xref: /netbsd-src/sys/kern/kern_descrip.c (revision 7d62b00eb9ad855ffcd7da46b41e23feb5476fac)
1 /*	$NetBSD: kern_descrip.c,v 1.255 2023/02/24 11:02:27 riastradh 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 1.255 2023/02/24 11:02:27 riastradh Exp $");
74 
75 #include <sys/param.h>
76 #include <sys/systm.h>
77 #include <sys/filedesc.h>
78 #include <sys/kernel.h>
79 #include <sys/proc.h>
80 #include <sys/file.h>
81 #include <sys/socket.h>
82 #include <sys/socketvar.h>
83 #include <sys/stat.h>
84 #include <sys/ioctl.h>
85 #include <sys/fcntl.h>
86 #include <sys/pool.h>
87 #include <sys/unistd.h>
88 #include <sys/resourcevar.h>
89 #include <sys/conf.h>
90 #include <sys/event.h>
91 #include <sys/kauth.h>
92 #include <sys/atomic.h>
93 #include <sys/syscallargs.h>
94 #include <sys/cpu.h>
95 #include <sys/kmem.h>
96 #include <sys/vnode.h>
97 #include <sys/sysctl.h>
98 #include <sys/ktrace.h>
99 
100 /*
101  * A list (head) of open files, counter, and lock protecting them.
102  */
103 struct filelist		filehead	__cacheline_aligned;
104 static u_int		nfiles		__cacheline_aligned;
105 kmutex_t		filelist_lock	__cacheline_aligned;
106 
107 static pool_cache_t	filedesc_cache	__read_mostly;
108 static pool_cache_t	file_cache	__read_mostly;
109 static pool_cache_t	fdfile_cache	__read_mostly;
110 
111 static int	file_ctor(void *, void *, int);
112 static void	file_dtor(void *, void *);
113 static int	fdfile_ctor(void *, void *, int);
114 static void	fdfile_dtor(void *, void *);
115 static int	filedesc_ctor(void *, void *, int);
116 static void	filedesc_dtor(void *, void *);
117 static int	filedescopen(dev_t, int, int, lwp_t *);
118 
119 static int sysctl_kern_file(SYSCTLFN_PROTO);
120 static int sysctl_kern_file2(SYSCTLFN_PROTO);
121 static void fill_file(struct file *, const struct file *);
122 static void fill_file2(struct kinfo_file *, const file_t *, const fdfile_t *,
123 		      int, pid_t);
124 
125 const struct cdevsw filedesc_cdevsw = {
126 	.d_open = filedescopen,
127 	.d_close = noclose,
128 	.d_read = noread,
129 	.d_write = nowrite,
130 	.d_ioctl = noioctl,
131 	.d_stop = nostop,
132 	.d_tty = notty,
133 	.d_poll = nopoll,
134 	.d_mmap = nommap,
135 	.d_kqfilter = nokqfilter,
136 	.d_discard = nodiscard,
137 	.d_flag = D_OTHER | D_MPSAFE
138 };
139 
140 /* For ease of reading. */
141 __strong_alias(fd_putvnode,fd_putfile)
142 __strong_alias(fd_putsock,fd_putfile)
143 
144 /*
145  * Initialize the descriptor system.
146  */
147 void
148 fd_sys_init(void)
149 {
150 	static struct sysctllog *clog;
151 
152 	mutex_init(&filelist_lock, MUTEX_DEFAULT, IPL_NONE);
153 
154 	LIST_INIT(&filehead);
155 
156 	file_cache = pool_cache_init(sizeof(file_t), coherency_unit, 0,
157 	    0, "file", NULL, IPL_NONE, file_ctor, file_dtor, NULL);
158 	KASSERT(file_cache != NULL);
159 
160 	fdfile_cache = pool_cache_init(sizeof(fdfile_t), coherency_unit, 0,
161 	    PR_LARGECACHE, "fdfile", NULL, IPL_NONE, fdfile_ctor, fdfile_dtor,
162 	    NULL);
163 	KASSERT(fdfile_cache != NULL);
164 
165 	filedesc_cache = pool_cache_init(sizeof(filedesc_t), coherency_unit,
166 	    0, 0, "filedesc", NULL, IPL_NONE, filedesc_ctor, filedesc_dtor,
167 	    NULL);
168 	KASSERT(filedesc_cache != NULL);
169 
170 	sysctl_createv(&clog, 0, NULL, NULL,
171 		       CTLFLAG_PERMANENT,
172 		       CTLTYPE_STRUCT, "file",
173 		       SYSCTL_DESCR("System open file table"),
174 		       sysctl_kern_file, 0, NULL, 0,
175 		       CTL_KERN, KERN_FILE, CTL_EOL);
176 	sysctl_createv(&clog, 0, NULL, NULL,
177 		       CTLFLAG_PERMANENT,
178 		       CTLTYPE_STRUCT, "file2",
179 		       SYSCTL_DESCR("System open file table"),
180 		       sysctl_kern_file2, 0, NULL, 0,
181 		       CTL_KERN, KERN_FILE2, CTL_EOL);
182 }
183 
184 static bool
185 fd_isused(filedesc_t *fdp, unsigned fd)
186 {
187 	u_int off = fd >> NDENTRYSHIFT;
188 
189 	KASSERT(fd < atomic_load_consume(&fdp->fd_dt)->dt_nfiles);
190 
191 	return (fdp->fd_lomap[off] & (1U << (fd & NDENTRYMASK))) != 0;
192 }
193 
194 /*
195  * Verify that the bitmaps match the descriptor table.
196  */
197 static inline void
198 fd_checkmaps(filedesc_t *fdp)
199 {
200 #ifdef DEBUG
201 	fdtab_t *dt;
202 	u_int fd;
203 
204 	KASSERT(fdp->fd_refcnt <= 1 || mutex_owned(&fdp->fd_lock));
205 
206 	dt = fdp->fd_dt;
207 	if (fdp->fd_refcnt == -1) {
208 		/*
209 		 * fd_free tears down the table without maintaining its bitmap.
210 		 */
211 		return;
212 	}
213 	for (fd = 0; fd < dt->dt_nfiles; fd++) {
214 		if (fd < NDFDFILE) {
215 			KASSERT(dt->dt_ff[fd] ==
216 			    (fdfile_t *)fdp->fd_dfdfile[fd]);
217 		}
218 		if (dt->dt_ff[fd] == NULL) {
219 			KASSERT(!fd_isused(fdp, fd));
220 		} else if (dt->dt_ff[fd]->ff_file != NULL) {
221 			KASSERT(fd_isused(fdp, fd));
222 		}
223 	}
224 #endif
225 }
226 
227 static int
228 fd_next_zero(filedesc_t *fdp, uint32_t *bitmap, int want, u_int bits)
229 {
230 	int i, off, maxoff;
231 	uint32_t sub;
232 
233 	KASSERT(mutex_owned(&fdp->fd_lock));
234 
235 	fd_checkmaps(fdp);
236 
237 	if (want > bits)
238 		return -1;
239 
240 	off = want >> NDENTRYSHIFT;
241 	i = want & NDENTRYMASK;
242 	if (i) {
243 		sub = bitmap[off] | ((u_int)~0 >> (NDENTRIES - i));
244 		if (sub != ~0)
245 			goto found;
246 		off++;
247 	}
248 
249 	maxoff = NDLOSLOTS(bits);
250 	while (off < maxoff) {
251 		if ((sub = bitmap[off]) != ~0)
252 			goto found;
253 		off++;
254 	}
255 
256 	return -1;
257 
258  found:
259 	return (off << NDENTRYSHIFT) + ffs(~sub) - 1;
260 }
261 
262 static int
263 fd_last_set(filedesc_t *fd, int last)
264 {
265 	int off, i;
266 	fdfile_t **ff = fd->fd_dt->dt_ff;
267 	uint32_t *bitmap = fd->fd_lomap;
268 
269 	KASSERT(mutex_owned(&fd->fd_lock));
270 
271 	fd_checkmaps(fd);
272 
273 	off = (last - 1) >> NDENTRYSHIFT;
274 
275 	while (off >= 0 && !bitmap[off])
276 		off--;
277 
278 	if (off < 0)
279 		return -1;
280 
281 	i = ((off + 1) << NDENTRYSHIFT) - 1;
282 	if (i >= last)
283 		i = last - 1;
284 
285 	/* XXX should use bitmap */
286 	while (i > 0 && (ff[i] == NULL || !ff[i]->ff_allocated))
287 		i--;
288 
289 	return i;
290 }
291 
292 static inline void
293 fd_used(filedesc_t *fdp, unsigned fd)
294 {
295 	u_int off = fd >> NDENTRYSHIFT;
296 	fdfile_t *ff;
297 
298 	ff = fdp->fd_dt->dt_ff[fd];
299 
300 	KASSERT(mutex_owned(&fdp->fd_lock));
301 	KASSERT((fdp->fd_lomap[off] & (1U << (fd & NDENTRYMASK))) == 0);
302 	KASSERT(ff != NULL);
303 	KASSERT(ff->ff_file == NULL);
304 	KASSERT(!ff->ff_allocated);
305 
306 	ff->ff_allocated = true;
307 	fdp->fd_lomap[off] |= 1U << (fd & NDENTRYMASK);
308 	if (__predict_false(fdp->fd_lomap[off] == ~0)) {
309 		KASSERT((fdp->fd_himap[off >> NDENTRYSHIFT] &
310 		    (1U << (off & NDENTRYMASK))) == 0);
311 		fdp->fd_himap[off >> NDENTRYSHIFT] |= 1U << (off & NDENTRYMASK);
312 	}
313 
314 	if ((int)fd > fdp->fd_lastfile) {
315 		fdp->fd_lastfile = fd;
316 	}
317 
318 	fd_checkmaps(fdp);
319 }
320 
321 static inline void
322 fd_unused(filedesc_t *fdp, unsigned fd)
323 {
324 	u_int off = fd >> NDENTRYSHIFT;
325 	fdfile_t *ff;
326 
327 	ff = fdp->fd_dt->dt_ff[fd];
328 
329 	KASSERT(mutex_owned(&fdp->fd_lock));
330 	KASSERT(ff != NULL);
331 	KASSERT(ff->ff_file == NULL);
332 	KASSERT(ff->ff_allocated);
333 
334 	if (fd < fdp->fd_freefile) {
335 		fdp->fd_freefile = fd;
336 	}
337 
338 	if (fdp->fd_lomap[off] == ~0) {
339 		KASSERT((fdp->fd_himap[off >> NDENTRYSHIFT] &
340 		    (1U << (off & NDENTRYMASK))) != 0);
341 		fdp->fd_himap[off >> NDENTRYSHIFT] &=
342 		    ~(1U << (off & NDENTRYMASK));
343 	}
344 	KASSERT((fdp->fd_lomap[off] & (1U << (fd & NDENTRYMASK))) != 0);
345 	fdp->fd_lomap[off] &= ~(1U << (fd & NDENTRYMASK));
346 	ff->ff_allocated = false;
347 
348 	KASSERT(fd <= fdp->fd_lastfile);
349 	if (fd == fdp->fd_lastfile) {
350 		fdp->fd_lastfile = fd_last_set(fdp, fd);
351 	}
352 	fd_checkmaps(fdp);
353 }
354 
355 /*
356  * Look up the file structure corresponding to a file descriptor
357  * and return the file, holding a reference on the descriptor.
358  */
359 file_t *
360 fd_getfile(unsigned fd)
361 {
362 	filedesc_t *fdp;
363 	fdfile_t *ff;
364 	file_t *fp;
365 	fdtab_t *dt;
366 
367 	/*
368 	 * Look up the fdfile structure representing this descriptor.
369 	 * We are doing this unlocked.  See fd_tryexpand().
370 	 */
371 	fdp = curlwp->l_fd;
372 	dt = atomic_load_consume(&fdp->fd_dt);
373 	if (__predict_false(fd >= dt->dt_nfiles)) {
374 		return NULL;
375 	}
376 	ff = dt->dt_ff[fd];
377 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
378 	if (__predict_false(ff == NULL)) {
379 		return NULL;
380 	}
381 
382 	/* Now get a reference to the descriptor. */
383 	if (fdp->fd_refcnt == 1) {
384 		/*
385 		 * Single threaded: don't need to worry about concurrent
386 		 * access (other than earlier calls to kqueue, which may
387 		 * hold a reference to the descriptor).
388 		 */
389 		ff->ff_refcnt++;
390 	} else {
391 		/*
392 		 * Multi threaded: issue a memory barrier to ensure that we
393 		 * acquire the file pointer _after_ adding a reference.  If
394 		 * no memory barrier, we could fetch a stale pointer.
395 		 *
396 		 * In particular, we must coordinate the following four
397 		 * memory operations:
398 		 *
399 		 *	A. fd_close store ff->ff_file = NULL
400 		 *	B. fd_close refcnt = atomic_dec_uint_nv(&ff->ff_refcnt)
401 		 *	C. fd_getfile atomic_inc_uint(&ff->ff_refcnt)
402 		 *	D. fd_getfile load fp = ff->ff_file
403 		 *
404 		 * If the order is D;A;B;C:
405 		 *
406 		 *	1. D: fp = ff->ff_file
407 		 *	2. A: ff->ff_file = NULL
408 		 *	3. B: refcnt = atomic_dec_uint_nv(&ff->ff_refcnt)
409 		 *	4. C: atomic_inc_uint(&ff->ff_refcnt)
410 		 *
411 		 * then fd_close determines that there are no more
412 		 * references and decides to free fp immediately, at
413 		 * the same that fd_getfile ends up with an fp that's
414 		 * about to be freed.  *boom*
415 		 *
416 		 * By making B a release operation in fd_close, and by
417 		 * making C an acquire operation in fd_getfile, since
418 		 * they are atomic operations on the same object, which
419 		 * has a total modification order, we guarantee either:
420 		 *
421 		 *	- B happens before C.  Then since A is
422 		 *	  sequenced before B in fd_close, and C is
423 		 *	  sequenced before D in fd_getfile, we
424 		 *	  guarantee A happens before D, so fd_getfile
425 		 *	  reads a null fp and safely fails.
426 		 *
427 		 *	- C happens before B.  Then fd_getfile may read
428 		 *	  null or nonnull, but either way, fd_close
429 		 *	  will safely wait for references to drain.
430 		 */
431 		atomic_inc_uint(&ff->ff_refcnt);
432 		membar_acquire();
433 	}
434 
435 	/*
436 	 * If the file is not open or is being closed then put the
437 	 * reference back.
438 	 */
439 	fp = atomic_load_consume(&ff->ff_file);
440 	if (__predict_true(fp != NULL)) {
441 		return fp;
442 	}
443 	fd_putfile(fd);
444 	return NULL;
445 }
446 
447 /*
448  * Release a reference to a file descriptor acquired with fd_getfile().
449  */
450 void
451 fd_putfile(unsigned fd)
452 {
453 	filedesc_t *fdp;
454 	fdfile_t *ff;
455 	u_int u, v;
456 
457 	fdp = curlwp->l_fd;
458 	KASSERT(fd < atomic_load_consume(&fdp->fd_dt)->dt_nfiles);
459 	ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
460 
461 	KASSERT(ff != NULL);
462 	KASSERT((ff->ff_refcnt & FR_MASK) > 0);
463 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
464 
465 	if (fdp->fd_refcnt == 1) {
466 		/*
467 		 * Single threaded: don't need to worry about concurrent
468 		 * access (other than earlier calls to kqueue, which may
469 		 * hold a reference to the descriptor).
470 		 */
471 		if (__predict_false((ff->ff_refcnt & FR_CLOSING) != 0)) {
472 			fd_close(fd);
473 			return;
474 		}
475 		ff->ff_refcnt--;
476 		return;
477 	}
478 
479 	/*
480 	 * Ensure that any use of the file is complete and globally
481 	 * visible before dropping the final reference.  If no membar,
482 	 * the current CPU could still access memory associated with
483 	 * the file after it has been freed or recycled by another
484 	 * CPU.
485 	 */
486 	membar_release();
487 
488 	/*
489 	 * Be optimistic and start out with the assumption that no other
490 	 * threads are trying to close the descriptor.  If the CAS fails,
491 	 * we lost a race and/or it's being closed.
492 	 */
493 	for (u = ff->ff_refcnt & FR_MASK;; u = v) {
494 		v = atomic_cas_uint(&ff->ff_refcnt, u, u - 1);
495 		if (__predict_true(u == v)) {
496 			return;
497 		}
498 		if (__predict_false((v & FR_CLOSING) != 0)) {
499 			break;
500 		}
501 	}
502 
503 	/* Another thread is waiting to close the file: join it. */
504 	(void)fd_close(fd);
505 }
506 
507 /*
508  * Convenience wrapper around fd_getfile() that returns reference
509  * to a vnode.
510  */
511 int
512 fd_getvnode(unsigned fd, file_t **fpp)
513 {
514 	vnode_t *vp;
515 	file_t *fp;
516 
517 	fp = fd_getfile(fd);
518 	if (__predict_false(fp == NULL)) {
519 		return EBADF;
520 	}
521 	if (__predict_false(fp->f_type != DTYPE_VNODE)) {
522 		fd_putfile(fd);
523 		return EINVAL;
524 	}
525 	vp = fp->f_vnode;
526 	if (__predict_false(vp->v_type == VBAD)) {
527 		/* XXX Is this case really necessary? */
528 		fd_putfile(fd);
529 		return EBADF;
530 	}
531 	*fpp = fp;
532 	return 0;
533 }
534 
535 /*
536  * Convenience wrapper around fd_getfile() that returns reference
537  * to a socket.
538  */
539 int
540 fd_getsock1(unsigned fd, struct socket **sop, file_t **fp)
541 {
542 	*fp = fd_getfile(fd);
543 	if (__predict_false(*fp == NULL)) {
544 		return EBADF;
545 	}
546 	if (__predict_false((*fp)->f_type != DTYPE_SOCKET)) {
547 		fd_putfile(fd);
548 		return ENOTSOCK;
549 	}
550 	*sop = (*fp)->f_socket;
551 	return 0;
552 }
553 
554 int
555 fd_getsock(unsigned fd, struct socket **sop)
556 {
557 	file_t *fp;
558 	return fd_getsock1(fd, sop, &fp);
559 }
560 
561 /*
562  * Look up the file structure corresponding to a file descriptor
563  * and return it with a reference held on the file, not the
564  * descriptor.
565  *
566  * This is heavyweight and only used when accessing descriptors
567  * from a foreign process.  The caller must ensure that `p' does
568  * not exit or fork across this call.
569  *
570  * To release the file (not descriptor) reference, use closef().
571  */
572 file_t *
573 fd_getfile2(proc_t *p, unsigned fd)
574 {
575 	filedesc_t *fdp;
576 	fdfile_t *ff;
577 	file_t *fp;
578 	fdtab_t *dt;
579 
580 	fdp = p->p_fd;
581 	mutex_enter(&fdp->fd_lock);
582 	dt = fdp->fd_dt;
583 	if (fd >= dt->dt_nfiles) {
584 		mutex_exit(&fdp->fd_lock);
585 		return NULL;
586 	}
587 	if ((ff = dt->dt_ff[fd]) == NULL) {
588 		mutex_exit(&fdp->fd_lock);
589 		return NULL;
590 	}
591 	if ((fp = atomic_load_consume(&ff->ff_file)) == NULL) {
592 		mutex_exit(&fdp->fd_lock);
593 		return NULL;
594 	}
595 	mutex_enter(&fp->f_lock);
596 	fp->f_count++;
597 	mutex_exit(&fp->f_lock);
598 	mutex_exit(&fdp->fd_lock);
599 
600 	return fp;
601 }
602 
603 /*
604  * Internal form of close.  Must be called with a reference to the
605  * descriptor, and will drop the reference.  When all descriptor
606  * references are dropped, releases the descriptor slot and a single
607  * reference to the file structure.
608  */
609 int
610 fd_close(unsigned fd)
611 {
612 	struct flock lf;
613 	filedesc_t *fdp;
614 	fdfile_t *ff;
615 	file_t *fp;
616 	proc_t *p;
617 	lwp_t *l;
618 	u_int refcnt;
619 
620 	l = curlwp;
621 	p = l->l_proc;
622 	fdp = l->l_fd;
623 	ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
624 
625 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
626 
627 	mutex_enter(&fdp->fd_lock);
628 	KASSERT((ff->ff_refcnt & FR_MASK) > 0);
629 	fp = atomic_load_consume(&ff->ff_file);
630 	if (__predict_false(fp == NULL)) {
631 		/*
632 		 * Another user of the file is already closing, and is
633 		 * waiting for other users of the file to drain.  Release
634 		 * our reference, and wake up the closer.
635 		 */
636 		membar_release();
637 		atomic_dec_uint(&ff->ff_refcnt);
638 		cv_broadcast(&ff->ff_closing);
639 		mutex_exit(&fdp->fd_lock);
640 
641 		/*
642 		 * An application error, so pretend that the descriptor
643 		 * was already closed.  We can't safely wait for it to
644 		 * be closed without potentially deadlocking.
645 		 */
646 		return (EBADF);
647 	}
648 	KASSERT((ff->ff_refcnt & FR_CLOSING) == 0);
649 
650 	/*
651 	 * There may be multiple users of this file within the process.
652 	 * Notify existing and new users that the file is closing.  This
653 	 * will prevent them from adding additional uses to this file
654 	 * while we are closing it.
655 	 */
656 	atomic_store_relaxed(&ff->ff_file, NULL);
657 	ff->ff_exclose = false;
658 
659 	/*
660 	 * We expect the caller to hold a descriptor reference - drop it.
661 	 * The reference count may increase beyond zero at this point due
662 	 * to an erroneous descriptor reference by an application, but
663 	 * fd_getfile() will notice that the file is being closed and drop
664 	 * the reference again.
665 	 */
666 	if (fdp->fd_refcnt == 1) {
667 		/* Single threaded. */
668 		refcnt = --(ff->ff_refcnt);
669 	} else {
670 		/* Multi threaded. */
671 		membar_release();
672 		refcnt = atomic_dec_uint_nv(&ff->ff_refcnt);
673 		membar_acquire();
674 	}
675 	if (__predict_false(refcnt != 0)) {
676 		/*
677 		 * Wait for other references to drain.  This is typically
678 		 * an application error - the descriptor is being closed
679 		 * while still in use.
680 		 * (Or just a threaded application trying to unblock its
681 		 * thread that sleeps in (say) accept()).
682 		 */
683 		atomic_or_uint(&ff->ff_refcnt, FR_CLOSING);
684 
685 		/*
686 		 * Remove any knotes attached to the file.  A knote
687 		 * attached to the descriptor can hold references on it.
688 		 */
689 		mutex_exit(&fdp->fd_lock);
690 		if (!SLIST_EMPTY(&ff->ff_knlist)) {
691 			knote_fdclose(fd);
692 		}
693 
694 		/*
695 		 * Since the file system code doesn't know which fd
696 		 * each request came from (think dup()), we have to
697 		 * ask it to return ERESTART for any long-term blocks.
698 		 * The re-entry through read/write/etc will detect the
699 		 * closed fd and return EBAFD.
700 		 * Blocked partial writes may return a short length.
701 		 */
702 		(*fp->f_ops->fo_restart)(fp);
703 		mutex_enter(&fdp->fd_lock);
704 
705 		/*
706 		 * We need to see the count drop to zero at least once,
707 		 * in order to ensure that all pre-existing references
708 		 * have been drained.  New references past this point are
709 		 * of no interest.
710 		 * XXX (dsl) this may need to call fo_restart() after a
711 		 * timeout to guarantee that all the system calls exit.
712 		 */
713 		while ((ff->ff_refcnt & FR_MASK) != 0) {
714 			cv_wait(&ff->ff_closing, &fdp->fd_lock);
715 		}
716 		atomic_and_uint(&ff->ff_refcnt, ~FR_CLOSING);
717 	} else {
718 		/* If no references, there must be no knotes. */
719 		KASSERT(SLIST_EMPTY(&ff->ff_knlist));
720 	}
721 
722 	/*
723 	 * POSIX record locking dictates that any close releases ALL
724 	 * locks owned by this process.  This is handled by setting
725 	 * a flag in the unlock to free ONLY locks obeying POSIX
726 	 * semantics, and not to free BSD-style file locks.
727 	 * If the descriptor was in a message, POSIX-style locks
728 	 * aren't passed with the descriptor.
729 	 */
730 	if (__predict_false((p->p_flag & PK_ADVLOCK) != 0 &&
731 	    fp->f_type == DTYPE_VNODE)) {
732 		lf.l_whence = SEEK_SET;
733 		lf.l_start = 0;
734 		lf.l_len = 0;
735 		lf.l_type = F_UNLCK;
736 		mutex_exit(&fdp->fd_lock);
737 		(void)VOP_ADVLOCK(fp->f_vnode, p, F_UNLCK, &lf, F_POSIX);
738 		mutex_enter(&fdp->fd_lock);
739 	}
740 
741 	/* Free descriptor slot. */
742 	fd_unused(fdp, fd);
743 	mutex_exit(&fdp->fd_lock);
744 
745 	/* Now drop reference to the file itself. */
746 	return closef(fp);
747 }
748 
749 /*
750  * Duplicate a file descriptor.
751  */
752 int
753 fd_dup(file_t *fp, int minfd, int *newp, bool exclose)
754 {
755 	proc_t *p = curproc;
756 	fdtab_t *dt;
757 	int error;
758 
759 	while ((error = fd_alloc(p, minfd, newp)) != 0) {
760 		if (error != ENOSPC) {
761 			return error;
762 		}
763 		fd_tryexpand(p);
764 	}
765 
766 	dt = atomic_load_consume(&curlwp->l_fd->fd_dt);
767 	dt->dt_ff[*newp]->ff_exclose = exclose;
768 	fd_affix(p, fp, *newp);
769 	return 0;
770 }
771 
772 /*
773  * dup2 operation.
774  */
775 int
776 fd_dup2(file_t *fp, unsigned newfd, int flags)
777 {
778 	filedesc_t *fdp = curlwp->l_fd;
779 	fdfile_t *ff;
780 	fdtab_t *dt;
781 
782 	if (flags & ~(O_CLOEXEC|O_NONBLOCK|O_NOSIGPIPE))
783 		return EINVAL;
784 	/*
785 	 * Ensure there are enough slots in the descriptor table,
786 	 * and allocate an fdfile_t up front in case we need it.
787 	 */
788 	while (newfd >= atomic_load_consume(&fdp->fd_dt)->dt_nfiles) {
789 		fd_tryexpand(curproc);
790 	}
791 	ff = pool_cache_get(fdfile_cache, PR_WAITOK);
792 
793 	/*
794 	 * If there is already a file open, close it.  If the file is
795 	 * half open, wait for it to be constructed before closing it.
796 	 * XXX Potential for deadlock here?
797 	 */
798 	mutex_enter(&fdp->fd_lock);
799 	while (fd_isused(fdp, newfd)) {
800 		mutex_exit(&fdp->fd_lock);
801 		if (fd_getfile(newfd) != NULL) {
802 			(void)fd_close(newfd);
803 		} else {
804 			/*
805 			 * Crummy, but unlikely to happen.
806 			 * Can occur if we interrupt another
807 			 * thread while it is opening a file.
808 			 */
809 			kpause("dup2", false, 1, NULL);
810 		}
811 		mutex_enter(&fdp->fd_lock);
812 	}
813 	dt = fdp->fd_dt;
814 	if (dt->dt_ff[newfd] == NULL) {
815 		KASSERT(newfd >= NDFDFILE);
816 		dt->dt_ff[newfd] = ff;
817 		ff = NULL;
818 	}
819 	fd_used(fdp, newfd);
820 	mutex_exit(&fdp->fd_lock);
821 
822 	dt->dt_ff[newfd]->ff_exclose = (flags & O_CLOEXEC) != 0;
823 	fp->f_flag |= flags & (FNONBLOCK|FNOSIGPIPE);
824 	/* Slot is now allocated.  Insert copy of the file. */
825 	fd_affix(curproc, fp, newfd);
826 	if (ff != NULL) {
827 		pool_cache_put(fdfile_cache, ff);
828 	}
829 	return 0;
830 }
831 
832 /*
833  * Drop reference to a file structure.
834  */
835 int
836 closef(file_t *fp)
837 {
838 	struct flock lf;
839 	int error;
840 
841 	/*
842 	 * Drop reference.  If referenced elsewhere it's still open
843 	 * and we have nothing more to do.
844 	 */
845 	mutex_enter(&fp->f_lock);
846 	KASSERT(fp->f_count > 0);
847 	if (--fp->f_count > 0) {
848 		mutex_exit(&fp->f_lock);
849 		return 0;
850 	}
851 	KASSERT(fp->f_count == 0);
852 	mutex_exit(&fp->f_lock);
853 
854 	/* We held the last reference - release locks, close and free. */
855 	if ((fp->f_flag & FHASLOCK) && fp->f_type == DTYPE_VNODE) {
856 		lf.l_whence = SEEK_SET;
857 		lf.l_start = 0;
858 		lf.l_len = 0;
859 		lf.l_type = F_UNLCK;
860 		(void)VOP_ADVLOCK(fp->f_vnode, fp, F_UNLCK, &lf, F_FLOCK);
861 	}
862 	if (fp->f_ops != NULL) {
863 		error = (*fp->f_ops->fo_close)(fp);
864 	} else {
865 		error = 0;
866 	}
867 	KASSERT(fp->f_count == 0);
868 	KASSERT(fp->f_cred != NULL);
869 	pool_cache_put(file_cache, fp);
870 
871 	return error;
872 }
873 
874 /*
875  * Allocate a file descriptor for the process.
876  */
877 int
878 fd_alloc(proc_t *p, int want, int *result)
879 {
880 	filedesc_t *fdp = p->p_fd;
881 	int i, lim, last, error, hi;
882 	u_int off;
883 	fdtab_t *dt;
884 
885 	KASSERT(p == curproc || p == &proc0);
886 
887 	/*
888 	 * Search for a free descriptor starting at the higher
889 	 * of want or fd_freefile.
890 	 */
891 	mutex_enter(&fdp->fd_lock);
892 	fd_checkmaps(fdp);
893 	dt = fdp->fd_dt;
894 	KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
895 	lim = uimin((int)p->p_rlimit[RLIMIT_NOFILE].rlim_cur, maxfiles);
896 	last = uimin(dt->dt_nfiles, lim);
897 	for (;;) {
898 		if ((i = want) < fdp->fd_freefile)
899 			i = fdp->fd_freefile;
900 		off = i >> NDENTRYSHIFT;
901 		hi = fd_next_zero(fdp, fdp->fd_himap, off,
902 		    (last + NDENTRIES - 1) >> NDENTRYSHIFT);
903 		if (hi == -1)
904 			break;
905 		i = fd_next_zero(fdp, &fdp->fd_lomap[hi],
906 		    hi > off ? 0 : i & NDENTRYMASK, NDENTRIES);
907 		if (i == -1) {
908 			/*
909 			 * Free file descriptor in this block was
910 			 * below want, try again with higher want.
911 			 */
912 			want = (hi + 1) << NDENTRYSHIFT;
913 			continue;
914 		}
915 		i += (hi << NDENTRYSHIFT);
916 		if (i >= last) {
917 			break;
918 		}
919 		if (dt->dt_ff[i] == NULL) {
920 			KASSERT(i >= NDFDFILE);
921 			dt->dt_ff[i] = pool_cache_get(fdfile_cache, PR_WAITOK);
922 		}
923 		KASSERT(dt->dt_ff[i]->ff_file == NULL);
924 		fd_used(fdp, i);
925 		if (want <= fdp->fd_freefile) {
926 			fdp->fd_freefile = i;
927 		}
928 		*result = i;
929 		KASSERT(i >= NDFDFILE ||
930 		    dt->dt_ff[i] == (fdfile_t *)fdp->fd_dfdfile[i]);
931 		fd_checkmaps(fdp);
932 		mutex_exit(&fdp->fd_lock);
933 		return 0;
934 	}
935 
936 	/* No space in current array.  Let the caller expand and retry. */
937 	error = (dt->dt_nfiles >= lim) ? EMFILE : ENOSPC;
938 	mutex_exit(&fdp->fd_lock);
939 	return error;
940 }
941 
942 /*
943  * Allocate memory for a descriptor table.
944  */
945 static fdtab_t *
946 fd_dtab_alloc(int n)
947 {
948 	fdtab_t *dt;
949 	size_t sz;
950 
951 	KASSERT(n > NDFILE);
952 
953 	sz = sizeof(*dt) + (n - NDFILE) * sizeof(dt->dt_ff[0]);
954 	dt = kmem_alloc(sz, KM_SLEEP);
955 #ifdef DIAGNOSTIC
956 	memset(dt, 0xff, sz);
957 #endif
958 	dt->dt_nfiles = n;
959 	dt->dt_link = NULL;
960 	return dt;
961 }
962 
963 /*
964  * Free a descriptor table, and all tables linked for deferred free.
965  */
966 static void
967 fd_dtab_free(fdtab_t *dt)
968 {
969 	fdtab_t *next;
970 	size_t sz;
971 
972 	do {
973 		next = dt->dt_link;
974 		KASSERT(dt->dt_nfiles > NDFILE);
975 		sz = sizeof(*dt) +
976 		    (dt->dt_nfiles - NDFILE) * sizeof(dt->dt_ff[0]);
977 #ifdef DIAGNOSTIC
978 		memset(dt, 0xff, sz);
979 #endif
980 		kmem_free(dt, sz);
981 		dt = next;
982 	} while (dt != NULL);
983 }
984 
985 /*
986  * Allocate descriptor bitmap.
987  */
988 static void
989 fd_map_alloc(int n, uint32_t **lo, uint32_t **hi)
990 {
991 	uint8_t *ptr;
992 	size_t szlo, szhi;
993 
994 	KASSERT(n > NDENTRIES);
995 
996 	szlo = NDLOSLOTS(n) * sizeof(uint32_t);
997 	szhi = NDHISLOTS(n) * sizeof(uint32_t);
998 	ptr = kmem_alloc(szlo + szhi, KM_SLEEP);
999 	*lo = (uint32_t *)ptr;
1000 	*hi = (uint32_t *)(ptr + szlo);
1001 }
1002 
1003 /*
1004  * Free descriptor bitmap.
1005  */
1006 static void
1007 fd_map_free(int n, uint32_t *lo, uint32_t *hi)
1008 {
1009 	size_t szlo, szhi;
1010 
1011 	KASSERT(n > NDENTRIES);
1012 
1013 	szlo = NDLOSLOTS(n) * sizeof(uint32_t);
1014 	szhi = NDHISLOTS(n) * sizeof(uint32_t);
1015 	KASSERT(hi == (uint32_t *)((uint8_t *)lo + szlo));
1016 	kmem_free(lo, szlo + szhi);
1017 }
1018 
1019 /*
1020  * Expand a process' descriptor table.
1021  */
1022 void
1023 fd_tryexpand(proc_t *p)
1024 {
1025 	filedesc_t *fdp;
1026 	int i, numfiles, oldnfiles;
1027 	fdtab_t *newdt, *dt;
1028 	uint32_t *newhimap, *newlomap;
1029 
1030 	KASSERT(p == curproc || p == &proc0);
1031 
1032 	fdp = p->p_fd;
1033 	newhimap = NULL;
1034 	newlomap = NULL;
1035 	oldnfiles = atomic_load_consume(&fdp->fd_dt)->dt_nfiles;
1036 
1037 	if (oldnfiles < NDEXTENT)
1038 		numfiles = NDEXTENT;
1039 	else
1040 		numfiles = 2 * oldnfiles;
1041 
1042 	newdt = fd_dtab_alloc(numfiles);
1043 	if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1044 		fd_map_alloc(numfiles, &newlomap, &newhimap);
1045 	}
1046 
1047 	mutex_enter(&fdp->fd_lock);
1048 	dt = fdp->fd_dt;
1049 	KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1050 	if (dt->dt_nfiles != oldnfiles) {
1051 		/* fdp changed; caller must retry */
1052 		mutex_exit(&fdp->fd_lock);
1053 		fd_dtab_free(newdt);
1054 		if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1055 			fd_map_free(numfiles, newlomap, newhimap);
1056 		}
1057 		return;
1058 	}
1059 
1060 	/* Copy the existing descriptor table and zero the new portion. */
1061 	i = sizeof(fdfile_t *) * oldnfiles;
1062 	memcpy(newdt->dt_ff, dt->dt_ff, i);
1063 	memset((uint8_t *)newdt->dt_ff + i, 0,
1064 	    numfiles * sizeof(fdfile_t *) - i);
1065 
1066 	/*
1067 	 * Link old descriptor array into list to be discarded.  We defer
1068 	 * freeing until the last reference to the descriptor table goes
1069 	 * away (usually process exit).  This allows us to do lockless
1070 	 * lookups in fd_getfile().
1071 	 */
1072 	if (oldnfiles > NDFILE) {
1073 		if (fdp->fd_refcnt > 1) {
1074 			newdt->dt_link = dt;
1075 		} else {
1076 			fd_dtab_free(dt);
1077 		}
1078 	}
1079 
1080 	if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1081 		i = NDHISLOTS(oldnfiles) * sizeof(uint32_t);
1082 		memcpy(newhimap, fdp->fd_himap, i);
1083 		memset((uint8_t *)newhimap + i, 0,
1084 		    NDHISLOTS(numfiles) * sizeof(uint32_t) - i);
1085 
1086 		i = NDLOSLOTS(oldnfiles) * sizeof(uint32_t);
1087 		memcpy(newlomap, fdp->fd_lomap, i);
1088 		memset((uint8_t *)newlomap + i, 0,
1089 		    NDLOSLOTS(numfiles) * sizeof(uint32_t) - i);
1090 
1091 		if (NDHISLOTS(oldnfiles) > NDHISLOTS(NDFILE)) {
1092 			fd_map_free(oldnfiles, fdp->fd_lomap, fdp->fd_himap);
1093 		}
1094 		fdp->fd_himap = newhimap;
1095 		fdp->fd_lomap = newlomap;
1096 	}
1097 
1098 	/*
1099 	 * All other modifications must become globally visible before
1100 	 * the change to fd_dt.  See fd_getfile().
1101 	 */
1102 	atomic_store_release(&fdp->fd_dt, newdt);
1103 	KASSERT(newdt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1104 	fd_checkmaps(fdp);
1105 	mutex_exit(&fdp->fd_lock);
1106 }
1107 
1108 /*
1109  * Create a new open file structure and allocate a file descriptor
1110  * for the current process.
1111  */
1112 int
1113 fd_allocfile(file_t **resultfp, int *resultfd)
1114 {
1115 	proc_t *p = curproc;
1116 	kauth_cred_t cred;
1117 	file_t *fp;
1118 	int error;
1119 
1120 	while ((error = fd_alloc(p, 0, resultfd)) != 0) {
1121 		if (error != ENOSPC) {
1122 			return error;
1123 		}
1124 		fd_tryexpand(p);
1125 	}
1126 
1127 	fp = pool_cache_get(file_cache, PR_WAITOK);
1128 	if (fp == NULL) {
1129 		fd_abort(p, NULL, *resultfd);
1130 		return ENFILE;
1131 	}
1132 	KASSERT(fp->f_count == 0);
1133 	KASSERT(fp->f_msgcount == 0);
1134 	KASSERT(fp->f_unpcount == 0);
1135 
1136 	/* Replace cached credentials if not what we need. */
1137 	cred = curlwp->l_cred;
1138 	if (__predict_false(cred != fp->f_cred)) {
1139 		kauth_cred_free(fp->f_cred);
1140 		kauth_cred_hold(cred);
1141 		fp->f_cred = cred;
1142 	}
1143 
1144 	/*
1145 	 * Don't allow recycled files to be scanned.
1146 	 * See uipc_usrreq.c.
1147 	 */
1148 	if (__predict_false((fp->f_flag & FSCAN) != 0)) {
1149 		mutex_enter(&fp->f_lock);
1150 		atomic_and_uint(&fp->f_flag, ~FSCAN);
1151 		mutex_exit(&fp->f_lock);
1152 	}
1153 
1154 	fp->f_advice = 0;
1155 	fp->f_offset = 0;
1156 	*resultfp = fp;
1157 
1158 	return 0;
1159 }
1160 
1161 /*
1162  * Successful creation of a new descriptor: make visible to the process.
1163  */
1164 void
1165 fd_affix(proc_t *p, file_t *fp, unsigned fd)
1166 {
1167 	fdfile_t *ff;
1168 	filedesc_t *fdp;
1169 	fdtab_t *dt;
1170 
1171 	KASSERT(p == curproc || p == &proc0);
1172 
1173 	/* Add a reference to the file structure. */
1174 	mutex_enter(&fp->f_lock);
1175 	fp->f_count++;
1176 	mutex_exit(&fp->f_lock);
1177 
1178 	/*
1179 	 * Insert the new file into the descriptor slot.
1180 	 */
1181 	fdp = p->p_fd;
1182 	dt = atomic_load_consume(&fdp->fd_dt);
1183 	ff = dt->dt_ff[fd];
1184 
1185 	KASSERT(ff != NULL);
1186 	KASSERT(ff->ff_file == NULL);
1187 	KASSERT(ff->ff_allocated);
1188 	KASSERT(fd_isused(fdp, fd));
1189 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1190 
1191 	/* No need to lock in order to make file initially visible. */
1192 	atomic_store_release(&ff->ff_file, fp);
1193 }
1194 
1195 /*
1196  * Abort creation of a new descriptor: free descriptor slot and file.
1197  */
1198 void
1199 fd_abort(proc_t *p, file_t *fp, unsigned fd)
1200 {
1201 	filedesc_t *fdp;
1202 	fdfile_t *ff;
1203 
1204 	KASSERT(p == curproc || p == &proc0);
1205 
1206 	fdp = p->p_fd;
1207 	ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
1208 	ff->ff_exclose = false;
1209 
1210 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1211 
1212 	mutex_enter(&fdp->fd_lock);
1213 	KASSERT(fd_isused(fdp, fd));
1214 	fd_unused(fdp, fd);
1215 	mutex_exit(&fdp->fd_lock);
1216 
1217 	if (fp != NULL) {
1218 		KASSERT(fp->f_count == 0);
1219 		KASSERT(fp->f_cred != NULL);
1220 		pool_cache_put(file_cache, fp);
1221 	}
1222 }
1223 
1224 static int
1225 file_ctor(void *arg, void *obj, int flags)
1226 {
1227 	file_t *fp = obj;
1228 
1229 	memset(fp, 0, sizeof(*fp));
1230 
1231 	mutex_enter(&filelist_lock);
1232 	if (__predict_false(nfiles >= maxfiles)) {
1233 		mutex_exit(&filelist_lock);
1234 		tablefull("file", "increase kern.maxfiles or MAXFILES");
1235 		return ENFILE;
1236 	}
1237 	nfiles++;
1238 	LIST_INSERT_HEAD(&filehead, fp, f_list);
1239 	mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1240 	fp->f_cred = curlwp->l_cred;
1241 	kauth_cred_hold(fp->f_cred);
1242 	mutex_exit(&filelist_lock);
1243 
1244 	return 0;
1245 }
1246 
1247 static void
1248 file_dtor(void *arg, void *obj)
1249 {
1250 	file_t *fp = obj;
1251 
1252 	mutex_enter(&filelist_lock);
1253 	nfiles--;
1254 	LIST_REMOVE(fp, f_list);
1255 	mutex_exit(&filelist_lock);
1256 
1257 	KASSERT(fp->f_count == 0);
1258 	kauth_cred_free(fp->f_cred);
1259 	mutex_destroy(&fp->f_lock);
1260 }
1261 
1262 static int
1263 fdfile_ctor(void *arg, void *obj, int flags)
1264 {
1265 	fdfile_t *ff = obj;
1266 
1267 	memset(ff, 0, sizeof(*ff));
1268 	cv_init(&ff->ff_closing, "fdclose");
1269 
1270 	return 0;
1271 }
1272 
1273 static void
1274 fdfile_dtor(void *arg, void *obj)
1275 {
1276 	fdfile_t *ff = obj;
1277 
1278 	cv_destroy(&ff->ff_closing);
1279 }
1280 
1281 file_t *
1282 fgetdummy(void)
1283 {
1284 	file_t *fp;
1285 
1286 	fp = kmem_zalloc(sizeof(*fp), KM_SLEEP);
1287 	mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1288 	return fp;
1289 }
1290 
1291 void
1292 fputdummy(file_t *fp)
1293 {
1294 
1295 	mutex_destroy(&fp->f_lock);
1296 	kmem_free(fp, sizeof(*fp));
1297 }
1298 
1299 /*
1300  * Create an initial filedesc structure.
1301  */
1302 filedesc_t *
1303 fd_init(filedesc_t *fdp)
1304 {
1305 #ifdef DIAGNOSTIC
1306 	unsigned fd;
1307 #endif
1308 
1309 	if (__predict_true(fdp == NULL)) {
1310 		fdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1311 	} else {
1312 		KASSERT(fdp == &filedesc0);
1313 		filedesc_ctor(NULL, fdp, PR_WAITOK);
1314 	}
1315 
1316 #ifdef DIAGNOSTIC
1317 	KASSERT(fdp->fd_lastfile == -1);
1318 	KASSERT(fdp->fd_lastkqfile == -1);
1319 	KASSERT(fdp->fd_knhash == NULL);
1320 	KASSERT(fdp->fd_freefile == 0);
1321 	KASSERT(fdp->fd_exclose == false);
1322 	KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1323 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1324 	for (fd = 0; fd < NDFDFILE; fd++) {
1325 		KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] ==
1326 		    (fdfile_t *)fdp->fd_dfdfile[fd]);
1327 	}
1328 	for (fd = NDFDFILE; fd < NDFILE; fd++) {
1329 		KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] == NULL);
1330 	}
1331 	KASSERT(fdp->fd_himap == fdp->fd_dhimap);
1332 	KASSERT(fdp->fd_lomap == fdp->fd_dlomap);
1333 #endif	/* DIAGNOSTIC */
1334 
1335 	fdp->fd_refcnt = 1;
1336 	fd_checkmaps(fdp);
1337 
1338 	return fdp;
1339 }
1340 
1341 /*
1342  * Initialize a file descriptor table.
1343  */
1344 static int
1345 filedesc_ctor(void *arg, void *obj, int flag)
1346 {
1347 	filedesc_t *fdp = obj;
1348 	fdfile_t **ffp;
1349 	int i;
1350 
1351 	memset(fdp, 0, sizeof(*fdp));
1352 	mutex_init(&fdp->fd_lock, MUTEX_DEFAULT, IPL_NONE);
1353 	fdp->fd_lastfile = -1;
1354 	fdp->fd_lastkqfile = -1;
1355 	fdp->fd_dt = &fdp->fd_dtbuiltin;
1356 	fdp->fd_dtbuiltin.dt_nfiles = NDFILE;
1357 	fdp->fd_himap = fdp->fd_dhimap;
1358 	fdp->fd_lomap = fdp->fd_dlomap;
1359 
1360 	CTASSERT(sizeof(fdp->fd_dfdfile[0]) >= sizeof(fdfile_t));
1361 	for (i = 0, ffp = fdp->fd_dt->dt_ff; i < NDFDFILE; i++, ffp++) {
1362 		*ffp = (fdfile_t *)fdp->fd_dfdfile[i];
1363 		(void)fdfile_ctor(NULL, fdp->fd_dfdfile[i], PR_WAITOK);
1364 	}
1365 
1366 	return 0;
1367 }
1368 
1369 static void
1370 filedesc_dtor(void *arg, void *obj)
1371 {
1372 	filedesc_t *fdp = obj;
1373 	int i;
1374 
1375 	for (i = 0; i < NDFDFILE; i++) {
1376 		fdfile_dtor(NULL, fdp->fd_dfdfile[i]);
1377 	}
1378 
1379 	mutex_destroy(&fdp->fd_lock);
1380 }
1381 
1382 /*
1383  * Make p share curproc's filedesc structure.
1384  */
1385 void
1386 fd_share(struct proc *p)
1387 {
1388 	filedesc_t *fdp;
1389 
1390 	fdp = curlwp->l_fd;
1391 	p->p_fd = fdp;
1392 	atomic_inc_uint(&fdp->fd_refcnt);
1393 }
1394 
1395 /*
1396  * Acquire a hold on a filedesc structure.
1397  */
1398 void
1399 fd_hold(lwp_t *l)
1400 {
1401 	filedesc_t *fdp = l->l_fd;
1402 
1403 	atomic_inc_uint(&fdp->fd_refcnt);
1404 }
1405 
1406 /*
1407  * Copy a filedesc structure.
1408  */
1409 filedesc_t *
1410 fd_copy(void)
1411 {
1412 	filedesc_t *newfdp, *fdp;
1413 	fdfile_t *ff, **ffp, **nffp, *ff2;
1414 	int i, j, numfiles, lastfile, newlast;
1415 	file_t *fp;
1416 	fdtab_t *newdt;
1417 
1418 	fdp = curproc->p_fd;
1419 	newfdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1420 	newfdp->fd_refcnt = 1;
1421 
1422 #ifdef DIAGNOSTIC
1423 	KASSERT(newfdp->fd_lastfile == -1);
1424 	KASSERT(newfdp->fd_lastkqfile == -1);
1425 	KASSERT(newfdp->fd_knhash == NULL);
1426 	KASSERT(newfdp->fd_freefile == 0);
1427 	KASSERT(newfdp->fd_exclose == false);
1428 	KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1429 	KASSERT(newfdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1430 	for (i = 0; i < NDFDFILE; i++) {
1431 		KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] ==
1432 		    (fdfile_t *)&newfdp->fd_dfdfile[i]);
1433 	}
1434 	for (i = NDFDFILE; i < NDFILE; i++) {
1435 		KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] == NULL);
1436 	}
1437 #endif	/* DIAGNOSTIC */
1438 
1439 	mutex_enter(&fdp->fd_lock);
1440 	fd_checkmaps(fdp);
1441 	numfiles = fdp->fd_dt->dt_nfiles;
1442 	lastfile = fdp->fd_lastfile;
1443 
1444 	/*
1445 	 * If the number of open files fits in the internal arrays
1446 	 * of the open file structure, use them, otherwise allocate
1447 	 * additional memory for the number of descriptors currently
1448 	 * in use.
1449 	 */
1450 	if (lastfile < NDFILE) {
1451 		i = NDFILE;
1452 		newdt = newfdp->fd_dt;
1453 		KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1454 	} else {
1455 		/*
1456 		 * Compute the smallest multiple of NDEXTENT needed
1457 		 * for the file descriptors currently in use,
1458 		 * allowing the table to shrink.
1459 		 */
1460 		i = numfiles;
1461 		while (i >= 2 * NDEXTENT && i > lastfile * 2) {
1462 			i /= 2;
1463 		}
1464 		KASSERT(i > NDFILE);
1465 		newdt = fd_dtab_alloc(i);
1466 		newfdp->fd_dt = newdt;
1467 		memcpy(newdt->dt_ff, newfdp->fd_dtbuiltin.dt_ff,
1468 		    NDFDFILE * sizeof(fdfile_t **));
1469 		memset(newdt->dt_ff + NDFDFILE, 0,
1470 		    (i - NDFDFILE) * sizeof(fdfile_t **));
1471 	}
1472 	if (NDHISLOTS(i) <= NDHISLOTS(NDFILE)) {
1473 		newfdp->fd_himap = newfdp->fd_dhimap;
1474 		newfdp->fd_lomap = newfdp->fd_dlomap;
1475 	} else {
1476 		fd_map_alloc(i, &newfdp->fd_lomap, &newfdp->fd_himap);
1477 		KASSERT(i >= NDENTRIES * NDENTRIES);
1478 		memset(newfdp->fd_himap, 0, NDHISLOTS(i)*sizeof(uint32_t));
1479 		memset(newfdp->fd_lomap, 0, NDLOSLOTS(i)*sizeof(uint32_t));
1480 	}
1481 	newfdp->fd_freefile = fdp->fd_freefile;
1482 	newfdp->fd_exclose = fdp->fd_exclose;
1483 
1484 	ffp = fdp->fd_dt->dt_ff;
1485 	nffp = newdt->dt_ff;
1486 	newlast = -1;
1487 	for (i = 0; i <= lastfile; i++, ffp++, nffp++) {
1488 		KASSERT(i >= NDFDFILE ||
1489 		    *nffp == (fdfile_t *)newfdp->fd_dfdfile[i]);
1490 		ff = *ffp;
1491 		if (ff == NULL ||
1492 		    (fp = atomic_load_consume(&ff->ff_file)) == NULL) {
1493 			/* Descriptor unused, or descriptor half open. */
1494 			KASSERT(!fd_isused(newfdp, i));
1495 			continue;
1496 		}
1497 		if (__predict_false(fp->f_type == DTYPE_KQUEUE)) {
1498 			/* kqueue descriptors cannot be copied. */
1499 			if (i < newfdp->fd_freefile) {
1500 				newfdp->fd_freefile = i;
1501 			}
1502 			continue;
1503 		}
1504 		/* It's active: add a reference to the file. */
1505 		mutex_enter(&fp->f_lock);
1506 		fp->f_count++;
1507 		mutex_exit(&fp->f_lock);
1508 
1509 		/* Allocate an fdfile_t to represent it. */
1510 		if (i >= NDFDFILE) {
1511 			ff2 = pool_cache_get(fdfile_cache, PR_WAITOK);
1512 			*nffp = ff2;
1513 		} else {
1514 			ff2 = newdt->dt_ff[i];
1515 		}
1516 		ff2->ff_file = fp;
1517 		ff2->ff_exclose = ff->ff_exclose;
1518 		ff2->ff_allocated = true;
1519 
1520 		/* Fix up bitmaps. */
1521 		j = i >> NDENTRYSHIFT;
1522 		KASSERT((newfdp->fd_lomap[j] & (1U << (i & NDENTRYMASK))) == 0);
1523 		newfdp->fd_lomap[j] |= 1U << (i & NDENTRYMASK);
1524 		if (__predict_false(newfdp->fd_lomap[j] == ~0)) {
1525 			KASSERT((newfdp->fd_himap[j >> NDENTRYSHIFT] &
1526 			    (1U << (j & NDENTRYMASK))) == 0);
1527 			newfdp->fd_himap[j >> NDENTRYSHIFT] |=
1528 			    1U << (j & NDENTRYMASK);
1529 		}
1530 		newlast = i;
1531 	}
1532 	KASSERT(newdt->dt_ff[0] == (fdfile_t *)newfdp->fd_dfdfile[0]);
1533 	newfdp->fd_lastfile = newlast;
1534 	fd_checkmaps(newfdp);
1535 	mutex_exit(&fdp->fd_lock);
1536 
1537 	return newfdp;
1538 }
1539 
1540 /*
1541  * Release a filedesc structure.
1542  */
1543 void
1544 fd_free(void)
1545 {
1546 	fdfile_t *ff;
1547 	file_t *fp;
1548 	int fd, nf;
1549 	fdtab_t *dt;
1550 	lwp_t * const l = curlwp;
1551 	filedesc_t * const fdp = l->l_fd;
1552 	const bool noadvlock = (l->l_proc->p_flag & PK_ADVLOCK) == 0;
1553 
1554 	KASSERT(atomic_load_consume(&fdp->fd_dt)->dt_ff[0] ==
1555 	    (fdfile_t *)fdp->fd_dfdfile[0]);
1556 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1557 	KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1558 
1559 	membar_release();
1560 	if (atomic_dec_uint_nv(&fdp->fd_refcnt) > 0)
1561 		return;
1562 	membar_acquire();
1563 
1564 	/*
1565 	 * Close any files that the process holds open.
1566 	 */
1567 	dt = fdp->fd_dt;
1568 	fd_checkmaps(fdp);
1569 #ifdef DEBUG
1570 	fdp->fd_refcnt = -1; /* see fd_checkmaps */
1571 #endif
1572 	for (fd = 0, nf = dt->dt_nfiles; fd < nf; fd++) {
1573 		ff = dt->dt_ff[fd];
1574 		KASSERT(fd >= NDFDFILE ||
1575 		    ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1576 		if (ff == NULL)
1577 			continue;
1578 		if ((fp = atomic_load_consume(&ff->ff_file)) != NULL) {
1579 			/*
1580 			 * Must use fd_close() here if there is
1581 			 * a reference from kqueue or we might have posix
1582 			 * advisory locks.
1583 			 */
1584 			if (__predict_true(ff->ff_refcnt == 0) &&
1585 			    (noadvlock || fp->f_type != DTYPE_VNODE)) {
1586 				ff->ff_file = NULL;
1587 				ff->ff_exclose = false;
1588 				ff->ff_allocated = false;
1589 				closef(fp);
1590 			} else {
1591 				ff->ff_refcnt++;
1592 				fd_close(fd);
1593 			}
1594 		}
1595 		KASSERT(ff->ff_refcnt == 0);
1596 		KASSERT(ff->ff_file == NULL);
1597 		KASSERT(!ff->ff_exclose);
1598 		KASSERT(!ff->ff_allocated);
1599 		if (fd >= NDFDFILE) {
1600 			pool_cache_put(fdfile_cache, ff);
1601 			dt->dt_ff[fd] = NULL;
1602 		}
1603 	}
1604 
1605 	/*
1606 	 * Clean out the descriptor table for the next user and return
1607 	 * to the cache.
1608 	 */
1609 	if (__predict_false(dt != &fdp->fd_dtbuiltin)) {
1610 		fd_dtab_free(fdp->fd_dt);
1611 		/* Otherwise, done above. */
1612 		memset(&fdp->fd_dtbuiltin.dt_ff[NDFDFILE], 0,
1613 		    (NDFILE - NDFDFILE) * sizeof(fdp->fd_dtbuiltin.dt_ff[0]));
1614 		fdp->fd_dt = &fdp->fd_dtbuiltin;
1615 	}
1616 	if (__predict_false(NDHISLOTS(nf) > NDHISLOTS(NDFILE))) {
1617 		KASSERT(fdp->fd_himap != fdp->fd_dhimap);
1618 		KASSERT(fdp->fd_lomap != fdp->fd_dlomap);
1619 		fd_map_free(nf, fdp->fd_lomap, fdp->fd_himap);
1620 	}
1621 	if (__predict_false(fdp->fd_knhash != NULL)) {
1622 		hashdone(fdp->fd_knhash, HASH_LIST, fdp->fd_knhashmask);
1623 		fdp->fd_knhash = NULL;
1624 		fdp->fd_knhashmask = 0;
1625 	} else {
1626 		KASSERT(fdp->fd_knhashmask == 0);
1627 	}
1628 	fdp->fd_dt = &fdp->fd_dtbuiltin;
1629 	fdp->fd_lastkqfile = -1;
1630 	fdp->fd_lastfile = -1;
1631 	fdp->fd_freefile = 0;
1632 	fdp->fd_exclose = false;
1633 	memset(&fdp->fd_startzero, 0, sizeof(*fdp) -
1634 	    offsetof(filedesc_t, fd_startzero));
1635 	fdp->fd_himap = fdp->fd_dhimap;
1636 	fdp->fd_lomap = fdp->fd_dlomap;
1637 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1638 	KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1639 	KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1640 #ifdef DEBUG
1641 	fdp->fd_refcnt = 0; /* see fd_checkmaps */
1642 #endif
1643 	fd_checkmaps(fdp);
1644 	pool_cache_put(filedesc_cache, fdp);
1645 }
1646 
1647 /*
1648  * File Descriptor pseudo-device driver (/dev/fd/).
1649  *
1650  * Opening minor device N dup()s the file (if any) connected to file
1651  * descriptor N belonging to the calling process.  Note that this driver
1652  * consists of only the ``open()'' routine, because all subsequent
1653  * references to this file will be direct to the other driver.
1654  */
1655 static int
1656 filedescopen(dev_t dev, int mode, int type, lwp_t *l)
1657 {
1658 
1659 	/*
1660 	 * XXX Kludge: set dupfd to contain the value of the
1661 	 * the file descriptor being sought for duplication. The error
1662 	 * return ensures that the vnode for this device will be released
1663 	 * by vn_open. Open will detect this special error and take the
1664 	 * actions in fd_dupopen below. Other callers of vn_open or VOP_OPEN
1665 	 * will simply report the error.
1666 	 */
1667 	l->l_dupfd = minor(dev);	/* XXX */
1668 	return EDUPFD;
1669 }
1670 
1671 /*
1672  * Duplicate the specified descriptor to a free descriptor.
1673  *
1674  * old is the original fd.
1675  * moveit is true if we should move rather than duplicate.
1676  * flags are the open flags (converted from O_* to F*).
1677  * newp returns the new fd on success.
1678  *
1679  * These two cases are produced by the EDUPFD and EMOVEFD magic
1680  * errnos, but in the interest of removing that regrettable interface,
1681  * vn_open has been changed to intercept them. Now vn_open returns
1682  * either a vnode or a filehandle, and the filehandle is accompanied
1683  * by a boolean that says whether we should dup (moveit == false) or
1684  * move (moveit == true) the fd.
1685  *
1686  * The dup case is used by /dev/stderr, /proc/self/fd, and such. The
1687  * move case is used by cloner devices that allocate a fd of their
1688  * own (a layering violation that should go away eventually) that
1689  * then needs to be put in the place open() expects it.
1690  */
1691 int
1692 fd_dupopen(int old, bool moveit, int flags, int *newp)
1693 {
1694 	filedesc_t *fdp;
1695 	fdfile_t *ff;
1696 	file_t *fp;
1697 	fdtab_t *dt;
1698 	int error;
1699 
1700 	if ((fp = fd_getfile(old)) == NULL) {
1701 		return EBADF;
1702 	}
1703 	fdp = curlwp->l_fd;
1704 	dt = atomic_load_consume(&fdp->fd_dt);
1705 	ff = dt->dt_ff[old];
1706 
1707 	/*
1708 	 * There are two cases of interest here.
1709 	 *
1710 	 * 1. moveit == false (used to be the EDUPFD magic errno):
1711 	 *    simply dup (old) to file descriptor (new) and return.
1712 	 *
1713 	 * 2. moveit == true (used to be the EMOVEFD magic errno):
1714 	 *    steal away the file structure from (old) and store it in
1715 	 *    (new).  (old) is effectively closed by this operation.
1716 	 */
1717 	if (moveit == false) {
1718 		/*
1719 		 * Check that the mode the file is being opened for is a
1720 		 * subset of the mode of the existing descriptor.
1721 		 */
1722 		if (((flags & (FREAD|FWRITE)) | fp->f_flag) != fp->f_flag) {
1723 			error = EACCES;
1724 			goto out;
1725 		}
1726 
1727 		/* Copy it. */
1728 		error = fd_dup(fp, 0, newp, ff->ff_exclose);
1729 	} else {
1730 		/* Copy it. */
1731 		error = fd_dup(fp, 0, newp, ff->ff_exclose);
1732 		if (error != 0) {
1733 			goto out;
1734 		}
1735 
1736 		/* Steal away the file pointer from 'old'. */
1737 		(void)fd_close(old);
1738 		return 0;
1739 	}
1740 
1741 out:
1742 	fd_putfile(old);
1743 	return error;
1744 }
1745 
1746 /*
1747  * Close open files on exec.
1748  */
1749 void
1750 fd_closeexec(void)
1751 {
1752 	proc_t *p;
1753 	filedesc_t *fdp;
1754 	fdfile_t *ff;
1755 	lwp_t *l;
1756 	fdtab_t *dt;
1757 	int fd;
1758 
1759 	l = curlwp;
1760 	p = l->l_proc;
1761 	fdp = p->p_fd;
1762 
1763 	if (fdp->fd_refcnt > 1) {
1764 		fdp = fd_copy();
1765 		fd_free();
1766 		p->p_fd = fdp;
1767 		l->l_fd = fdp;
1768 	}
1769 	if (!fdp->fd_exclose) {
1770 		return;
1771 	}
1772 	fdp->fd_exclose = false;
1773 	dt = atomic_load_consume(&fdp->fd_dt);
1774 
1775 	for (fd = 0; fd <= fdp->fd_lastfile; fd++) {
1776 		if ((ff = dt->dt_ff[fd]) == NULL) {
1777 			KASSERT(fd >= NDFDFILE);
1778 			continue;
1779 		}
1780 		KASSERT(fd >= NDFDFILE ||
1781 		    ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1782 		if (ff->ff_file == NULL)
1783 			continue;
1784 		if (ff->ff_exclose) {
1785 			/*
1786 			 * We need a reference to close the file.
1787 			 * No other threads can see the fdfile_t at
1788 			 * this point, so don't bother locking.
1789 			 */
1790 			KASSERT((ff->ff_refcnt & FR_CLOSING) == 0);
1791 			ff->ff_refcnt++;
1792 			fd_close(fd);
1793 		}
1794 	}
1795 }
1796 
1797 /*
1798  * Sets descriptor owner. If the owner is a process, 'pgid'
1799  * is set to positive value, process ID. If the owner is process group,
1800  * 'pgid' is set to -pg_id.
1801  */
1802 int
1803 fsetown(pid_t *pgid, u_long cmd, const void *data)
1804 {
1805 	pid_t id = *(const pid_t *)data;
1806 	int error;
1807 
1808 	if (id == INT_MIN)
1809 		return EINVAL;
1810 
1811 	switch (cmd) {
1812 	case TIOCSPGRP:
1813 		if (id < 0)
1814 			return EINVAL;
1815 		id = -id;
1816 		break;
1817 	default:
1818 		break;
1819 	}
1820 	if (id > 0) {
1821 		mutex_enter(&proc_lock);
1822 		error = proc_find(id) ? 0 : ESRCH;
1823 		mutex_exit(&proc_lock);
1824 	} else if (id < 0) {
1825 		error = pgid_in_session(curproc, -id);
1826 	} else {
1827 		error = 0;
1828 	}
1829 	if (!error) {
1830 		*pgid = id;
1831 	}
1832 	return error;
1833 }
1834 
1835 void
1836 fd_set_exclose(struct lwp *l, int fd, bool exclose)
1837 {
1838 	filedesc_t *fdp = l->l_fd;
1839 	fdfile_t *ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
1840 
1841 	ff->ff_exclose = exclose;
1842 	if (exclose)
1843 		fdp->fd_exclose = true;
1844 }
1845 
1846 /*
1847  * Return descriptor owner information. If the value is positive,
1848  * it's process ID. If it's negative, it's process group ID and
1849  * needs the sign removed before use.
1850  */
1851 int
1852 fgetown(pid_t pgid, u_long cmd, void *data)
1853 {
1854 
1855 	switch (cmd) {
1856 	case TIOCGPGRP:
1857 		*(int *)data = -pgid;
1858 		break;
1859 	default:
1860 		*(int *)data = pgid;
1861 		break;
1862 	}
1863 	return 0;
1864 }
1865 
1866 /*
1867  * Send signal to descriptor owner, either process or process group.
1868  */
1869 void
1870 fownsignal(pid_t pgid, int signo, int code, int band, void *fdescdata)
1871 {
1872 	ksiginfo_t ksi;
1873 
1874 	KASSERT(!cpu_intr_p());
1875 
1876 	if (pgid == 0) {
1877 		return;
1878 	}
1879 
1880 	KSI_INIT(&ksi);
1881 	ksi.ksi_signo = signo;
1882 	ksi.ksi_code = code;
1883 	ksi.ksi_band = band;
1884 
1885 	mutex_enter(&proc_lock);
1886 	if (pgid > 0) {
1887 		struct proc *p1;
1888 
1889 		p1 = proc_find(pgid);
1890 		if (p1 != NULL) {
1891 			kpsignal(p1, &ksi, fdescdata);
1892 		}
1893 	} else {
1894 		struct pgrp *pgrp;
1895 
1896 		KASSERT(pgid < 0);
1897 		pgrp = pgrp_find(-pgid);
1898 		if (pgrp != NULL) {
1899 			kpgsignal(pgrp, &ksi, fdescdata, 0);
1900 		}
1901 	}
1902 	mutex_exit(&proc_lock);
1903 }
1904 
1905 int
1906 fd_clone(file_t *fp, unsigned fd, int flag, const struct fileops *fops,
1907 	 void *data)
1908 {
1909 	fdfile_t *ff;
1910 	filedesc_t *fdp;
1911 
1912 	fp->f_flag = flag & FMASK;
1913 	fdp = curproc->p_fd;
1914 	ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
1915 	KASSERT(ff != NULL);
1916 	ff->ff_exclose = (flag & O_CLOEXEC) != 0;
1917 	fp->f_type = DTYPE_MISC;
1918 	fp->f_ops = fops;
1919 	fp->f_data = data;
1920 	curlwp->l_dupfd = fd;
1921 	fd_affix(curproc, fp, fd);
1922 
1923 	return EMOVEFD;
1924 }
1925 
1926 int
1927 fnullop_fcntl(file_t *fp, u_int cmd, void *data)
1928 {
1929 
1930 	if (cmd == F_SETFL)
1931 		return 0;
1932 
1933 	return EOPNOTSUPP;
1934 }
1935 
1936 int
1937 fnullop_poll(file_t *fp, int which)
1938 {
1939 
1940 	return 0;
1941 }
1942 
1943 int
1944 fnullop_kqfilter(file_t *fp, struct knote *kn)
1945 {
1946 
1947 	return EOPNOTSUPP;
1948 }
1949 
1950 void
1951 fnullop_restart(file_t *fp)
1952 {
1953 
1954 }
1955 
1956 int
1957 fbadop_read(file_t *fp, off_t *offset, struct uio *uio,
1958 	    kauth_cred_t cred, int flags)
1959 {
1960 
1961 	return EOPNOTSUPP;
1962 }
1963 
1964 int
1965 fbadop_write(file_t *fp, off_t *offset, struct uio *uio,
1966 	     kauth_cred_t cred, int flags)
1967 {
1968 
1969 	return EOPNOTSUPP;
1970 }
1971 
1972 int
1973 fbadop_ioctl(file_t *fp, u_long com, void *data)
1974 {
1975 
1976 	return EOPNOTSUPP;
1977 }
1978 
1979 int
1980 fbadop_stat(file_t *fp, struct stat *sb)
1981 {
1982 
1983 	return EOPNOTSUPP;
1984 }
1985 
1986 int
1987 fbadop_close(file_t *fp)
1988 {
1989 
1990 	return EOPNOTSUPP;
1991 }
1992 
1993 /*
1994  * sysctl routines pertaining to file descriptors
1995  */
1996 
1997 /* Initialized in sysctl_init() for now... */
1998 extern kmutex_t sysctl_file_marker_lock;
1999 static u_int sysctl_file_marker = 1;
2000 
2001 /*
2002  * Expects to be called with proc_lock and sysctl_file_marker_lock locked.
2003  */
2004 static void
2005 sysctl_file_marker_reset(void)
2006 {
2007 	struct proc *p;
2008 
2009 	PROCLIST_FOREACH(p, &allproc) {
2010 		struct filedesc *fd = p->p_fd;
2011 		fdtab_t *dt;
2012 		u_int i;
2013 
2014 		mutex_enter(&fd->fd_lock);
2015 		dt = fd->fd_dt;
2016 		for (i = 0; i < dt->dt_nfiles; i++) {
2017 			struct file *fp;
2018 			fdfile_t *ff;
2019 
2020 			if ((ff = dt->dt_ff[i]) == NULL) {
2021 				continue;
2022 			}
2023 			if ((fp = atomic_load_consume(&ff->ff_file)) == NULL) {
2024 				continue;
2025 			}
2026 			fp->f_marker = 0;
2027 		}
2028 		mutex_exit(&fd->fd_lock);
2029 	}
2030 }
2031 
2032 /*
2033  * sysctl helper routine for kern.file pseudo-subtree.
2034  */
2035 static int
2036 sysctl_kern_file(SYSCTLFN_ARGS)
2037 {
2038 	const bool allowaddr = get_expose_address(curproc);
2039 	struct filelist flist;
2040 	int error;
2041 	size_t buflen;
2042 	struct file *fp, fbuf;
2043 	char *start, *where;
2044 	struct proc *p;
2045 
2046 	start = where = oldp;
2047 	buflen = *oldlenp;
2048 
2049 	if (where == NULL) {
2050 		/*
2051 		 * overestimate by 10 files
2052 		 */
2053 		*oldlenp = sizeof(filehead) + (nfiles + 10) *
2054 		    sizeof(struct file);
2055 		return 0;
2056 	}
2057 
2058 	/*
2059 	 * first sysctl_copyout filehead
2060 	 */
2061 	if (buflen < sizeof(filehead)) {
2062 		*oldlenp = 0;
2063 		return 0;
2064 	}
2065 	sysctl_unlock();
2066 	if (allowaddr) {
2067 		memcpy(&flist, &filehead, sizeof(flist));
2068 	} else {
2069 		memset(&flist, 0, sizeof(flist));
2070 	}
2071 	error = sysctl_copyout(l, &flist, where, sizeof(flist));
2072 	if (error) {
2073 		sysctl_relock();
2074 		return error;
2075 	}
2076 	buflen -= sizeof(flist);
2077 	where += sizeof(flist);
2078 
2079 	/*
2080 	 * followed by an array of file structures
2081 	 */
2082 	mutex_enter(&sysctl_file_marker_lock);
2083 	mutex_enter(&proc_lock);
2084 	PROCLIST_FOREACH(p, &allproc) {
2085 		struct filedesc *fd;
2086 		fdtab_t *dt;
2087 		u_int i;
2088 
2089 		if (p->p_stat == SIDL) {
2090 			/* skip embryonic processes */
2091 			continue;
2092 		}
2093 		mutex_enter(p->p_lock);
2094 		error = kauth_authorize_process(l->l_cred,
2095 		    KAUTH_PROCESS_CANSEE, p,
2096 		    KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2097 		    NULL, NULL);
2098 		mutex_exit(p->p_lock);
2099 		if (error != 0) {
2100 			/*
2101 			 * Don't leak kauth retval if we're silently
2102 			 * skipping this entry.
2103 			 */
2104 			error = 0;
2105 			continue;
2106 		}
2107 
2108 		/*
2109 		 * Grab a hold on the process.
2110 		 */
2111 		if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2112 			continue;
2113 		}
2114 		mutex_exit(&proc_lock);
2115 
2116 		fd = p->p_fd;
2117 		mutex_enter(&fd->fd_lock);
2118 		dt = fd->fd_dt;
2119 		for (i = 0; i < dt->dt_nfiles; i++) {
2120 			fdfile_t *ff;
2121 
2122 			if ((ff = dt->dt_ff[i]) == NULL) {
2123 				continue;
2124 			}
2125 			if ((fp = atomic_load_consume(&ff->ff_file)) == NULL) {
2126 				continue;
2127 			}
2128 
2129 			mutex_enter(&fp->f_lock);
2130 
2131 			if ((fp->f_count == 0) ||
2132 			    (fp->f_marker == sysctl_file_marker)) {
2133 				mutex_exit(&fp->f_lock);
2134 				continue;
2135 			}
2136 
2137 			/* Check that we have enough space. */
2138 			if (buflen < sizeof(struct file)) {
2139 				*oldlenp = where - start;
2140 				mutex_exit(&fp->f_lock);
2141 				error = ENOMEM;
2142 				break;
2143 			}
2144 
2145 			fill_file(&fbuf, fp);
2146 			mutex_exit(&fp->f_lock);
2147 			error = sysctl_copyout(l, &fbuf, where, sizeof(fbuf));
2148 			if (error) {
2149 				break;
2150 			}
2151 			buflen -= sizeof(struct file);
2152 			where += sizeof(struct file);
2153 
2154 			fp->f_marker = sysctl_file_marker;
2155 		}
2156 		mutex_exit(&fd->fd_lock);
2157 
2158 		/*
2159 		 * Release reference to process.
2160 		 */
2161 		mutex_enter(&proc_lock);
2162 		rw_exit(&p->p_reflock);
2163 
2164 		if (error)
2165 			break;
2166 	}
2167 
2168 	sysctl_file_marker++;
2169 	/* Reset all markers if wrapped. */
2170 	if (sysctl_file_marker == 0) {
2171 		sysctl_file_marker_reset();
2172 		sysctl_file_marker++;
2173 	}
2174 
2175 	mutex_exit(&proc_lock);
2176 	mutex_exit(&sysctl_file_marker_lock);
2177 
2178 	*oldlenp = where - start;
2179 	sysctl_relock();
2180 	return error;
2181 }
2182 
2183 /*
2184  * sysctl helper function for kern.file2
2185  */
2186 static int
2187 sysctl_kern_file2(SYSCTLFN_ARGS)
2188 {
2189 	struct proc *p;
2190 	struct file *fp;
2191 	struct filedesc *fd;
2192 	struct kinfo_file kf;
2193 	char *dp;
2194 	u_int i, op;
2195 	size_t len, needed, elem_size, out_size;
2196 	int error, arg, elem_count;
2197 	fdfile_t *ff;
2198 	fdtab_t *dt;
2199 
2200 	if (namelen == 1 && name[0] == CTL_QUERY)
2201 		return sysctl_query(SYSCTLFN_CALL(rnode));
2202 
2203 	if (namelen != 4)
2204 		return EINVAL;
2205 
2206 	error = 0;
2207 	dp = oldp;
2208 	len = (oldp != NULL) ? *oldlenp : 0;
2209 	op = name[0];
2210 	arg = name[1];
2211 	elem_size = name[2];
2212 	elem_count = name[3];
2213 	out_size = MIN(sizeof(kf), elem_size);
2214 	needed = 0;
2215 
2216 	if (elem_size < 1 || elem_count < 0)
2217 		return EINVAL;
2218 
2219 	switch (op) {
2220 	case KERN_FILE_BYFILE:
2221 	case KERN_FILE_BYPID:
2222 		/*
2223 		 * We're traversing the process list in both cases; the BYFILE
2224 		 * case does additional work of keeping track of files already
2225 		 * looked at.
2226 		 */
2227 
2228 		/* doesn't use arg so it must be zero */
2229 		if ((op == KERN_FILE_BYFILE) && (arg != 0))
2230 			return EINVAL;
2231 
2232 		if ((op == KERN_FILE_BYPID) && (arg < -1))
2233 			/* -1 means all processes */
2234 			return EINVAL;
2235 
2236 		sysctl_unlock();
2237 		if (op == KERN_FILE_BYFILE)
2238 			mutex_enter(&sysctl_file_marker_lock);
2239 		mutex_enter(&proc_lock);
2240 		PROCLIST_FOREACH(p, &allproc) {
2241 			if (p->p_stat == SIDL) {
2242 				/* skip embryonic processes */
2243 				continue;
2244 			}
2245 			if (arg > 0 && p->p_pid != arg) {
2246 				/* pick only the one we want */
2247 				/* XXX want 0 to mean "kernel files" */
2248 				continue;
2249 			}
2250 			mutex_enter(p->p_lock);
2251 			error = kauth_authorize_process(l->l_cred,
2252 			    KAUTH_PROCESS_CANSEE, p,
2253 			    KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2254 			    NULL, NULL);
2255 			mutex_exit(p->p_lock);
2256 			if (error != 0) {
2257 				/*
2258 				 * Don't leak kauth retval if we're silently
2259 				 * skipping this entry.
2260 				 */
2261 				error = 0;
2262 				continue;
2263 			}
2264 
2265 			/*
2266 			 * Grab a hold on the process.
2267 			 */
2268 			if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2269 				continue;
2270 			}
2271 			mutex_exit(&proc_lock);
2272 
2273 			fd = p->p_fd;
2274 			mutex_enter(&fd->fd_lock);
2275 			dt = fd->fd_dt;
2276 			for (i = 0; i < dt->dt_nfiles; i++) {
2277 				if ((ff = dt->dt_ff[i]) == NULL) {
2278 					continue;
2279 				}
2280 				if ((fp = atomic_load_consume(&ff->ff_file)) ==
2281 				    NULL) {
2282 					continue;
2283 				}
2284 
2285 				if ((op == KERN_FILE_BYFILE) &&
2286 				    (fp->f_marker == sysctl_file_marker)) {
2287 					continue;
2288 				}
2289 				if (len >= elem_size && elem_count > 0) {
2290 					mutex_enter(&fp->f_lock);
2291 					fill_file2(&kf, fp, ff, i, p->p_pid);
2292 					mutex_exit(&fp->f_lock);
2293 					mutex_exit(&fd->fd_lock);
2294 					error = sysctl_copyout(l,
2295 					    &kf, dp, out_size);
2296 					mutex_enter(&fd->fd_lock);
2297 					if (error)
2298 						break;
2299 					dp += elem_size;
2300 					len -= elem_size;
2301 				}
2302 				if (op == KERN_FILE_BYFILE)
2303 					fp->f_marker = sysctl_file_marker;
2304 				needed += elem_size;
2305 				if (elem_count > 0 && elem_count != INT_MAX)
2306 					elem_count--;
2307 			}
2308 			mutex_exit(&fd->fd_lock);
2309 
2310 			/*
2311 			 * Release reference to process.
2312 			 */
2313 			mutex_enter(&proc_lock);
2314 			rw_exit(&p->p_reflock);
2315 		}
2316 		if (op == KERN_FILE_BYFILE) {
2317 			sysctl_file_marker++;
2318 
2319 			/* Reset all markers if wrapped. */
2320 			if (sysctl_file_marker == 0) {
2321 				sysctl_file_marker_reset();
2322 				sysctl_file_marker++;
2323 			}
2324 		}
2325 		mutex_exit(&proc_lock);
2326 		if (op == KERN_FILE_BYFILE)
2327 			mutex_exit(&sysctl_file_marker_lock);
2328 		sysctl_relock();
2329 		break;
2330 	default:
2331 		return EINVAL;
2332 	}
2333 
2334 	if (oldp == NULL)
2335 		needed += KERN_FILESLOP * elem_size;
2336 	*oldlenp = needed;
2337 
2338 	return error;
2339 }
2340 
2341 static void
2342 fill_file(struct file *fp, const struct file *fpsrc)
2343 {
2344 	const bool allowaddr = get_expose_address(curproc);
2345 
2346 	memset(fp, 0, sizeof(*fp));
2347 
2348 	fp->f_offset = fpsrc->f_offset;
2349 	COND_SET_PTR(fp->f_cred, fpsrc->f_cred, allowaddr);
2350 	COND_SET_CPTR(fp->f_ops, fpsrc->f_ops, allowaddr);
2351 	COND_SET_STRUCT(fp->f_undata, fpsrc->f_undata, allowaddr);
2352 	COND_SET_STRUCT(fp->f_list, fpsrc->f_list, allowaddr);
2353 	fp->f_flag = fpsrc->f_flag;
2354 	fp->f_marker = fpsrc->f_marker;
2355 	fp->f_type = fpsrc->f_type;
2356 	fp->f_advice = fpsrc->f_advice;
2357 	fp->f_count = fpsrc->f_count;
2358 	fp->f_msgcount = fpsrc->f_msgcount;
2359 	fp->f_unpcount = fpsrc->f_unpcount;
2360 	COND_SET_STRUCT(fp->f_unplist, fpsrc->f_unplist, allowaddr);
2361 }
2362 
2363 static void
2364 fill_file2(struct kinfo_file *kp, const file_t *fp, const fdfile_t *ff,
2365 	  int i, pid_t pid)
2366 {
2367 	const bool allowaddr = get_expose_address(curproc);
2368 
2369 	memset(kp, 0, sizeof(*kp));
2370 
2371 	COND_SET_VALUE(kp->ki_fileaddr, PTRTOUINT64(fp), allowaddr);
2372 	kp->ki_flag =		fp->f_flag;
2373 	kp->ki_iflags =		0;
2374 	kp->ki_ftype =		fp->f_type;
2375 	kp->ki_count =		fp->f_count;
2376 	kp->ki_msgcount =	fp->f_msgcount;
2377 	COND_SET_VALUE(kp->ki_fucred, PTRTOUINT64(fp->f_cred), allowaddr);
2378 	kp->ki_fuid =		kauth_cred_geteuid(fp->f_cred);
2379 	kp->ki_fgid =		kauth_cred_getegid(fp->f_cred);
2380 	COND_SET_VALUE(kp->ki_fops, PTRTOUINT64(fp->f_ops), allowaddr);
2381 	kp->ki_foffset =	fp->f_offset;
2382 	COND_SET_VALUE(kp->ki_fdata, PTRTOUINT64(fp->f_data), allowaddr);
2383 
2384 	/* vnode information to glue this file to something */
2385 	if (fp->f_type == DTYPE_VNODE) {
2386 		struct vnode *vp = fp->f_vnode;
2387 
2388 		COND_SET_VALUE(kp->ki_vun, PTRTOUINT64(vp->v_un.vu_socket),
2389 		    allowaddr);
2390 		kp->ki_vsize =	vp->v_size;
2391 		kp->ki_vtype =	vp->v_type;
2392 		kp->ki_vtag =	vp->v_tag;
2393 		COND_SET_VALUE(kp->ki_vdata, PTRTOUINT64(vp->v_data),
2394 		    allowaddr);
2395 	}
2396 
2397 	/* process information when retrieved via KERN_FILE_BYPID */
2398 	if (ff != NULL) {
2399 		kp->ki_pid =		pid;
2400 		kp->ki_fd =		i;
2401 		kp->ki_ofileflags =	ff->ff_exclose;
2402 		kp->ki_usecount =	ff->ff_refcnt;
2403 	}
2404 }
2405