki_fileaddr =	PTRTOUINT64(fp);
2272 	kp->ki_flag =		fp->f_flag;
2273 	kp->ki_iflags =		0;
2274 	kp->ki_ftype =		fp->f_type;
2275 	kp->ki_count =		fp->f_count;
2276 	kp->ki_msgcount =	fp->f_msgcount;
2277 	kp->ki_fucred =		PTRTOUINT64(fp->f_cred);
2278 	kp->ki_fuid =		kauth_cred_geteuid(fp->f_cred);
2279 	kp->ki_fgid =		kauth_cred_getegid(fp->f_cred);
2280 	kp->ki_fops =		PTRTOUINT64(fp->f_ops);
2281 	kp->ki_foffset =	fp->f_offset;
2282 	kp->ki_fdata =		PTRTOUINT64(fp->f_data);
2283 
2284 	/* vnode information to glue this file to something */
2285 	if (fp->f_type == DTYPE_VNODE) {
2286 		struct vnode *vp = (struct vnode *)fp->f_data;
2287 
2288 		kp->ki_vun =	PTRTOUINT64(vp->v_un.vu_socket);
2289 		kp->ki_vsize =	vp->v_size;
2290 		kp->ki_vtype =	vp->v_type;
2291 		kp->ki_vtag =	vp->v_tag;
2292 		kp->ki_vdata =	PTRTOUINT64(vp->v_data);
2293 	}
2294 
2295 	/* process information when retrieved via KERN_FILE_BYPID */
2296 	if (ff != NULL) {
2297 		kp->ki_pid =		pid;
2298 		kp->ki_fd =		i;
2299 		kp->ki_ofileflags =	ff->ff_exclose;
2300 		kp->ki_usecount =	ff->ff_refcnt;
2301 	}
2302 }
2303