xref: /netbsd-src/sys/kern/kern_descrip.c (revision ca453df649ce9db45b64d73678ba06cbccf9aa11)
1 /*	$NetBSD: kern_descrip.c,v 1.216 2011/07/15 14:50:19 christos Exp $	*/
2 
3 /*-
4  * Copyright (c) 2008, 2009 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Andrew Doran.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 /*
33  * Copyright (c) 1982, 1986, 1989, 1991, 1993
34  *	The Regents of the University of California.  All rights reserved.
35  * (c) UNIX System Laboratories, Inc.
36  * All or some portions of this file are derived from material licensed
37  * to the University of California by American Telephone and Telegraph
38  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
39  * the permission of UNIX System Laboratories, Inc.
40  *
41  * Redistribution and use in source and binary forms, with or without
42  * modification, are permitted provided that the following conditions
43  * are met:
44  * 1. Redistributions of source code must retain the above copyright
45  *    notice, this list of conditions and the following disclaimer.
46  * 2. Redistributions in binary form must reproduce the above copyright
47  *    notice, this list of conditions and the following disclaimer in the
48  *    documentation and/or other materials provided with the distribution.
49  * 3. Neither the name of the University nor the names of its contributors
50  *    may be used to endorse or promote products derived from this software
51  *    without specific prior written permission.
52  *
53  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
54  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
55  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
56  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
57  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
58  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
59  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
60  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
61  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
62  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
63  * SUCH DAMAGE.
64  *
65  *	@(#)kern_descrip.c	8.8 (Berkeley) 2/14/95
66  */
67 
68 /*
69  * File descriptor management.
70  */
71 
72 #include <sys/cdefs.h>
73 __KERNEL_RCSID(0, "$NetBSD: kern_descrip.c,v 1.216 2011/07/15 14:50:19 christos Exp $");
74 
75 #include <sys/param.h>
76 #include <sys/systm.h>
77 #include <sys/filedesc.h>
78 #include <sys/kernel.h>
79 #include <sys/proc.h>
80 #include <sys/file.h>
81 #include <sys/socket.h>
82 #include <sys/socketvar.h>
83 #include <sys/stat.h>
84 #include <sys/ioctl.h>
85 #include <sys/fcntl.h>
86 #include <sys/pool.h>
87 #include <sys/unistd.h>
88 #include <sys/resourcevar.h>
89 #include <sys/conf.h>
90 #include <sys/event.h>
91 #include <sys/kauth.h>
92 #include <sys/atomic.h>
93 #include <sys/syscallargs.h>
94 #include <sys/cpu.h>
95 #include <sys/kmem.h>
96 #include <sys/vnode.h>
97 #include <sys/sysctl.h>
98 #include <sys/ktrace.h>
99 
100 /*
101  * A list (head) of open files, counter, and lock protecting them.
102  */
103 struct filelist		filehead	__cacheline_aligned;
104 static u_int		nfiles		__cacheline_aligned;
105 kmutex_t		filelist_lock	__cacheline_aligned;
106 
107 static pool_cache_t	filedesc_cache	__read_mostly;
108 static pool_cache_t	file_cache	__read_mostly;
109 static pool_cache_t	fdfile_cache	__read_mostly;
110 
111 static int	file_ctor(void *, void *, int);
112 static void	file_dtor(void *, void *);
113 static int	fdfile_ctor(void *, void *, int);
114 static void	fdfile_dtor(void *, void *);
115 static int	filedesc_ctor(void *, void *, int);
116 static void	filedesc_dtor(void *, void *);
117 static int	filedescopen(dev_t, int, int, lwp_t *);
118 
119 static int sysctl_kern_file(SYSCTLFN_PROTO);
120 static int sysctl_kern_file2(SYSCTLFN_PROTO);
121 static void fill_file(struct 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, int flags)
733 {
734 	filedesc_t *fdp = curlwp->l_fd;
735 	fdfile_t *ff;
736 	fdtab_t *dt;
737 
738 	if (flags & ~(O_CLOEXEC|O_NONBLOCK))
739 		return EINVAL;
740 	/*
741 	 * Ensure there are enough slots in the descriptor table,
742 	 * and allocate an fdfile_t up front in case we need it.
743 	 */
744 	while (new >= fdp->fd_dt->dt_nfiles) {
745 		fd_tryexpand(curproc);
746 	}
747 	ff = pool_cache_get(fdfile_cache, PR_WAITOK);
748 
749 	/*
750 	 * If there is already a file open, close it.  If the file is
751 	 * half open, wait for it to be constructed before closing it.
752 	 * XXX Potential for deadlock here?
753 	 */
754 	mutex_enter(&fdp->fd_lock);
755 	while (fd_isused(fdp, new)) {
756 		mutex_exit(&fdp->fd_lock);
757 		if (fd_getfile(new) != NULL) {
758 			(void)fd_close(new);
759 		} else {
760 			/*
761 			 * Crummy, but unlikely to happen.
762 			 * Can occur if we interrupt another
763 			 * thread while it is opening a file.
764 			 */
765 			kpause("dup2", false, 1, NULL);
766 		}
767 		mutex_enter(&fdp->fd_lock);
768 	}
769 	dt = fdp->fd_dt;
770 	if (dt->dt_ff[new] == NULL) {
771 		KASSERT(new >= NDFDFILE);
772 		dt->dt_ff[new] = ff;
773 		ff = NULL;
774 	}
775 	fd_used(fdp, new);
776 	mutex_exit(&fdp->fd_lock);
777 
778 	dt->dt_ff[new]->ff_exclose = (flags & O_CLOEXEC) != 0;
779 	fp->f_flag |= flags & FNONBLOCK;
780 	/* Slot is now allocated.  Insert copy of the file. */
781 	fd_affix(curproc, fp, new);
782 	if (ff != NULL) {
783 		pool_cache_put(fdfile_cache, ff);
784 	}
785 	return 0;
786 }
787 
788 /*
789  * Drop reference to a file structure.
790  */
791 int
792 closef(file_t *fp)
793 {
794 	struct flock lf;
795 	int error;
796 
797 	/*
798 	 * Drop reference.  If referenced elsewhere it's still open
799 	 * and we have nothing more to do.
800 	 */
801 	mutex_enter(&fp->f_lock);
802 	KASSERT(fp->f_count > 0);
803 	if (--fp->f_count > 0) {
804 		mutex_exit(&fp->f_lock);
805 		return 0;
806 	}
807 	KASSERT(fp->f_count == 0);
808 	mutex_exit(&fp->f_lock);
809 
810 	/* We held the last reference - release locks, close and free. */
811 	if ((fp->f_flag & FHASLOCK) && fp->f_type == DTYPE_VNODE) {
812 		lf.l_whence = SEEK_SET;
813 		lf.l_start = 0;
814 		lf.l_len = 0;
815 		lf.l_type = F_UNLCK;
816 		(void)VOP_ADVLOCK(fp->f_data, fp, F_UNLCK, &lf, F_FLOCK);
817 	}
818 	if (fp->f_ops != NULL) {
819 		error = (*fp->f_ops->fo_close)(fp);
820 	} else {
821 		error = 0;
822 	}
823 	KASSERT(fp->f_count == 0);
824 	KASSERT(fp->f_cred != NULL);
825 	pool_cache_put(file_cache, fp);
826 
827 	return error;
828 }
829 
830 /*
831  * Allocate a file descriptor for the process.
832  */
833 int
834 fd_alloc(proc_t *p, int want, int *result)
835 {
836 	filedesc_t *fdp = p->p_fd;
837 	int i, lim, last, error;
838 	u_int off, new;
839 	fdtab_t *dt;
840 
841 	KASSERT(p == curproc || p == &proc0);
842 
843 	/*
844 	 * Search for a free descriptor starting at the higher
845 	 * of want or fd_freefile.
846 	 */
847 	mutex_enter(&fdp->fd_lock);
848 	fd_checkmaps(fdp);
849 	dt = fdp->fd_dt;
850 	KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
851 	lim = min((int)p->p_rlimit[RLIMIT_NOFILE].rlim_cur, maxfiles);
852 	last = min(dt->dt_nfiles, lim);
853 	for (;;) {
854 		if ((i = want) < fdp->fd_freefile)
855 			i = fdp->fd_freefile;
856 		off = i >> NDENTRYSHIFT;
857 		new = fd_next_zero(fdp, fdp->fd_himap, off,
858 		    (last + NDENTRIES - 1) >> NDENTRYSHIFT);
859 		if (new == -1)
860 			break;
861 		i = fd_next_zero(fdp, &fdp->fd_lomap[new],
862 		    new > off ? 0 : i & NDENTRYMASK, NDENTRIES);
863 		if (i == -1) {
864 			/*
865 			 * Free file descriptor in this block was
866 			 * below want, try again with higher want.
867 			 */
868 			want = (new + 1) << NDENTRYSHIFT;
869 			continue;
870 		}
871 		i += (new << NDENTRYSHIFT);
872 		if (i >= last) {
873 			break;
874 		}
875 		if (dt->dt_ff[i] == NULL) {
876 			KASSERT(i >= NDFDFILE);
877 			dt->dt_ff[i] = pool_cache_get(fdfile_cache, PR_WAITOK);
878 		}
879 		KASSERT(dt->dt_ff[i]->ff_file == NULL);
880 		fd_used(fdp, i);
881 		if (want <= fdp->fd_freefile) {
882 			fdp->fd_freefile = i;
883 		}
884 		*result = i;
885 		KASSERT(i >= NDFDFILE ||
886 		    dt->dt_ff[i] == (fdfile_t *)fdp->fd_dfdfile[i]);
887 		fd_checkmaps(fdp);
888 		mutex_exit(&fdp->fd_lock);
889 		return 0;
890 	}
891 
892 	/* No space in current array.  Let the caller expand and retry. */
893 	error = (dt->dt_nfiles >= lim) ? EMFILE : ENOSPC;
894 	mutex_exit(&fdp->fd_lock);
895 	return error;
896 }
897 
898 /*
899  * Allocate memory for a descriptor table.
900  */
901 static fdtab_t *
902 fd_dtab_alloc(int n)
903 {
904 	fdtab_t *dt;
905 	size_t sz;
906 
907 	KASSERT(n > NDFILE);
908 
909 	sz = sizeof(*dt) + (n - NDFILE) * sizeof(dt->dt_ff[0]);
910 	dt = kmem_alloc(sz, KM_SLEEP);
911 #ifdef DIAGNOSTIC
912 	memset(dt, 0xff, sz);
913 #endif
914 	dt->dt_nfiles = n;
915 	dt->dt_link = NULL;
916 	return dt;
917 }
918 
919 /*
920  * Free a descriptor table, and all tables linked for deferred free.
921  */
922 static void
923 fd_dtab_free(fdtab_t *dt)
924 {
925 	fdtab_t *next;
926 	size_t sz;
927 
928 	do {
929 		next = dt->dt_link;
930 		KASSERT(dt->dt_nfiles > NDFILE);
931 		sz = sizeof(*dt) +
932 		    (dt->dt_nfiles - NDFILE) * sizeof(dt->dt_ff[0]);
933 #ifdef DIAGNOSTIC
934 		memset(dt, 0xff, sz);
935 #endif
936 		kmem_free(dt, sz);
937 		dt = next;
938 	} while (dt != NULL);
939 }
940 
941 /*
942  * Allocate descriptor bitmap.
943  */
944 static void
945 fd_map_alloc(int n, uint32_t **lo, uint32_t **hi)
946 {
947 	uint8_t *ptr;
948 	size_t szlo, szhi;
949 
950 	KASSERT(n > NDENTRIES);
951 
952 	szlo = NDLOSLOTS(n) * sizeof(uint32_t);
953 	szhi = NDHISLOTS(n) * sizeof(uint32_t);
954 	ptr = kmem_alloc(szlo + szhi, KM_SLEEP);
955 	*lo = (uint32_t *)ptr;
956 	*hi = (uint32_t *)(ptr + szlo);
957 }
958 
959 /*
960  * Free descriptor bitmap.
961  */
962 static void
963 fd_map_free(int n, uint32_t *lo, uint32_t *hi)
964 {
965 	size_t szlo, szhi;
966 
967 	KASSERT(n > NDENTRIES);
968 
969 	szlo = NDLOSLOTS(n) * sizeof(uint32_t);
970 	szhi = NDHISLOTS(n) * sizeof(uint32_t);
971 	KASSERT(hi == (uint32_t *)((uint8_t *)lo + szlo));
972 	kmem_free(lo, szlo + szhi);
973 }
974 
975 /*
976  * Expand a process' descriptor table.
977  */
978 void
979 fd_tryexpand(proc_t *p)
980 {
981 	filedesc_t *fdp;
982 	int i, numfiles, oldnfiles;
983 	fdtab_t *newdt, *dt;
984 	uint32_t *newhimap, *newlomap;
985 
986 	KASSERT(p == curproc || p == &proc0);
987 
988 	fdp = p->p_fd;
989 	newhimap = NULL;
990 	newlomap = NULL;
991 	oldnfiles = fdp->fd_dt->dt_nfiles;
992 
993 	if (oldnfiles < NDEXTENT)
994 		numfiles = NDEXTENT;
995 	else
996 		numfiles = 2 * oldnfiles;
997 
998 	newdt = fd_dtab_alloc(numfiles);
999 	if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1000 		fd_map_alloc(numfiles, &newlomap, &newhimap);
1001 	}
1002 
1003 	mutex_enter(&fdp->fd_lock);
1004 	dt = fdp->fd_dt;
1005 	KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1006 	if (dt->dt_nfiles != oldnfiles) {
1007 		/* fdp changed; caller must retry */
1008 		mutex_exit(&fdp->fd_lock);
1009 		fd_dtab_free(newdt);
1010 		if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1011 			fd_map_free(numfiles, newlomap, newhimap);
1012 		}
1013 		return;
1014 	}
1015 
1016 	/* Copy the existing descriptor table and zero the new portion. */
1017 	i = sizeof(fdfile_t *) * oldnfiles;
1018 	memcpy(newdt->dt_ff, dt->dt_ff, i);
1019 	memset((uint8_t *)newdt->dt_ff + i, 0,
1020 	    numfiles * sizeof(fdfile_t *) - i);
1021 
1022 	/*
1023 	 * Link old descriptor array into list to be discarded.  We defer
1024 	 * freeing until the last reference to the descriptor table goes
1025 	 * away (usually process exit).  This allows us to do lockless
1026 	 * lookups in fd_getfile().
1027 	 */
1028 	if (oldnfiles > NDFILE) {
1029 		if (fdp->fd_refcnt > 1) {
1030 			newdt->dt_link = dt;
1031 		} else {
1032 			fd_dtab_free(dt);
1033 		}
1034 	}
1035 
1036 	if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1037 		i = NDHISLOTS(oldnfiles) * sizeof(uint32_t);
1038 		memcpy(newhimap, fdp->fd_himap, i);
1039 		memset((uint8_t *)newhimap + i, 0,
1040 		    NDHISLOTS(numfiles) * sizeof(uint32_t) - i);
1041 
1042 		i = NDLOSLOTS(oldnfiles) * sizeof(uint32_t);
1043 		memcpy(newlomap, fdp->fd_lomap, i);
1044 		memset((uint8_t *)newlomap + i, 0,
1045 		    NDLOSLOTS(numfiles) * sizeof(uint32_t) - i);
1046 
1047 		if (NDHISLOTS(oldnfiles) > NDHISLOTS(NDFILE)) {
1048 			fd_map_free(oldnfiles, fdp->fd_lomap, fdp->fd_himap);
1049 		}
1050 		fdp->fd_himap = newhimap;
1051 		fdp->fd_lomap = newlomap;
1052 	}
1053 
1054 	/*
1055 	 * All other modifications must become globally visible before
1056 	 * the change to fd_dt.  See fd_getfile().
1057 	 */
1058 	membar_producer();
1059 	fdp->fd_dt = newdt;
1060 	KASSERT(newdt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1061 	fd_checkmaps(fdp);
1062 	mutex_exit(&fdp->fd_lock);
1063 }
1064 
1065 /*
1066  * Create a new open file structure and allocate a file descriptor
1067  * for the current process.
1068  */
1069 int
1070 fd_allocfile(file_t **resultfp, int *resultfd)
1071 {
1072 	proc_t *p = curproc;
1073 	kauth_cred_t cred;
1074 	file_t *fp;
1075 	int error;
1076 
1077 	while ((error = fd_alloc(p, 0, resultfd)) != 0) {
1078 		if (error != ENOSPC) {
1079 			return error;
1080 		}
1081 		fd_tryexpand(p);
1082 	}
1083 
1084 	fp = pool_cache_get(file_cache, PR_WAITOK);
1085 	if (fp == NULL) {
1086 		return ENFILE;
1087 	}
1088 	KASSERT(fp->f_count == 0);
1089 	KASSERT(fp->f_msgcount == 0);
1090 	KASSERT(fp->f_unpcount == 0);
1091 
1092 	/* Replace cached credentials if not what we need. */
1093 	cred = curlwp->l_cred;
1094 	if (__predict_false(cred != fp->f_cred)) {
1095 		kauth_cred_free(fp->f_cred);
1096 		kauth_cred_hold(cred);
1097 		fp->f_cred = cred;
1098 	}
1099 
1100 	/*
1101 	 * Don't allow recycled files to be scanned.
1102 	 * See uipc_usrreq.c.
1103 	 */
1104 	if (__predict_false((fp->f_flag & FSCAN) != 0)) {
1105 		mutex_enter(&fp->f_lock);
1106 		atomic_and_uint(&fp->f_flag, ~FSCAN);
1107 		mutex_exit(&fp->f_lock);
1108 	}
1109 
1110 	fp->f_advice = 0;
1111 	fp->f_offset = 0;
1112 	*resultfp = fp;
1113 
1114 	return 0;
1115 }
1116 
1117 /*
1118  * Successful creation of a new descriptor: make visible to the process.
1119  */
1120 void
1121 fd_affix(proc_t *p, file_t *fp, unsigned fd)
1122 {
1123 	fdfile_t *ff;
1124 	filedesc_t *fdp;
1125 
1126 	KASSERT(p == curproc || p == &proc0);
1127 
1128 	/* Add a reference to the file structure. */
1129 	mutex_enter(&fp->f_lock);
1130 	fp->f_count++;
1131 	mutex_exit(&fp->f_lock);
1132 
1133 	/*
1134 	 * Insert the new file into the descriptor slot.
1135 	 *
1136 	 * The memory barriers provided by lock activity in this routine
1137 	 * ensure that any updates to the file structure become globally
1138 	 * visible before the file becomes visible to other LWPs in the
1139 	 * current process.
1140 	 */
1141 	fdp = p->p_fd;
1142 	ff = fdp->fd_dt->dt_ff[fd];
1143 
1144 	KASSERT(ff != NULL);
1145 	KASSERT(ff->ff_file == NULL);
1146 	KASSERT(ff->ff_allocated);
1147 	KASSERT(fd_isused(fdp, fd));
1148 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1149 
1150 	/* No need to lock in order to make file initially visible. */
1151 	ff->ff_file = fp;
1152 }
1153 
1154 /*
1155  * Abort creation of a new descriptor: free descriptor slot and file.
1156  */
1157 void
1158 fd_abort(proc_t *p, file_t *fp, unsigned fd)
1159 {
1160 	filedesc_t *fdp;
1161 	fdfile_t *ff;
1162 
1163 	KASSERT(p == curproc || p == &proc0);
1164 
1165 	fdp = p->p_fd;
1166 	ff = fdp->fd_dt->dt_ff[fd];
1167 
1168 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1169 
1170 	mutex_enter(&fdp->fd_lock);
1171 	KASSERT(fd_isused(fdp, fd));
1172 	fd_unused(fdp, fd);
1173 	mutex_exit(&fdp->fd_lock);
1174 
1175 	if (fp != NULL) {
1176 		KASSERT(fp->f_count == 0);
1177 		KASSERT(fp->f_cred != NULL);
1178 		pool_cache_put(file_cache, fp);
1179 	}
1180 }
1181 
1182 static int
1183 file_ctor(void *arg, void *obj, int flags)
1184 {
1185 	file_t *fp = obj;
1186 
1187 	memset(fp, 0, sizeof(*fp));
1188 
1189 	mutex_enter(&filelist_lock);
1190 	if (__predict_false(nfiles >= maxfiles)) {
1191 		mutex_exit(&filelist_lock);
1192 		tablefull("file", "increase kern.maxfiles or MAXFILES");
1193 		return ENFILE;
1194 	}
1195 	nfiles++;
1196 	LIST_INSERT_HEAD(&filehead, fp, f_list);
1197 	mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1198 	fp->f_cred = curlwp->l_cred;
1199 	kauth_cred_hold(fp->f_cred);
1200 	mutex_exit(&filelist_lock);
1201 
1202 	return 0;
1203 }
1204 
1205 static void
1206 file_dtor(void *arg, void *obj)
1207 {
1208 	file_t *fp = obj;
1209 
1210 	mutex_enter(&filelist_lock);
1211 	nfiles--;
1212 	LIST_REMOVE(fp, f_list);
1213 	mutex_exit(&filelist_lock);
1214 
1215 	kauth_cred_free(fp->f_cred);
1216 	mutex_destroy(&fp->f_lock);
1217 }
1218 
1219 static int
1220 fdfile_ctor(void *arg, void *obj, int flags)
1221 {
1222 	fdfile_t *ff = obj;
1223 
1224 	memset(ff, 0, sizeof(*ff));
1225 	cv_init(&ff->ff_closing, "fdclose");
1226 
1227 	return 0;
1228 }
1229 
1230 static void
1231 fdfile_dtor(void *arg, void *obj)
1232 {
1233 	fdfile_t *ff = obj;
1234 
1235 	cv_destroy(&ff->ff_closing);
1236 }
1237 
1238 file_t *
1239 fgetdummy(void)
1240 {
1241 	file_t *fp;
1242 
1243 	fp = kmem_zalloc(sizeof(*fp), KM_SLEEP);
1244 	if (fp != NULL) {
1245 		mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1246 	}
1247 	return fp;
1248 }
1249 
1250 void
1251 fputdummy(file_t *fp)
1252 {
1253 
1254 	mutex_destroy(&fp->f_lock);
1255 	kmem_free(fp, sizeof(*fp));
1256 }
1257 
1258 /*
1259  * Create an initial filedesc structure.
1260  */
1261 filedesc_t *
1262 fd_init(filedesc_t *fdp)
1263 {
1264 #ifdef DIAGNOSTIC
1265 	unsigned fd;
1266 #endif
1267 
1268 	if (__predict_true(fdp == NULL)) {
1269 		fdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1270 	} else {
1271 		KASSERT(fdp == &filedesc0);
1272 		filedesc_ctor(NULL, fdp, PR_WAITOK);
1273 	}
1274 
1275 #ifdef DIAGNOSTIC
1276 	KASSERT(fdp->fd_lastfile == -1);
1277 	KASSERT(fdp->fd_lastkqfile == -1);
1278 	KASSERT(fdp->fd_knhash == NULL);
1279 	KASSERT(fdp->fd_freefile == 0);
1280 	KASSERT(fdp->fd_exclose == false);
1281 	KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1282 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1283 	for (fd = 0; fd < NDFDFILE; fd++) {
1284 		KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] ==
1285 		    (fdfile_t *)fdp->fd_dfdfile[fd]);
1286 	}
1287 	for (fd = NDFDFILE; fd < NDFILE; fd++) {
1288 		KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] == NULL);
1289 	}
1290 	KASSERT(fdp->fd_himap == fdp->fd_dhimap);
1291 	KASSERT(fdp->fd_lomap == fdp->fd_dlomap);
1292 #endif	/* DIAGNOSTIC */
1293 
1294 	fdp->fd_refcnt = 1;
1295 	fd_checkmaps(fdp);
1296 
1297 	return fdp;
1298 }
1299 
1300 /*
1301  * Initialize a file descriptor table.
1302  */
1303 static int
1304 filedesc_ctor(void *arg, void *obj, int flag)
1305 {
1306 	filedesc_t *fdp = obj;
1307 	fdfile_t **ffp;
1308 	int i;
1309 
1310 	memset(fdp, 0, sizeof(*fdp));
1311 	mutex_init(&fdp->fd_lock, MUTEX_DEFAULT, IPL_NONE);
1312 	fdp->fd_lastfile = -1;
1313 	fdp->fd_lastkqfile = -1;
1314 	fdp->fd_dt = &fdp->fd_dtbuiltin;
1315 	fdp->fd_dtbuiltin.dt_nfiles = NDFILE;
1316 	fdp->fd_himap = fdp->fd_dhimap;
1317 	fdp->fd_lomap = fdp->fd_dlomap;
1318 
1319 	CTASSERT(sizeof(fdp->fd_dfdfile[0]) >= sizeof(fdfile_t));
1320 	for (i = 0, ffp = fdp->fd_dt->dt_ff; i < NDFDFILE; i++, ffp++) {
1321 		*ffp = (fdfile_t *)fdp->fd_dfdfile[i];
1322 		(void)fdfile_ctor(NULL, fdp->fd_dfdfile[i], PR_WAITOK);
1323 	}
1324 
1325 	return 0;
1326 }
1327 
1328 static void
1329 filedesc_dtor(void *arg, void *obj)
1330 {
1331 	filedesc_t *fdp = obj;
1332 	int i;
1333 
1334 	for (i = 0; i < NDFDFILE; i++) {
1335 		fdfile_dtor(NULL, fdp->fd_dfdfile[i]);
1336 	}
1337 
1338 	mutex_destroy(&fdp->fd_lock);
1339 }
1340 
1341 /*
1342  * Make p share curproc's filedesc structure.
1343  */
1344 void
1345 fd_share(struct proc *p)
1346 {
1347 	filedesc_t *fdp;
1348 
1349 	fdp = curlwp->l_fd;
1350 	p->p_fd = fdp;
1351 	atomic_inc_uint(&fdp->fd_refcnt);
1352 }
1353 
1354 /*
1355  * Acquire a hold on a filedesc structure.
1356  */
1357 void
1358 fd_hold(lwp_t *l)
1359 {
1360 	filedesc_t *fdp = l->l_fd;
1361 
1362 	atomic_inc_uint(&fdp->fd_refcnt);
1363 }
1364 
1365 /*
1366  * Copy a filedesc structure.
1367  */
1368 filedesc_t *
1369 fd_copy(void)
1370 {
1371 	filedesc_t *newfdp, *fdp;
1372 	fdfile_t *ff, **ffp, **nffp, *ff2;
1373 	int i, j, numfiles, lastfile, newlast;
1374 	file_t *fp;
1375 	fdtab_t *newdt;
1376 
1377 	fdp = curproc->p_fd;
1378 	newfdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1379 	newfdp->fd_refcnt = 1;
1380 
1381 #ifdef DIAGNOSTIC
1382 	KASSERT(newfdp->fd_lastfile == -1);
1383 	KASSERT(newfdp->fd_lastkqfile == -1);
1384 	KASSERT(newfdp->fd_knhash == NULL);
1385 	KASSERT(newfdp->fd_freefile == 0);
1386 	KASSERT(newfdp->fd_exclose == false);
1387 	KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1388 	KASSERT(newfdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1389 	for (i = 0; i < NDFDFILE; i++) {
1390 		KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] ==
1391 		    (fdfile_t *)&newfdp->fd_dfdfile[i]);
1392 	}
1393 	for (i = NDFDFILE; i < NDFILE; i++) {
1394 		KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] == NULL);
1395 	}
1396 #endif	/* DIAGNOSTIC */
1397 
1398 	mutex_enter(&fdp->fd_lock);
1399 	fd_checkmaps(fdp);
1400 	numfiles = fdp->fd_dt->dt_nfiles;
1401 	lastfile = fdp->fd_lastfile;
1402 
1403 	/*
1404 	 * If the number of open files fits in the internal arrays
1405 	 * of the open file structure, use them, otherwise allocate
1406 	 * additional memory for the number of descriptors currently
1407 	 * in use.
1408 	 */
1409 	if (lastfile < NDFILE) {
1410 		i = NDFILE;
1411 		newdt = newfdp->fd_dt;
1412 		KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1413 	} else {
1414 		/*
1415 		 * Compute the smallest multiple of NDEXTENT needed
1416 		 * for the file descriptors currently in use,
1417 		 * allowing the table to shrink.
1418 		 */
1419 		i = numfiles;
1420 		while (i >= 2 * NDEXTENT && i > lastfile * 2) {
1421 			i /= 2;
1422 		}
1423 		KASSERT(i > NDFILE);
1424 		newdt = fd_dtab_alloc(i);
1425 		newfdp->fd_dt = newdt;
1426 		memcpy(newdt->dt_ff, newfdp->fd_dtbuiltin.dt_ff,
1427 		    NDFDFILE * sizeof(fdfile_t **));
1428 		memset(newdt->dt_ff + NDFDFILE, 0,
1429 		    (i - NDFDFILE) * sizeof(fdfile_t **));
1430 	}
1431 	if (NDHISLOTS(i) <= NDHISLOTS(NDFILE)) {
1432 		newfdp->fd_himap = newfdp->fd_dhimap;
1433 		newfdp->fd_lomap = newfdp->fd_dlomap;
1434 	} else {
1435 		fd_map_alloc(i, &newfdp->fd_lomap, &newfdp->fd_himap);
1436 		KASSERT(i >= NDENTRIES * NDENTRIES);
1437 		memset(newfdp->fd_himap, 0, NDHISLOTS(i)*sizeof(uint32_t));
1438 		memset(newfdp->fd_lomap, 0, NDLOSLOTS(i)*sizeof(uint32_t));
1439 	}
1440 	newfdp->fd_freefile = fdp->fd_freefile;
1441 	newfdp->fd_exclose = fdp->fd_exclose;
1442 
1443 	ffp = fdp->fd_dt->dt_ff;
1444 	nffp = newdt->dt_ff;
1445 	newlast = -1;
1446 	for (i = 0; i <= (int)lastfile; i++, ffp++, nffp++) {
1447 		KASSERT(i >= NDFDFILE ||
1448 		    *nffp == (fdfile_t *)newfdp->fd_dfdfile[i]);
1449 		ff = *ffp;
1450 		if (ff == NULL || (fp = ff->ff_file) == NULL) {
1451 			/* Descriptor unused, or descriptor half open. */
1452 			KASSERT(!fd_isused(newfdp, i));
1453 			continue;
1454 		}
1455 		if (__predict_false(fp->f_type == DTYPE_KQUEUE)) {
1456 			/* kqueue descriptors cannot be copied. */
1457 			if (i < newfdp->fd_freefile) {
1458 				newfdp->fd_freefile = i;
1459 			}
1460 			continue;
1461 		}
1462 		/* It's active: add a reference to the file. */
1463 		mutex_enter(&fp->f_lock);
1464 		fp->f_count++;
1465 		mutex_exit(&fp->f_lock);
1466 
1467 		/* Allocate an fdfile_t to represent it. */
1468 		if (i >= NDFDFILE) {
1469 			ff2 = pool_cache_get(fdfile_cache, PR_WAITOK);
1470 			*nffp = ff2;
1471 		} else {
1472 			ff2 = newdt->dt_ff[i];
1473 		}
1474 		ff2->ff_file = fp;
1475 		ff2->ff_exclose = ff->ff_exclose;
1476 		ff2->ff_allocated = true;
1477 
1478 		/* Fix up bitmaps. */
1479 		j = i >> NDENTRYSHIFT;
1480 		KASSERT((newfdp->fd_lomap[j] & (1 << (i & NDENTRYMASK))) == 0);
1481 		newfdp->fd_lomap[j] |= 1 << (i & NDENTRYMASK);
1482 		if (__predict_false(newfdp->fd_lomap[j] == ~0)) {
1483 			KASSERT((newfdp->fd_himap[j >> NDENTRYSHIFT] &
1484 			    (1 << (j & NDENTRYMASK))) == 0);
1485 			newfdp->fd_himap[j >> NDENTRYSHIFT] |=
1486 			    1 << (j & NDENTRYMASK);
1487 		}
1488 		newlast = i;
1489 	}
1490 	KASSERT(newdt->dt_ff[0] == (fdfile_t *)newfdp->fd_dfdfile[0]);
1491 	newfdp->fd_lastfile = newlast;
1492 	fd_checkmaps(newfdp);
1493 	mutex_exit(&fdp->fd_lock);
1494 
1495 	return newfdp;
1496 }
1497 
1498 /*
1499  * Release a filedesc structure.
1500  */
1501 void
1502 fd_free(void)
1503 {
1504 	fdfile_t *ff;
1505 	file_t *fp;
1506 	int fd, nf;
1507 	fdtab_t *dt;
1508 	lwp_t * const l = curlwp;
1509 	filedesc_t * const fdp = l->l_fd;
1510 	const bool noadvlock = (l->l_proc->p_flag & PK_ADVLOCK) == 0;
1511 
1512 	KASSERT(fdp->fd_dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1513 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1514 	KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1515 
1516 #ifndef __HAVE_ATOMIC_AS_MEMBAR
1517 	membar_exit();
1518 #endif
1519 	if (atomic_dec_uint_nv(&fdp->fd_refcnt) > 0)
1520 		return;
1521 
1522 	/*
1523 	 * Close any files that the process holds open.
1524 	 */
1525 	dt = fdp->fd_dt;
1526 	fd_checkmaps(fdp);
1527 #ifdef DEBUG
1528 	fdp->fd_refcnt = -1; /* see fd_checkmaps */
1529 #endif
1530 	for (fd = 0, nf = dt->dt_nfiles; fd < nf; fd++) {
1531 		ff = dt->dt_ff[fd];
1532 		KASSERT(fd >= NDFDFILE ||
1533 		    ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1534 		if (ff == NULL)
1535 			continue;
1536 		if ((fp = ff->ff_file) != NULL) {
1537 			/*
1538 			 * Must use fd_close() here if there is
1539 			 * a reference from kqueue or we might have posix
1540 			 * advisory locks.
1541 			 */
1542 			if (__predict_true(ff->ff_refcnt == 0) &&
1543 			    (noadvlock || fp->f_type != DTYPE_VNODE)) {
1544 				ff->ff_file = NULL;
1545 				ff->ff_exclose = false;
1546 				ff->ff_allocated = false;
1547 				closef(fp);
1548 			} else {
1549 				ff->ff_refcnt++;
1550 				fd_close(fd);
1551 			}
1552 		}
1553 		KASSERT(ff->ff_refcnt == 0);
1554 		KASSERT(ff->ff_file == NULL);
1555 		KASSERT(!ff->ff_exclose);
1556 		KASSERT(!ff->ff_allocated);
1557 		if (fd >= NDFDFILE) {
1558 			pool_cache_put(fdfile_cache, ff);
1559 			dt->dt_ff[fd] = NULL;
1560 		}
1561 	}
1562 
1563 	/*
1564 	 * Clean out the descriptor table for the next user and return
1565 	 * to the cache.
1566 	 */
1567 	if (__predict_false(dt != &fdp->fd_dtbuiltin)) {
1568 		fd_dtab_free(fdp->fd_dt);
1569 		/* Otherwise, done above. */
1570 		memset(&fdp->fd_dtbuiltin.dt_ff[NDFDFILE], 0,
1571 		    (NDFILE - NDFDFILE) * sizeof(fdp->fd_dtbuiltin.dt_ff[0]));
1572 		fdp->fd_dt = &fdp->fd_dtbuiltin;
1573 	}
1574 	if (__predict_false(NDHISLOTS(nf) > NDHISLOTS(NDFILE))) {
1575 		KASSERT(fdp->fd_himap != fdp->fd_dhimap);
1576 		KASSERT(fdp->fd_lomap != fdp->fd_dlomap);
1577 		fd_map_free(nf, fdp->fd_lomap, fdp->fd_himap);
1578 	}
1579 	if (__predict_false(fdp->fd_knhash != NULL)) {
1580 		hashdone(fdp->fd_knhash, HASH_LIST, fdp->fd_knhashmask);
1581 		fdp->fd_knhash = NULL;
1582 		fdp->fd_knhashmask = 0;
1583 	} else {
1584 		KASSERT(fdp->fd_knhashmask == 0);
1585 	}
1586 	fdp->fd_dt = &fdp->fd_dtbuiltin;
1587 	fdp->fd_lastkqfile = -1;
1588 	fdp->fd_lastfile = -1;
1589 	fdp->fd_freefile = 0;
1590 	fdp->fd_exclose = false;
1591 	memset(&fdp->fd_startzero, 0, sizeof(*fdp) -
1592 	    offsetof(filedesc_t, fd_startzero));
1593 	fdp->fd_himap = fdp->fd_dhimap;
1594 	fdp->fd_lomap = fdp->fd_dlomap;
1595 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1596 	KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1597 	KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1598 #ifdef DEBUG
1599 	fdp->fd_refcnt = 0; /* see fd_checkmaps */
1600 #endif
1601 	fd_checkmaps(fdp);
1602 	pool_cache_put(filedesc_cache, fdp);
1603 }
1604 
1605 /*
1606  * File Descriptor pseudo-device driver (/dev/fd/).
1607  *
1608  * Opening minor device N dup()s the file (if any) connected to file
1609  * descriptor N belonging to the calling process.  Note that this driver
1610  * consists of only the ``open()'' routine, because all subsequent
1611  * references to this file will be direct to the other driver.
1612  */
1613 static int
1614 filedescopen(dev_t dev, int mode, int type, lwp_t *l)
1615 {
1616 
1617 	/*
1618 	 * XXX Kludge: set dupfd to contain the value of the
1619 	 * the file descriptor being sought for duplication. The error
1620 	 * return ensures that the vnode for this device will be released
1621 	 * by vn_open. Open will detect this special error and take the
1622 	 * actions in fd_dupopen below. Other callers of vn_open or VOP_OPEN
1623 	 * will simply report the error.
1624 	 */
1625 	l->l_dupfd = minor(dev);	/* XXX */
1626 	return EDUPFD;
1627 }
1628 
1629 /*
1630  * Duplicate the specified descriptor to a free descriptor.
1631  */
1632 int
1633 fd_dupopen(int old, int *new, int mode, int error)
1634 {
1635 	filedesc_t *fdp;
1636 	fdfile_t *ff;
1637 	file_t *fp;
1638 	fdtab_t *dt;
1639 
1640 	if ((fp = fd_getfile(old)) == NULL) {
1641 		return EBADF;
1642 	}
1643 	fdp = curlwp->l_fd;
1644 	dt = fdp->fd_dt;
1645 	ff = dt->dt_ff[old];
1646 
1647 	/*
1648 	 * There are two cases of interest here.
1649 	 *
1650 	 * For EDUPFD simply dup (old) to file descriptor
1651 	 * (new) and return.
1652 	 *
1653 	 * For EMOVEFD steal away the file structure from (old) and
1654 	 * store it in (new).  (old) is effectively closed by
1655 	 * this operation.
1656 	 *
1657 	 * Any other error code is just returned.
1658 	 */
1659 	switch (error) {
1660 	case EDUPFD:
1661 		/*
1662 		 * Check that the mode the file is being opened for is a
1663 		 * subset of the mode of the existing descriptor.
1664 		 */
1665 		if (((mode & (FREAD|FWRITE)) | fp->f_flag) != fp->f_flag) {
1666 			error = EACCES;
1667 			break;
1668 		}
1669 
1670 		/* Copy it. */
1671 		error = fd_dup(fp, 0, new, ff->ff_exclose);
1672 		break;
1673 
1674 	case EMOVEFD:
1675 		/* Copy it. */
1676 		error = fd_dup(fp, 0, new, ff->ff_exclose);
1677 		if (error != 0) {
1678 			break;
1679 		}
1680 
1681 		/* Steal away the file pointer from 'old'. */
1682 		(void)fd_close(old);
1683 		return 0;
1684 	}
1685 
1686 	fd_putfile(old);
1687 	return error;
1688 }
1689 
1690 /*
1691  * Close open files on exec.
1692  */
1693 void
1694 fd_closeexec(void)
1695 {
1696 	proc_t *p;
1697 	filedesc_t *fdp;
1698 	fdfile_t *ff;
1699 	lwp_t *l;
1700 	fdtab_t *dt;
1701 	int fd;
1702 
1703 	l = curlwp;
1704 	p = l->l_proc;
1705 	fdp = p->p_fd;
1706 
1707 	if (fdp->fd_refcnt > 1) {
1708 		fdp = fd_copy();
1709 		fd_free();
1710 		p->p_fd = fdp;
1711 		l->l_fd = fdp;
1712 	}
1713 	if (!fdp->fd_exclose) {
1714 		return;
1715 	}
1716 	fdp->fd_exclose = false;
1717 	dt = fdp->fd_dt;
1718 
1719 	for (fd = 0; fd <= fdp->fd_lastfile; fd++) {
1720 		if ((ff = dt->dt_ff[fd]) == NULL) {
1721 			KASSERT(fd >= NDFDFILE);
1722 			continue;
1723 		}
1724 		KASSERT(fd >= NDFDFILE ||
1725 		    ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1726 		if (ff->ff_file == NULL)
1727 			continue;
1728 		if (ff->ff_exclose) {
1729 			/*
1730 			 * We need a reference to close the file.
1731 			 * No other threads can see the fdfile_t at
1732 			 * this point, so don't bother locking.
1733 			 */
1734 			KASSERT((ff->ff_refcnt & FR_CLOSING) == 0);
1735 			ff->ff_refcnt++;
1736 			fd_close(fd);
1737 		}
1738 	}
1739 }
1740 
1741 /*
1742  * Sets descriptor owner. If the owner is a process, 'pgid'
1743  * is set to positive value, process ID. If the owner is process group,
1744  * 'pgid' is set to -pg_id.
1745  */
1746 int
1747 fsetown(pid_t *pgid, u_long cmd, const void *data)
1748 {
1749 	pid_t id = *(const pid_t *)data;
1750 	int error;
1751 
1752 	switch (cmd) {
1753 	case TIOCSPGRP:
1754 		if (id < 0)
1755 			return EINVAL;
1756 		id = -id;
1757 		break;
1758 	default:
1759 		break;
1760 	}
1761 	if (id > 0) {
1762 		mutex_enter(proc_lock);
1763 		error = proc_find(id) ? 0 : ESRCH;
1764 		mutex_exit(proc_lock);
1765 	} else if (id < 0) {
1766 		error = pgid_in_session(curproc, -id);
1767 	} else {
1768 		error = 0;
1769 	}
1770 	if (!error) {
1771 		*pgid = id;
1772 	}
1773 	return error;
1774 }
1775 
1776 void
1777 fd_set_exclose(struct lwp *l, int fd, bool exclose)
1778 {
1779 	filedesc_t *fdp = l->l_fd;
1780 	fdfile_t *ff = fdp->fd_dt->dt_ff[fd];
1781 
1782 	ff->ff_exclose = exclose;
1783 	if (exclose)
1784 		fdp->fd_exclose = true;
1785 }
1786 
1787 /*
1788  * Return descriptor owner information. If the value is positive,
1789  * it's process ID. If it's negative, it's process group ID and
1790  * needs the sign removed before use.
1791  */
1792 int
1793 fgetown(pid_t pgid, u_long cmd, void *data)
1794 {
1795 
1796 	switch (cmd) {
1797 	case TIOCGPGRP:
1798 		*(int *)data = -pgid;
1799 		break;
1800 	default:
1801 		*(int *)data = pgid;
1802 		break;
1803 	}
1804 	return 0;
1805 }
1806 
1807 /*
1808  * Send signal to descriptor owner, either process or process group.
1809  */
1810 void
1811 fownsignal(pid_t pgid, int signo, int code, int band, void *fdescdata)
1812 {
1813 	ksiginfo_t ksi;
1814 
1815 	KASSERT(!cpu_intr_p());
1816 
1817 	if (pgid == 0) {
1818 		return;
1819 	}
1820 
1821 	KSI_INIT(&ksi);
1822 	ksi.ksi_signo = signo;
1823 	ksi.ksi_code = code;
1824 	ksi.ksi_band = band;
1825 
1826 	mutex_enter(proc_lock);
1827 	if (pgid > 0) {
1828 		struct proc *p1;
1829 
1830 		p1 = proc_find(pgid);
1831 		if (p1 != NULL) {
1832 			kpsignal(p1, &ksi, fdescdata);
1833 		}
1834 	} else {
1835 		struct pgrp *pgrp;
1836 
1837 		KASSERT(pgid < 0);
1838 		pgrp = pgrp_find(-pgid);
1839 		if (pgrp != NULL) {
1840 			kpgsignal(pgrp, &ksi, fdescdata, 0);
1841 		}
1842 	}
1843 	mutex_exit(proc_lock);
1844 }
1845 
1846 int
1847 fd_clone(file_t *fp, unsigned fd, int flag, const struct fileops *fops,
1848 	 void *data)
1849 {
1850 
1851 	fp->f_flag = flag;
1852 	fp->f_type = DTYPE_MISC;
1853 	fp->f_ops = fops;
1854 	fp->f_data = data;
1855 	curlwp->l_dupfd = fd;
1856 	fd_affix(curproc, fp, fd);
1857 
1858 	return EMOVEFD;
1859 }
1860 
1861 int
1862 fnullop_fcntl(file_t *fp, u_int cmd, void *data)
1863 {
1864 
1865 	if (cmd == F_SETFL)
1866 		return 0;
1867 
1868 	return EOPNOTSUPP;
1869 }
1870 
1871 int
1872 fnullop_poll(file_t *fp, int which)
1873 {
1874 
1875 	return 0;
1876 }
1877 
1878 int
1879 fnullop_kqfilter(file_t *fp, struct knote *kn)
1880 {
1881 
1882 	return 0;
1883 }
1884 
1885 void
1886 fnullop_restart(file_t *fp)
1887 {
1888 
1889 }
1890 
1891 int
1892 fbadop_read(file_t *fp, off_t *offset, struct uio *uio,
1893 	    kauth_cred_t cred, int flags)
1894 {
1895 
1896 	return EOPNOTSUPP;
1897 }
1898 
1899 int
1900 fbadop_write(file_t *fp, off_t *offset, struct uio *uio,
1901 	     kauth_cred_t cred, int flags)
1902 {
1903 
1904 	return EOPNOTSUPP;
1905 }
1906 
1907 int
1908 fbadop_ioctl(file_t *fp, u_long com, void *data)
1909 {
1910 
1911 	return EOPNOTSUPP;
1912 }
1913 
1914 int
1915 fbadop_stat(file_t *fp, struct stat *sb)
1916 {
1917 
1918 	return EOPNOTSUPP;
1919 }
1920 
1921 int
1922 fbadop_close(file_t *fp)
1923 {
1924 
1925 	return EOPNOTSUPP;
1926 }
1927 
1928 /*
1929  * sysctl routines pertaining to file descriptors
1930  */
1931 
1932 /* Initialized in sysctl_init() for now... */
1933 extern kmutex_t sysctl_file_marker_lock;
1934 static u_int sysctl_file_marker = 1;
1935 
1936 /*
1937  * Expects to be called with proc_lock and sysctl_file_marker_lock locked.
1938  */
1939 static void
1940 sysctl_file_marker_reset(void)
1941 {
1942 	struct proc *p;
1943 
1944 	PROCLIST_FOREACH(p, &allproc) {
1945 		struct filedesc *fd = p->p_fd;
1946 		fdtab_t *dt;
1947 		u_int i;
1948 
1949 		mutex_enter(&fd->fd_lock);
1950 		dt = fd->fd_dt;
1951 		for (i = 0; i < dt->dt_nfiles; i++) {
1952 			struct file *fp;
1953 			fdfile_t *ff;
1954 
1955 			if ((ff = dt->dt_ff[i]) == NULL) {
1956 				continue;
1957 			}
1958 			if ((fp = ff->ff_file) == NULL) {
1959 				continue;
1960 			}
1961 			fp->f_marker = 0;
1962 		}
1963 		mutex_exit(&fd->fd_lock);
1964 	}
1965 }
1966 
1967 /*
1968  * sysctl helper routine for kern.file pseudo-subtree.
1969  */
1970 static int
1971 sysctl_kern_file(SYSCTLFN_ARGS)
1972 {
1973 	int error;
1974 	size_t buflen;
1975 	struct file *fp, fbuf;
1976 	char *start, *where;
1977 	struct proc *p;
1978 
1979 	start = where = oldp;
1980 	buflen = *oldlenp;
1981 
1982 	if (where == NULL) {
1983 		/*
1984 		 * overestimate by 10 files
1985 		 */
1986 		*oldlenp = sizeof(filehead) + (nfiles + 10) *
1987 		    sizeof(struct file);
1988 		return 0;
1989 	}
1990 
1991 	/*
1992 	 * first sysctl_copyout filehead
1993 	 */
1994 	if (buflen < sizeof(filehead)) {
1995 		*oldlenp = 0;
1996 		return 0;
1997 	}
1998 	sysctl_unlock();
1999 	error = sysctl_copyout(l, &filehead, where, sizeof(filehead));
2000 	if (error) {
2001 		sysctl_relock();
2002 		return error;
2003 	}
2004 	buflen -= sizeof(filehead);
2005 	where += sizeof(filehead);
2006 
2007 	/*
2008 	 * followed by an array of file structures
2009 	 */
2010 	mutex_enter(&sysctl_file_marker_lock);
2011 	mutex_enter(proc_lock);
2012 	PROCLIST_FOREACH(p, &allproc) {
2013 		struct filedesc *fd;
2014 		fdtab_t *dt;
2015 		u_int i;
2016 
2017 		if (p->p_stat == SIDL) {
2018 			/* skip embryonic processes */
2019 			continue;
2020 		}
2021 		mutex_enter(p->p_lock);
2022 		error = kauth_authorize_process(l->l_cred,
2023 		    KAUTH_PROCESS_CANSEE, p,
2024 		    KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2025 		    NULL, NULL);
2026 		mutex_exit(p->p_lock);
2027 		if (error != 0) {
2028 			/*
2029 			 * Don't leak kauth retval if we're silently
2030 			 * skipping this entry.
2031 			 */
2032 			error = 0;
2033 			continue;
2034 		}
2035 
2036 		/*
2037 		 * Grab a hold on the process.
2038 		 */
2039 		if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2040 			continue;
2041 		}
2042 		mutex_exit(proc_lock);
2043 
2044 		fd = p->p_fd;
2045 		mutex_enter(&fd->fd_lock);
2046 		dt = fd->fd_dt;
2047 		for (i = 0; i < dt->dt_nfiles; i++) {
2048 			fdfile_t *ff;
2049 
2050 			if ((ff = dt->dt_ff[i]) == NULL) {
2051 				continue;
2052 			}
2053 			if ((fp = ff->ff_file) == NULL) {
2054 				continue;
2055 			}
2056 
2057 			mutex_enter(&fp->f_lock);
2058 
2059 			if ((fp->f_count == 0) ||
2060 			    (fp->f_marker == sysctl_file_marker)) {
2061 				mutex_exit(&fp->f_lock);
2062 				continue;
2063 			}
2064 
2065 			/* Check that we have enough space. */
2066 			if (buflen < sizeof(struct file)) {
2067 				*oldlenp = where - start;
2068 				mutex_exit(&fp->f_lock);
2069 				error = ENOMEM;
2070 				break;
2071 			}
2072 
2073 			memcpy(&fbuf, fp, sizeof(fbuf));
2074 			mutex_exit(&fp->f_lock);
2075 			error = sysctl_copyout(l, &fbuf, where, sizeof(fbuf));
2076 			if (error) {
2077 				break;
2078 			}
2079 			buflen -= sizeof(struct file);
2080 			where += sizeof(struct file);
2081 
2082 			fp->f_marker = sysctl_file_marker;
2083 		}
2084 		mutex_exit(&fd->fd_lock);
2085 
2086 		/*
2087 		 * Release reference to process.
2088 		 */
2089 		mutex_enter(proc_lock);
2090 		rw_exit(&p->p_reflock);
2091 
2092 		if (error)
2093 			break;
2094 	}
2095 
2096 	sysctl_file_marker++;
2097 	/* Reset all markers if wrapped. */
2098 	if (sysctl_file_marker == 0) {
2099 		sysctl_file_marker_reset();
2100 		sysctl_file_marker++;
2101 	}
2102 
2103 	mutex_exit(proc_lock);
2104 	mutex_exit(&sysctl_file_marker_lock);
2105 
2106 	*oldlenp = where - start;
2107 	sysctl_relock();
2108 	return error;
2109 }
2110 
2111 /*
2112  * sysctl helper function for kern.file2
2113  */
2114 static int
2115 sysctl_kern_file2(SYSCTLFN_ARGS)
2116 {
2117 	struct proc *p;
2118 	struct file *fp;
2119 	struct filedesc *fd;
2120 	struct kinfo_file kf;
2121 	char *dp;
2122 	u_int i, op;
2123 	size_t len, needed, elem_size, out_size;
2124 	int error, arg, elem_count;
2125 	fdfile_t *ff;
2126 	fdtab_t *dt;
2127 
2128 	if (namelen == 1 && name[0] == CTL_QUERY)
2129 		return sysctl_query(SYSCTLFN_CALL(rnode));
2130 
2131 	if (namelen != 4)
2132 		return EINVAL;
2133 
2134 	error = 0;
2135 	dp = oldp;
2136 	len = (oldp != NULL) ? *oldlenp : 0;
2137 	op = name[0];
2138 	arg = name[1];
2139 	elem_size = name[2];
2140 	elem_count = name[3];
2141 	out_size = MIN(sizeof(kf), elem_size);
2142 	needed = 0;
2143 
2144 	if (elem_size < 1 || elem_count < 0)
2145 		return EINVAL;
2146 
2147 	switch (op) {
2148 	case KERN_FILE_BYFILE:
2149 	case KERN_FILE_BYPID:
2150 		/*
2151 		 * We're traversing the process list in both cases; the BYFILE
2152 		 * case does additional work of keeping track of files already
2153 		 * looked at.
2154 		 */
2155 
2156 		/* doesn't use arg so it must be zero */
2157 		if ((op == KERN_FILE_BYFILE) && (arg != 0))
2158 			return EINVAL;
2159 
2160 		if ((op == KERN_FILE_BYPID) && (arg < -1))
2161 			/* -1 means all processes */
2162 			return EINVAL;
2163 
2164 		sysctl_unlock();
2165 		if (op == KERN_FILE_BYFILE)
2166 			mutex_enter(&sysctl_file_marker_lock);
2167 		mutex_enter(proc_lock);
2168 		PROCLIST_FOREACH(p, &allproc) {
2169 			if (p->p_stat == SIDL) {
2170 				/* skip embryonic processes */
2171 				continue;
2172 			}
2173 			if (arg > 0 && p->p_pid != arg) {
2174 				/* pick only the one we want */
2175 				/* XXX want 0 to mean "kernel files" */
2176 				continue;
2177 			}
2178 			mutex_enter(p->p_lock);
2179 			error = kauth_authorize_process(l->l_cred,
2180 			    KAUTH_PROCESS_CANSEE, p,
2181 			    KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2182 			    NULL, NULL);
2183 			mutex_exit(p->p_lock);
2184 			if (error != 0) {
2185 				/*
2186 				 * Don't leak kauth retval if we're silently
2187 				 * skipping this entry.
2188 				 */
2189 				error = 0;
2190 				continue;
2191 			}
2192 
2193 			/*
2194 			 * Grab a hold on the process.
2195 			 */
2196 			if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2197 				continue;
2198 			}
2199 			mutex_exit(proc_lock);
2200 
2201 			fd = p->p_fd;
2202 			mutex_enter(&fd->fd_lock);
2203 			dt = fd->fd_dt;
2204 			for (i = 0; i < dt->dt_nfiles; i++) {
2205 				if ((ff = dt->dt_ff[i]) == NULL) {
2206 					continue;
2207 				}
2208 				if ((fp = ff->ff_file) == NULL) {
2209 					continue;
2210 				}
2211 
2212 				if ((op == KERN_FILE_BYFILE) &&
2213 				    (fp->f_marker == sysctl_file_marker)) {
2214 					continue;
2215 				}
2216 				if (len >= elem_size && elem_count > 0) {
2217 					mutex_enter(&fp->f_lock);
2218 					fill_file(&kf, fp, ff, i, p->p_pid);
2219 					mutex_exit(&fp->f_lock);
2220 					mutex_exit(&fd->fd_lock);
2221 					error = sysctl_copyout(l,
2222 					    &kf, dp, out_size);
2223 					mutex_enter(&fd->fd_lock);
2224 					if (error)
2225 						break;
2226 					dp += elem_size;
2227 					len -= elem_size;
2228 				}
2229 				if (op == KERN_FILE_BYFILE)
2230 					fp->f_marker = sysctl_file_marker;
2231 				needed += elem_size;
2232 				if (elem_count > 0 && elem_count != INT_MAX)
2233 					elem_count--;
2234 			}
2235 			mutex_exit(&fd->fd_lock);
2236 
2237 			/*
2238 			 * Release reference to process.
2239 			 */
2240 			mutex_enter(proc_lock);
2241 			rw_exit(&p->p_reflock);
2242 		}
2243 		if (op == KERN_FILE_BYFILE) {
2244 			sysctl_file_marker++;
2245 
2246 			/* Reset all markers if wrapped. */
2247 			if (sysctl_file_marker == 0) {
2248 				sysctl_file_marker_reset();
2249 				sysctl_file_marker++;
2250 			}
2251 		}
2252 		mutex_exit(proc_lock);
2253 		if (op == KERN_FILE_BYFILE)
2254 			mutex_exit(&sysctl_file_marker_lock);
2255 		sysctl_relock();
2256 		break;
2257 	default:
2258 		return EINVAL;
2259 	}
2260 
2261 	if (oldp == NULL)
2262 		needed += KERN_FILESLOP * elem_size;
2263 	*oldlenp = needed;
2264 
2265 	return error;
2266 }
2267 
2268 static void
2269 fill_file(struct kinfo_file *kp, const file_t *fp, const fdfile_t *ff,
2270 	  int i, pid_t pid)
2271 {
2272 
2273 	memset(kp, 0, sizeof(*kp));
2274 
2275 	kp->ki_fileaddr =	PTRTOUINT64(fp);
2276 	kp->ki_flag =		fp->f_flag;
2277 	kp->ki_iflags =		0;
2278 	kp->ki_ftype =		fp->f_type;
2279 	kp->ki_count =		fp->f_count;
2280 	kp->ki_msgcount =	fp->f_msgcount;
2281 	kp->ki_fucred =		PTRTOUINT64(fp->f_cred);
2282 	kp->ki_fuid =		kauth_cred_geteuid(fp->f_cred);
2283 	kp->ki_fgid =		kauth_cred_getegid(fp->f_cred);
2284 	kp->ki_fops =		PTRTOUINT64(fp->f_ops);
2285 	kp->ki_foffset =	fp->f_offset;
2286 	kp->ki_fdata =		PTRTOUINT64(fp->f_data);
2287 
2288 	/* vnode information to glue this file to something */
2289 	if (fp->f_type == DTYPE_VNODE) {
2290 		struct vnode *vp = (struct vnode *)fp->f_data;
2291 
2292 		kp->ki_vun =	PTRTOUINT64(vp->v_un.vu_socket);
2293 		kp->ki_vsize =	vp->v_size;
2294 		kp->ki_vtype =	vp->v_type;
2295 		kp->ki_vtag =	vp->v_tag;
2296 		kp->ki_vdata =	PTRTOUINT64(vp->v_data);
2297 	}
2298 
2299 	/* process information when retrieved via KERN_FILE_BYPID */
2300 	if (ff != NULL) {
2301 		kp->ki_pid =		pid;
2302 		kp->ki_fd =		i;
2303 		kp->ki_ofileflags =	ff->ff_exclose;
2304 		kp->ki_usecount =	ff->ff_refcnt;
2305 	}
2306 }
2307