xref: /netbsd/sys/kern/kern_descrip.c (revision 1a58fce8)
1 /*	$NetBSD: kern_descrip.c,v 1.257 2023/04/22 14:23:59 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 1.257 2023/04/22 14:23:59 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. */
__strong_alias(fd_putvnode,fd_putfile)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
fd_isused(filedesc_t * fdp,unsigned fd)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
fd_checkmaps(filedesc_t * fdp)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
fd_next_zero(filedesc_t * fdp,uint32_t * bitmap,int want,u_int bits)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
fd_last_set(filedesc_t * fd,int last)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
fd_used(filedesc_t * fdp,unsigned fd)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
fd_unused(filedesc_t * fdp,unsigned fd)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 *
fd_getfile(unsigned fd)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
fd_putfile(unsigned fd)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
fd_getvnode(unsigned fd,file_t ** fpp)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
fd_getsock1(unsigned fd,struct socket ** sop,file_t ** fp)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
fd_getsock(unsigned fd,struct socket ** sop)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 *
fd_getfile2(proc_t * p,unsigned fd)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
fd_close(unsigned fd)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_ops->fo_advlock != NULL) {
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)(*fp->f_ops->fo_advlock)(fp, 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
fd_dup(file_t * fp,int minfd,int * newp,bool exclose)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
fd_dup2(file_t * fp,unsigned newfd,int flags)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
closef(file_t * fp)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_ops->fo_advlock == NULL) {
856 		KASSERT((fp->f_flag & FHASLOCK) == 0);
857 	} else if (fp->f_flag & FHASLOCK) {
858 		lf.l_whence = SEEK_SET;
859 		lf.l_start = 0;
860 		lf.l_len = 0;
861 		lf.l_type = F_UNLCK;
862 		(void)(*fp->f_ops->fo_advlock)(fp, fp, F_UNLCK, &lf, F_FLOCK);
863 	}
864 	if (fp->f_ops != NULL) {
865 		error = (*fp->f_ops->fo_close)(fp);
866 	} else {
867 		error = 0;
868 	}
869 	KASSERT(fp->f_count == 0);
870 	KASSERT(fp->f_cred != NULL);
871 	pool_cache_put(file_cache, fp);
872 
873 	return error;
874 }
875 
876 /*
877  * Allocate a file descriptor for the process.
878  */
879 int
fd_alloc(proc_t * p,int want,int * result)880 fd_alloc(proc_t *p, int want, int *result)
881 {
882 	filedesc_t *fdp = p->p_fd;
883 	int i, lim, last, error, hi;
884 	u_int off;
885 	fdtab_t *dt;
886 
887 	KASSERT(p == curproc || p == &proc0);
888 
889 	/*
890 	 * Search for a free descriptor starting at the higher
891 	 * of want or fd_freefile.
892 	 */
893 	mutex_enter(&fdp->fd_lock);
894 	fd_checkmaps(fdp);
895 	dt = fdp->fd_dt;
896 	KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
897 	lim = uimin((int)p->p_rlimit[RLIMIT_NOFILE].rlim_cur, maxfiles);
898 	last = uimin(dt->dt_nfiles, lim);
899 	for (;;) {
900 		if ((i = want) < fdp->fd_freefile)
901 			i = fdp->fd_freefile;
902 		off = i >> NDENTRYSHIFT;
903 		hi = fd_next_zero(fdp, fdp->fd_himap, off,
904 		    (last + NDENTRIES - 1) >> NDENTRYSHIFT);
905 		if (hi == -1)
906 			break;
907 		i = fd_next_zero(fdp, &fdp->fd_lomap[hi],
908 		    hi > off ? 0 : i & NDENTRYMASK, NDENTRIES);
909 		if (i == -1) {
910 			/*
911 			 * Free file descriptor in this block was
912 			 * below want, try again with higher want.
913 			 */
914 			want = (hi + 1) << NDENTRYSHIFT;
915 			continue;
916 		}
917 		i += (hi << NDENTRYSHIFT);
918 		if (i >= last) {
919 			break;
920 		}
921 		if (dt->dt_ff[i] == NULL) {
922 			KASSERT(i >= NDFDFILE);
923 			dt->dt_ff[i] = pool_cache_get(fdfile_cache, PR_WAITOK);
924 		}
925 		KASSERT(dt->dt_ff[i]->ff_file == NULL);
926 		fd_used(fdp, i);
927 		if (want <= fdp->fd_freefile) {
928 			fdp->fd_freefile = i;
929 		}
930 		*result = i;
931 		KASSERT(i >= NDFDFILE ||
932 		    dt->dt_ff[i] == (fdfile_t *)fdp->fd_dfdfile[i]);
933 		fd_checkmaps(fdp);
934 		mutex_exit(&fdp->fd_lock);
935 		return 0;
936 	}
937 
938 	/* No space in current array.  Let the caller expand and retry. */
939 	error = (dt->dt_nfiles >= lim) ? EMFILE : ENOSPC;
940 	mutex_exit(&fdp->fd_lock);
941 	return error;
942 }
943 
944 /*
945  * Allocate memory for a descriptor table.
946  */
947 static fdtab_t *
fd_dtab_alloc(int n)948 fd_dtab_alloc(int n)
949 {
950 	fdtab_t *dt;
951 	size_t sz;
952 
953 	KASSERT(n > NDFILE);
954 
955 	sz = sizeof(*dt) + (n - NDFILE) * sizeof(dt->dt_ff[0]);
956 	dt = kmem_alloc(sz, KM_SLEEP);
957 #ifdef DIAGNOSTIC
958 	memset(dt, 0xff, sz);
959 #endif
960 	dt->dt_nfiles = n;
961 	dt->dt_link = NULL;
962 	return dt;
963 }
964 
965 /*
966  * Free a descriptor table, and all tables linked for deferred free.
967  */
968 static void
fd_dtab_free(fdtab_t * dt)969 fd_dtab_free(fdtab_t *dt)
970 {
971 	fdtab_t *next;
972 	size_t sz;
973 
974 	do {
975 		next = dt->dt_link;
976 		KASSERT(dt->dt_nfiles > NDFILE);
977 		sz = sizeof(*dt) +
978 		    (dt->dt_nfiles - NDFILE) * sizeof(dt->dt_ff[0]);
979 #ifdef DIAGNOSTIC
980 		memset(dt, 0xff, sz);
981 #endif
982 		kmem_free(dt, sz);
983 		dt = next;
984 	} while (dt != NULL);
985 }
986 
987 /*
988  * Allocate descriptor bitmap.
989  */
990 static void
fd_map_alloc(int n,uint32_t ** lo,uint32_t ** hi)991 fd_map_alloc(int n, uint32_t **lo, uint32_t **hi)
992 {
993 	uint8_t *ptr;
994 	size_t szlo, szhi;
995 
996 	KASSERT(n > NDENTRIES);
997 
998 	szlo = NDLOSLOTS(n) * sizeof(uint32_t);
999 	szhi = NDHISLOTS(n) * sizeof(uint32_t);
1000 	ptr = kmem_alloc(szlo + szhi, KM_SLEEP);
1001 	*lo = (uint32_t *)ptr;
1002 	*hi = (uint32_t *)(ptr + szlo);
1003 }
1004 
1005 /*
1006  * Free descriptor bitmap.
1007  */
1008 static void
fd_map_free(int n,uint32_t * lo,uint32_t * hi)1009 fd_map_free(int n, uint32_t *lo, uint32_t *hi)
1010 {
1011 	size_t szlo, szhi;
1012 
1013 	KASSERT(n > NDENTRIES);
1014 
1015 	szlo = NDLOSLOTS(n) * sizeof(uint32_t);
1016 	szhi = NDHISLOTS(n) * sizeof(uint32_t);
1017 	KASSERT(hi == (uint32_t *)((uint8_t *)lo + szlo));
1018 	kmem_free(lo, szlo + szhi);
1019 }
1020 
1021 /*
1022  * Expand a process' descriptor table.
1023  */
1024 void
fd_tryexpand(proc_t * p)1025 fd_tryexpand(proc_t *p)
1026 {
1027 	filedesc_t *fdp;
1028 	int i, numfiles, oldnfiles;
1029 	fdtab_t *newdt, *dt;
1030 	uint32_t *newhimap, *newlomap;
1031 
1032 	KASSERT(p == curproc || p == &proc0);
1033 
1034 	fdp = p->p_fd;
1035 	newhimap = NULL;
1036 	newlomap = NULL;
1037 	oldnfiles = atomic_load_consume(&fdp->fd_dt)->dt_nfiles;
1038 
1039 	if (oldnfiles < NDEXTENT)
1040 		numfiles = NDEXTENT;
1041 	else
1042 		numfiles = 2 * oldnfiles;
1043 
1044 	newdt = fd_dtab_alloc(numfiles);
1045 	if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1046 		fd_map_alloc(numfiles, &newlomap, &newhimap);
1047 	}
1048 
1049 	mutex_enter(&fdp->fd_lock);
1050 	dt = fdp->fd_dt;
1051 	KASSERT(dt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1052 	if (dt->dt_nfiles != oldnfiles) {
1053 		/* fdp changed; caller must retry */
1054 		mutex_exit(&fdp->fd_lock);
1055 		fd_dtab_free(newdt);
1056 		if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1057 			fd_map_free(numfiles, newlomap, newhimap);
1058 		}
1059 		return;
1060 	}
1061 
1062 	/* Copy the existing descriptor table and zero the new portion. */
1063 	i = sizeof(fdfile_t *) * oldnfiles;
1064 	memcpy(newdt->dt_ff, dt->dt_ff, i);
1065 	memset((uint8_t *)newdt->dt_ff + i, 0,
1066 	    numfiles * sizeof(fdfile_t *) - i);
1067 
1068 	/*
1069 	 * Link old descriptor array into list to be discarded.  We defer
1070 	 * freeing until the last reference to the descriptor table goes
1071 	 * away (usually process exit).  This allows us to do lockless
1072 	 * lookups in fd_getfile().
1073 	 */
1074 	if (oldnfiles > NDFILE) {
1075 		if (fdp->fd_refcnt > 1) {
1076 			newdt->dt_link = dt;
1077 		} else {
1078 			fd_dtab_free(dt);
1079 		}
1080 	}
1081 
1082 	if (NDHISLOTS(numfiles) > NDHISLOTS(oldnfiles)) {
1083 		i = NDHISLOTS(oldnfiles) * sizeof(uint32_t);
1084 		memcpy(newhimap, fdp->fd_himap, i);
1085 		memset((uint8_t *)newhimap + i, 0,
1086 		    NDHISLOTS(numfiles) * sizeof(uint32_t) - i);
1087 
1088 		i = NDLOSLOTS(oldnfiles) * sizeof(uint32_t);
1089 		memcpy(newlomap, fdp->fd_lomap, i);
1090 		memset((uint8_t *)newlomap + i, 0,
1091 		    NDLOSLOTS(numfiles) * sizeof(uint32_t) - i);
1092 
1093 		if (NDHISLOTS(oldnfiles) > NDHISLOTS(NDFILE)) {
1094 			fd_map_free(oldnfiles, fdp->fd_lomap, fdp->fd_himap);
1095 		}
1096 		fdp->fd_himap = newhimap;
1097 		fdp->fd_lomap = newlomap;
1098 	}
1099 
1100 	/*
1101 	 * All other modifications must become globally visible before
1102 	 * the change to fd_dt.  See fd_getfile().
1103 	 */
1104 	atomic_store_release(&fdp->fd_dt, newdt);
1105 	KASSERT(newdt->dt_ff[0] == (fdfile_t *)fdp->fd_dfdfile[0]);
1106 	fd_checkmaps(fdp);
1107 	mutex_exit(&fdp->fd_lock);
1108 }
1109 
1110 /*
1111  * Create a new open file structure and allocate a file descriptor
1112  * for the current process.
1113  */
1114 int
fd_allocfile(file_t ** resultfp,int * resultfd)1115 fd_allocfile(file_t **resultfp, int *resultfd)
1116 {
1117 	proc_t *p = curproc;
1118 	kauth_cred_t cred;
1119 	file_t *fp;
1120 	int error;
1121 
1122 	while ((error = fd_alloc(p, 0, resultfd)) != 0) {
1123 		if (error != ENOSPC) {
1124 			return error;
1125 		}
1126 		fd_tryexpand(p);
1127 	}
1128 
1129 	fp = pool_cache_get(file_cache, PR_WAITOK);
1130 	if (fp == NULL) {
1131 		fd_abort(p, NULL, *resultfd);
1132 		return ENFILE;
1133 	}
1134 	KASSERT(fp->f_count == 0);
1135 	KASSERT(fp->f_msgcount == 0);
1136 	KASSERT(fp->f_unpcount == 0);
1137 
1138 	/* Replace cached credentials if not what we need. */
1139 	cred = curlwp->l_cred;
1140 	if (__predict_false(cred != fp->f_cred)) {
1141 		kauth_cred_free(fp->f_cred);
1142 		kauth_cred_hold(cred);
1143 		fp->f_cred = cred;
1144 	}
1145 
1146 	/*
1147 	 * Don't allow recycled files to be scanned.
1148 	 * See uipc_usrreq.c.
1149 	 */
1150 	if (__predict_false((fp->f_flag & FSCAN) != 0)) {
1151 		mutex_enter(&fp->f_lock);
1152 		atomic_and_uint(&fp->f_flag, ~FSCAN);
1153 		mutex_exit(&fp->f_lock);
1154 	}
1155 
1156 	fp->f_advice = 0;
1157 	fp->f_offset = 0;
1158 	*resultfp = fp;
1159 
1160 	return 0;
1161 }
1162 
1163 /*
1164  * Successful creation of a new descriptor: make visible to the process.
1165  */
1166 void
fd_affix(proc_t * p,file_t * fp,unsigned fd)1167 fd_affix(proc_t *p, file_t *fp, unsigned fd)
1168 {
1169 	fdfile_t *ff;
1170 	filedesc_t *fdp;
1171 	fdtab_t *dt;
1172 
1173 	KASSERT(p == curproc || p == &proc0);
1174 
1175 	/* Add a reference to the file structure. */
1176 	mutex_enter(&fp->f_lock);
1177 	fp->f_count++;
1178 	mutex_exit(&fp->f_lock);
1179 
1180 	/*
1181 	 * Insert the new file into the descriptor slot.
1182 	 */
1183 	fdp = p->p_fd;
1184 	dt = atomic_load_consume(&fdp->fd_dt);
1185 	ff = dt->dt_ff[fd];
1186 
1187 	KASSERT(ff != NULL);
1188 	KASSERT(ff->ff_file == NULL);
1189 	KASSERT(ff->ff_allocated);
1190 	KASSERT(fd_isused(fdp, fd));
1191 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1192 
1193 	/* No need to lock in order to make file initially visible. */
1194 	atomic_store_release(&ff->ff_file, fp);
1195 }
1196 
1197 /*
1198  * Abort creation of a new descriptor: free descriptor slot and file.
1199  */
1200 void
fd_abort(proc_t * p,file_t * fp,unsigned fd)1201 fd_abort(proc_t *p, file_t *fp, unsigned fd)
1202 {
1203 	filedesc_t *fdp;
1204 	fdfile_t *ff;
1205 
1206 	KASSERT(p == curproc || p == &proc0);
1207 
1208 	fdp = p->p_fd;
1209 	ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
1210 	ff->ff_exclose = false;
1211 
1212 	KASSERT(fd >= NDFDFILE || ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1213 
1214 	mutex_enter(&fdp->fd_lock);
1215 	KASSERT(fd_isused(fdp, fd));
1216 	fd_unused(fdp, fd);
1217 	mutex_exit(&fdp->fd_lock);
1218 
1219 	if (fp != NULL) {
1220 		KASSERT(fp->f_count == 0);
1221 		KASSERT(fp->f_cred != NULL);
1222 		pool_cache_put(file_cache, fp);
1223 	}
1224 }
1225 
1226 static int
file_ctor(void * arg,void * obj,int flags)1227 file_ctor(void *arg, void *obj, int flags)
1228 {
1229 	file_t *fp = obj;
1230 
1231 	memset(fp, 0, sizeof(*fp));
1232 
1233 	mutex_enter(&filelist_lock);
1234 	if (__predict_false(nfiles >= maxfiles)) {
1235 		mutex_exit(&filelist_lock);
1236 		tablefull("file", "increase kern.maxfiles or MAXFILES");
1237 		return ENFILE;
1238 	}
1239 	nfiles++;
1240 	LIST_INSERT_HEAD(&filehead, fp, f_list);
1241 	mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1242 	fp->f_cred = curlwp->l_cred;
1243 	kauth_cred_hold(fp->f_cred);
1244 	mutex_exit(&filelist_lock);
1245 
1246 	return 0;
1247 }
1248 
1249 static void
file_dtor(void * arg,void * obj)1250 file_dtor(void *arg, void *obj)
1251 {
1252 	file_t *fp = obj;
1253 
1254 	mutex_enter(&filelist_lock);
1255 	nfiles--;
1256 	LIST_REMOVE(fp, f_list);
1257 	mutex_exit(&filelist_lock);
1258 
1259 	KASSERT(fp->f_count == 0);
1260 	kauth_cred_free(fp->f_cred);
1261 	mutex_destroy(&fp->f_lock);
1262 }
1263 
1264 static int
fdfile_ctor(void * arg,void * obj,int flags)1265 fdfile_ctor(void *arg, void *obj, int flags)
1266 {
1267 	fdfile_t *ff = obj;
1268 
1269 	memset(ff, 0, sizeof(*ff));
1270 	cv_init(&ff->ff_closing, "fdclose");
1271 
1272 	return 0;
1273 }
1274 
1275 static void
fdfile_dtor(void * arg,void * obj)1276 fdfile_dtor(void *arg, void *obj)
1277 {
1278 	fdfile_t *ff = obj;
1279 
1280 	cv_destroy(&ff->ff_closing);
1281 }
1282 
1283 file_t *
fgetdummy(void)1284 fgetdummy(void)
1285 {
1286 	file_t *fp;
1287 
1288 	fp = kmem_zalloc(sizeof(*fp), KM_SLEEP);
1289 	mutex_init(&fp->f_lock, MUTEX_DEFAULT, IPL_NONE);
1290 	return fp;
1291 }
1292 
1293 void
fputdummy(file_t * fp)1294 fputdummy(file_t *fp)
1295 {
1296 
1297 	mutex_destroy(&fp->f_lock);
1298 	kmem_free(fp, sizeof(*fp));
1299 }
1300 
1301 /*
1302  * Create an initial filedesc structure.
1303  */
1304 filedesc_t *
fd_init(filedesc_t * fdp)1305 fd_init(filedesc_t *fdp)
1306 {
1307 #ifdef DIAGNOSTIC
1308 	unsigned fd;
1309 #endif
1310 
1311 	if (__predict_true(fdp == NULL)) {
1312 		fdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1313 	} else {
1314 		KASSERT(fdp == &filedesc0);
1315 		filedesc_ctor(NULL, fdp, PR_WAITOK);
1316 	}
1317 
1318 #ifdef DIAGNOSTIC
1319 	KASSERT(fdp->fd_lastfile == -1);
1320 	KASSERT(fdp->fd_lastkqfile == -1);
1321 	KASSERT(fdp->fd_knhash == NULL);
1322 	KASSERT(fdp->fd_freefile == 0);
1323 	KASSERT(fdp->fd_exclose == false);
1324 	KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1325 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1326 	for (fd = 0; fd < NDFDFILE; fd++) {
1327 		KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] ==
1328 		    (fdfile_t *)fdp->fd_dfdfile[fd]);
1329 	}
1330 	for (fd = NDFDFILE; fd < NDFILE; fd++) {
1331 		KASSERT(fdp->fd_dtbuiltin.dt_ff[fd] == NULL);
1332 	}
1333 	KASSERT(fdp->fd_himap == fdp->fd_dhimap);
1334 	KASSERT(fdp->fd_lomap == fdp->fd_dlomap);
1335 #endif	/* DIAGNOSTIC */
1336 
1337 	fdp->fd_refcnt = 1;
1338 	fd_checkmaps(fdp);
1339 
1340 	return fdp;
1341 }
1342 
1343 /*
1344  * Initialize a file descriptor table.
1345  */
1346 static int
filedesc_ctor(void * arg,void * obj,int flag)1347 filedesc_ctor(void *arg, void *obj, int flag)
1348 {
1349 	filedesc_t *fdp = obj;
1350 	fdfile_t **ffp;
1351 	int i;
1352 
1353 	memset(fdp, 0, sizeof(*fdp));
1354 	mutex_init(&fdp->fd_lock, MUTEX_DEFAULT, IPL_NONE);
1355 	fdp->fd_lastfile = -1;
1356 	fdp->fd_lastkqfile = -1;
1357 	fdp->fd_dt = &fdp->fd_dtbuiltin;
1358 	fdp->fd_dtbuiltin.dt_nfiles = NDFILE;
1359 	fdp->fd_himap = fdp->fd_dhimap;
1360 	fdp->fd_lomap = fdp->fd_dlomap;
1361 
1362 	CTASSERT(sizeof(fdp->fd_dfdfile[0]) >= sizeof(fdfile_t));
1363 	for (i = 0, ffp = fdp->fd_dt->dt_ff; i < NDFDFILE; i++, ffp++) {
1364 		*ffp = (fdfile_t *)fdp->fd_dfdfile[i];
1365 		(void)fdfile_ctor(NULL, fdp->fd_dfdfile[i], PR_WAITOK);
1366 	}
1367 
1368 	return 0;
1369 }
1370 
1371 static void
filedesc_dtor(void * arg,void * obj)1372 filedesc_dtor(void *arg, void *obj)
1373 {
1374 	filedesc_t *fdp = obj;
1375 	int i;
1376 
1377 	for (i = 0; i < NDFDFILE; i++) {
1378 		fdfile_dtor(NULL, fdp->fd_dfdfile[i]);
1379 	}
1380 
1381 	mutex_destroy(&fdp->fd_lock);
1382 }
1383 
1384 /*
1385  * Make p share curproc's filedesc structure.
1386  */
1387 void
fd_share(struct proc * p)1388 fd_share(struct proc *p)
1389 {
1390 	filedesc_t *fdp;
1391 
1392 	fdp = curlwp->l_fd;
1393 	p->p_fd = fdp;
1394 	atomic_inc_uint(&fdp->fd_refcnt);
1395 }
1396 
1397 /*
1398  * Acquire a hold on a filedesc structure.
1399  */
1400 void
fd_hold(lwp_t * l)1401 fd_hold(lwp_t *l)
1402 {
1403 	filedesc_t *fdp = l->l_fd;
1404 
1405 	atomic_inc_uint(&fdp->fd_refcnt);
1406 }
1407 
1408 /*
1409  * Copy a filedesc structure.
1410  */
1411 filedesc_t *
fd_copy(void)1412 fd_copy(void)
1413 {
1414 	filedesc_t *newfdp, *fdp;
1415 	fdfile_t *ff, **ffp, **nffp, *ff2;
1416 	int i, j, numfiles, lastfile, newlast;
1417 	file_t *fp;
1418 	fdtab_t *newdt;
1419 
1420 	fdp = curproc->p_fd;
1421 	newfdp = pool_cache_get(filedesc_cache, PR_WAITOK);
1422 	newfdp->fd_refcnt = 1;
1423 
1424 #ifdef DIAGNOSTIC
1425 	KASSERT(newfdp->fd_lastfile == -1);
1426 	KASSERT(newfdp->fd_lastkqfile == -1);
1427 	KASSERT(newfdp->fd_knhash == NULL);
1428 	KASSERT(newfdp->fd_freefile == 0);
1429 	KASSERT(newfdp->fd_exclose == false);
1430 	KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1431 	KASSERT(newfdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1432 	for (i = 0; i < NDFDFILE; i++) {
1433 		KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] ==
1434 		    (fdfile_t *)&newfdp->fd_dfdfile[i]);
1435 	}
1436 	for (i = NDFDFILE; i < NDFILE; i++) {
1437 		KASSERT(newfdp->fd_dtbuiltin.dt_ff[i] == NULL);
1438 	}
1439 #endif	/* DIAGNOSTIC */
1440 
1441 	mutex_enter(&fdp->fd_lock);
1442 	fd_checkmaps(fdp);
1443 	numfiles = fdp->fd_dt->dt_nfiles;
1444 	lastfile = fdp->fd_lastfile;
1445 
1446 	/*
1447 	 * If the number of open files fits in the internal arrays
1448 	 * of the open file structure, use them, otherwise allocate
1449 	 * additional memory for the number of descriptors currently
1450 	 * in use.
1451 	 */
1452 	if (lastfile < NDFILE) {
1453 		i = NDFILE;
1454 		newdt = newfdp->fd_dt;
1455 		KASSERT(newfdp->fd_dt == &newfdp->fd_dtbuiltin);
1456 	} else {
1457 		/*
1458 		 * Compute the smallest multiple of NDEXTENT needed
1459 		 * for the file descriptors currently in use,
1460 		 * allowing the table to shrink.
1461 		 */
1462 		i = numfiles;
1463 		while (i >= 2 * NDEXTENT && i > lastfile * 2) {
1464 			i /= 2;
1465 		}
1466 		KASSERT(i > NDFILE);
1467 		newdt = fd_dtab_alloc(i);
1468 		newfdp->fd_dt = newdt;
1469 		memcpy(newdt->dt_ff, newfdp->fd_dtbuiltin.dt_ff,
1470 		    NDFDFILE * sizeof(fdfile_t **));
1471 		memset(newdt->dt_ff + NDFDFILE, 0,
1472 		    (i - NDFDFILE) * sizeof(fdfile_t **));
1473 	}
1474 	if (NDHISLOTS(i) <= NDHISLOTS(NDFILE)) {
1475 		newfdp->fd_himap = newfdp->fd_dhimap;
1476 		newfdp->fd_lomap = newfdp->fd_dlomap;
1477 	} else {
1478 		fd_map_alloc(i, &newfdp->fd_lomap, &newfdp->fd_himap);
1479 		KASSERT(i >= NDENTRIES * NDENTRIES);
1480 		memset(newfdp->fd_himap, 0, NDHISLOTS(i)*sizeof(uint32_t));
1481 		memset(newfdp->fd_lomap, 0, NDLOSLOTS(i)*sizeof(uint32_t));
1482 	}
1483 	newfdp->fd_freefile = fdp->fd_freefile;
1484 	newfdp->fd_exclose = fdp->fd_exclose;
1485 
1486 	ffp = fdp->fd_dt->dt_ff;
1487 	nffp = newdt->dt_ff;
1488 	newlast = -1;
1489 	for (i = 0; i <= lastfile; i++, ffp++, nffp++) {
1490 		KASSERT(i >= NDFDFILE ||
1491 		    *nffp == (fdfile_t *)newfdp->fd_dfdfile[i]);
1492 		ff = *ffp;
1493 		if (ff == NULL ||
1494 		    (fp = atomic_load_consume(&ff->ff_file)) == NULL) {
1495 			/* Descriptor unused, or descriptor half open. */
1496 			KASSERT(!fd_isused(newfdp, i));
1497 			continue;
1498 		}
1499 		if (__predict_false(fp->f_type == DTYPE_KQUEUE)) {
1500 			/* kqueue descriptors cannot be copied. */
1501 			if (i < newfdp->fd_freefile) {
1502 				newfdp->fd_freefile = i;
1503 			}
1504 			continue;
1505 		}
1506 		/* It's active: add a reference to the file. */
1507 		mutex_enter(&fp->f_lock);
1508 		fp->f_count++;
1509 		mutex_exit(&fp->f_lock);
1510 
1511 		/* Allocate an fdfile_t to represent it. */
1512 		if (i >= NDFDFILE) {
1513 			ff2 = pool_cache_get(fdfile_cache, PR_WAITOK);
1514 			*nffp = ff2;
1515 		} else {
1516 			ff2 = newdt->dt_ff[i];
1517 		}
1518 		ff2->ff_file = fp;
1519 		ff2->ff_exclose = ff->ff_exclose;
1520 		ff2->ff_allocated = true;
1521 
1522 		/* Fix up bitmaps. */
1523 		j = i >> NDENTRYSHIFT;
1524 		KASSERT((newfdp->fd_lomap[j] & (1U << (i & NDENTRYMASK))) == 0);
1525 		newfdp->fd_lomap[j] |= 1U << (i & NDENTRYMASK);
1526 		if (__predict_false(newfdp->fd_lomap[j] == ~0)) {
1527 			KASSERT((newfdp->fd_himap[j >> NDENTRYSHIFT] &
1528 			    (1U << (j & NDENTRYMASK))) == 0);
1529 			newfdp->fd_himap[j >> NDENTRYSHIFT] |=
1530 			    1U << (j & NDENTRYMASK);
1531 		}
1532 		newlast = i;
1533 	}
1534 	KASSERT(newdt->dt_ff[0] == (fdfile_t *)newfdp->fd_dfdfile[0]);
1535 	newfdp->fd_lastfile = newlast;
1536 	fd_checkmaps(newfdp);
1537 	mutex_exit(&fdp->fd_lock);
1538 
1539 	return newfdp;
1540 }
1541 
1542 /*
1543  * Release a filedesc structure.
1544  */
1545 void
fd_free(void)1546 fd_free(void)
1547 {
1548 	fdfile_t *ff;
1549 	file_t *fp;
1550 	int fd, nf;
1551 	fdtab_t *dt;
1552 	lwp_t * const l = curlwp;
1553 	filedesc_t * const fdp = l->l_fd;
1554 	const bool noadvlock = (l->l_proc->p_flag & PK_ADVLOCK) == 0;
1555 
1556 	KASSERT(atomic_load_consume(&fdp->fd_dt)->dt_ff[0] ==
1557 	    (fdfile_t *)fdp->fd_dfdfile[0]);
1558 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1559 	KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1560 
1561 	membar_release();
1562 	if (atomic_dec_uint_nv(&fdp->fd_refcnt) > 0)
1563 		return;
1564 	membar_acquire();
1565 
1566 	/*
1567 	 * Close any files that the process holds open.
1568 	 */
1569 	dt = fdp->fd_dt;
1570 	fd_checkmaps(fdp);
1571 #ifdef DEBUG
1572 	fdp->fd_refcnt = -1; /* see fd_checkmaps */
1573 #endif
1574 	for (fd = 0, nf = dt->dt_nfiles; fd < nf; fd++) {
1575 		ff = dt->dt_ff[fd];
1576 		KASSERT(fd >= NDFDFILE ||
1577 		    ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1578 		if (ff == NULL)
1579 			continue;
1580 		if ((fp = atomic_load_consume(&ff->ff_file)) != NULL) {
1581 			/*
1582 			 * Must use fd_close() here if there is
1583 			 * a reference from kqueue or we might have posix
1584 			 * advisory locks.
1585 			 */
1586 			if (__predict_true(ff->ff_refcnt == 0) &&
1587 			    (noadvlock || fp->f_type != DTYPE_VNODE)) {
1588 				ff->ff_file = NULL;
1589 				ff->ff_exclose = false;
1590 				ff->ff_allocated = false;
1591 				closef(fp);
1592 			} else {
1593 				ff->ff_refcnt++;
1594 				fd_close(fd);
1595 			}
1596 		}
1597 		KASSERT(ff->ff_refcnt == 0);
1598 		KASSERT(ff->ff_file == NULL);
1599 		KASSERT(!ff->ff_exclose);
1600 		KASSERT(!ff->ff_allocated);
1601 		if (fd >= NDFDFILE) {
1602 			pool_cache_put(fdfile_cache, ff);
1603 			dt->dt_ff[fd] = NULL;
1604 		}
1605 	}
1606 
1607 	/*
1608 	 * Clean out the descriptor table for the next user and return
1609 	 * to the cache.
1610 	 */
1611 	if (__predict_false(dt != &fdp->fd_dtbuiltin)) {
1612 		fd_dtab_free(fdp->fd_dt);
1613 		/* Otherwise, done above. */
1614 		memset(&fdp->fd_dtbuiltin.dt_ff[NDFDFILE], 0,
1615 		    (NDFILE - NDFDFILE) * sizeof(fdp->fd_dtbuiltin.dt_ff[0]));
1616 		fdp->fd_dt = &fdp->fd_dtbuiltin;
1617 	}
1618 	if (__predict_false(NDHISLOTS(nf) > NDHISLOTS(NDFILE))) {
1619 		KASSERT(fdp->fd_himap != fdp->fd_dhimap);
1620 		KASSERT(fdp->fd_lomap != fdp->fd_dlomap);
1621 		fd_map_free(nf, fdp->fd_lomap, fdp->fd_himap);
1622 	}
1623 	if (__predict_false(fdp->fd_knhash != NULL)) {
1624 		hashdone(fdp->fd_knhash, HASH_LIST, fdp->fd_knhashmask);
1625 		fdp->fd_knhash = NULL;
1626 		fdp->fd_knhashmask = 0;
1627 	} else {
1628 		KASSERT(fdp->fd_knhashmask == 0);
1629 	}
1630 	fdp->fd_dt = &fdp->fd_dtbuiltin;
1631 	fdp->fd_lastkqfile = -1;
1632 	fdp->fd_lastfile = -1;
1633 	fdp->fd_freefile = 0;
1634 	fdp->fd_exclose = false;
1635 	memset(&fdp->fd_startzero, 0, sizeof(*fdp) -
1636 	    offsetof(filedesc_t, fd_startzero));
1637 	fdp->fd_himap = fdp->fd_dhimap;
1638 	fdp->fd_lomap = fdp->fd_dlomap;
1639 	KASSERT(fdp->fd_dtbuiltin.dt_nfiles == NDFILE);
1640 	KASSERT(fdp->fd_dtbuiltin.dt_link == NULL);
1641 	KASSERT(fdp->fd_dt == &fdp->fd_dtbuiltin);
1642 #ifdef DEBUG
1643 	fdp->fd_refcnt = 0; /* see fd_checkmaps */
1644 #endif
1645 	fd_checkmaps(fdp);
1646 	pool_cache_put(filedesc_cache, fdp);
1647 }
1648 
1649 /*
1650  * File Descriptor pseudo-device driver (/dev/fd/).
1651  *
1652  * Opening minor device N dup()s the file (if any) connected to file
1653  * descriptor N belonging to the calling process.  Note that this driver
1654  * consists of only the ``open()'' routine, because all subsequent
1655  * references to this file will be direct to the other driver.
1656  */
1657 static int
filedescopen(dev_t dev,int mode,int type,lwp_t * l)1658 filedescopen(dev_t dev, int mode, int type, lwp_t *l)
1659 {
1660 
1661 	/*
1662 	 * XXX Kludge: set dupfd to contain the value of the
1663 	 * the file descriptor being sought for duplication. The error
1664 	 * return ensures that the vnode for this device will be released
1665 	 * by vn_open. Open will detect this special error and take the
1666 	 * actions in fd_dupopen below. Other callers of vn_open or VOP_OPEN
1667 	 * will simply report the error.
1668 	 */
1669 	l->l_dupfd = minor(dev);	/* XXX */
1670 	return EDUPFD;
1671 }
1672 
1673 /*
1674  * Duplicate the specified descriptor to a free descriptor.
1675  *
1676  * old is the original fd.
1677  * moveit is true if we should move rather than duplicate.
1678  * flags are the open flags (converted from O_* to F*).
1679  * newp returns the new fd on success.
1680  *
1681  * These two cases are produced by the EDUPFD and EMOVEFD magic
1682  * errnos, but in the interest of removing that regrettable interface,
1683  * vn_open has been changed to intercept them. Now vn_open returns
1684  * either a vnode or a filehandle, and the filehandle is accompanied
1685  * by a boolean that says whether we should dup (moveit == false) or
1686  * move (moveit == true) the fd.
1687  *
1688  * The dup case is used by /dev/stderr, /proc/self/fd, and such. The
1689  * move case is used by cloner devices that allocate a fd of their
1690  * own (a layering violation that should go away eventually) that
1691  * then needs to be put in the place open() expects it.
1692  */
1693 int
fd_dupopen(int old,bool moveit,int flags,int * newp)1694 fd_dupopen(int old, bool moveit, int flags, int *newp)
1695 {
1696 	filedesc_t *fdp;
1697 	fdfile_t *ff;
1698 	file_t *fp;
1699 	fdtab_t *dt;
1700 	int error;
1701 
1702 	if ((fp = fd_getfile(old)) == NULL) {
1703 		return EBADF;
1704 	}
1705 	fdp = curlwp->l_fd;
1706 	dt = atomic_load_consume(&fdp->fd_dt);
1707 	ff = dt->dt_ff[old];
1708 
1709 	/*
1710 	 * There are two cases of interest here.
1711 	 *
1712 	 * 1. moveit == false (used to be the EDUPFD magic errno):
1713 	 *    simply dup (old) to file descriptor (new) and return.
1714 	 *
1715 	 * 2. moveit == true (used to be the EMOVEFD magic errno):
1716 	 *    steal away the file structure from (old) and store it in
1717 	 *    (new).  (old) is effectively closed by this operation.
1718 	 */
1719 	if (moveit == false) {
1720 		/*
1721 		 * Check that the mode the file is being opened for is a
1722 		 * subset of the mode of the existing descriptor.
1723 		 */
1724 		if (((flags & (FREAD|FWRITE)) | fp->f_flag) != fp->f_flag) {
1725 			error = EACCES;
1726 			goto out;
1727 		}
1728 
1729 		/* Copy it. */
1730 		error = fd_dup(fp, 0, newp, ff->ff_exclose);
1731 	} else {
1732 		/* Copy it. */
1733 		error = fd_dup(fp, 0, newp, ff->ff_exclose);
1734 		if (error != 0) {
1735 			goto out;
1736 		}
1737 
1738 		/* Steal away the file pointer from 'old'. */
1739 		(void)fd_close(old);
1740 		return 0;
1741 	}
1742 
1743 out:
1744 	fd_putfile(old);
1745 	return error;
1746 }
1747 
1748 /*
1749  * Close open files on exec.
1750  */
1751 void
fd_closeexec(void)1752 fd_closeexec(void)
1753 {
1754 	proc_t *p;
1755 	filedesc_t *fdp;
1756 	fdfile_t *ff;
1757 	lwp_t *l;
1758 	fdtab_t *dt;
1759 	int fd;
1760 
1761 	l = curlwp;
1762 	p = l->l_proc;
1763 	fdp = p->p_fd;
1764 
1765 	if (fdp->fd_refcnt > 1) {
1766 		fdp = fd_copy();
1767 		fd_free();
1768 		p->p_fd = fdp;
1769 		l->l_fd = fdp;
1770 	}
1771 	if (!fdp->fd_exclose) {
1772 		return;
1773 	}
1774 	fdp->fd_exclose = false;
1775 	dt = atomic_load_consume(&fdp->fd_dt);
1776 
1777 	for (fd = 0; fd <= fdp->fd_lastfile; fd++) {
1778 		if ((ff = dt->dt_ff[fd]) == NULL) {
1779 			KASSERT(fd >= NDFDFILE);
1780 			continue;
1781 		}
1782 		KASSERT(fd >= NDFDFILE ||
1783 		    ff == (fdfile_t *)fdp->fd_dfdfile[fd]);
1784 		if (ff->ff_file == NULL)
1785 			continue;
1786 		if (ff->ff_exclose) {
1787 			/*
1788 			 * We need a reference to close the file.
1789 			 * No other threads can see the fdfile_t at
1790 			 * this point, so don't bother locking.
1791 			 */
1792 			KASSERT((ff->ff_refcnt & FR_CLOSING) == 0);
1793 			ff->ff_refcnt++;
1794 			fd_close(fd);
1795 		}
1796 	}
1797 }
1798 
1799 /*
1800  * Sets descriptor owner. If the owner is a process, 'pgid'
1801  * is set to positive value, process ID. If the owner is process group,
1802  * 'pgid' is set to -pg_id.
1803  */
1804 int
fsetown(pid_t * pgid,u_long cmd,const void * data)1805 fsetown(pid_t *pgid, u_long cmd, const void *data)
1806 {
1807 	pid_t id = *(const pid_t *)data;
1808 	int error;
1809 
1810 	if (id == INT_MIN)
1811 		return EINVAL;
1812 
1813 	switch (cmd) {
1814 	case TIOCSPGRP:
1815 		if (id < 0)
1816 			return EINVAL;
1817 		id = -id;
1818 		break;
1819 	default:
1820 		break;
1821 	}
1822 	if (id > 0) {
1823 		mutex_enter(&proc_lock);
1824 		error = proc_find(id) ? 0 : ESRCH;
1825 		mutex_exit(&proc_lock);
1826 	} else if (id < 0) {
1827 		error = pgid_in_session(curproc, -id);
1828 	} else {
1829 		error = 0;
1830 	}
1831 	if (!error) {
1832 		*pgid = id;
1833 	}
1834 	return error;
1835 }
1836 
1837 void
fd_set_exclose(struct lwp * l,int fd,bool exclose)1838 fd_set_exclose(struct lwp *l, int fd, bool exclose)
1839 {
1840 	filedesc_t *fdp = l->l_fd;
1841 	fdfile_t *ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
1842 
1843 	ff->ff_exclose = exclose;
1844 	if (exclose)
1845 		fdp->fd_exclose = true;
1846 }
1847 
1848 /*
1849  * Return descriptor owner information. If the value is positive,
1850  * it's process ID. If it's negative, it's process group ID and
1851  * needs the sign removed before use.
1852  */
1853 int
fgetown(pid_t pgid,u_long cmd,void * data)1854 fgetown(pid_t pgid, u_long cmd, void *data)
1855 {
1856 
1857 	switch (cmd) {
1858 	case TIOCGPGRP:
1859 		*(int *)data = -pgid;
1860 		break;
1861 	default:
1862 		*(int *)data = pgid;
1863 		break;
1864 	}
1865 	return 0;
1866 }
1867 
1868 /*
1869  * Send signal to descriptor owner, either process or process group.
1870  */
1871 void
fownsignal(pid_t pgid,int signo,int code,int band,void * fdescdata)1872 fownsignal(pid_t pgid, int signo, int code, int band, void *fdescdata)
1873 {
1874 	ksiginfo_t ksi;
1875 
1876 	KASSERT(!cpu_intr_p());
1877 
1878 	if (pgid == 0) {
1879 		return;
1880 	}
1881 
1882 	KSI_INIT(&ksi);
1883 	ksi.ksi_signo = signo;
1884 	ksi.ksi_code = code;
1885 	ksi.ksi_band = band;
1886 
1887 	mutex_enter(&proc_lock);
1888 	if (pgid > 0) {
1889 		struct proc *p1;
1890 
1891 		p1 = proc_find(pgid);
1892 		if (p1 != NULL) {
1893 			kpsignal(p1, &ksi, fdescdata);
1894 		}
1895 	} else {
1896 		struct pgrp *pgrp;
1897 
1898 		KASSERT(pgid < 0);
1899 		pgrp = pgrp_find(-pgid);
1900 		if (pgrp != NULL) {
1901 			kpgsignal(pgrp, &ksi, fdescdata, 0);
1902 		}
1903 	}
1904 	mutex_exit(&proc_lock);
1905 }
1906 
1907 int
fd_clone(file_t * fp,unsigned fd,int flag,const struct fileops * fops,void * data)1908 fd_clone(file_t *fp, unsigned fd, int flag, const struct fileops *fops,
1909 	 void *data)
1910 {
1911 	fdfile_t *ff;
1912 	filedesc_t *fdp;
1913 
1914 	fp->f_flag = flag & FMASK;
1915 	fdp = curproc->p_fd;
1916 	ff = atomic_load_consume(&fdp->fd_dt)->dt_ff[fd];
1917 	KASSERT(ff != NULL);
1918 	ff->ff_exclose = (flag & O_CLOEXEC) != 0;
1919 	fp->f_type = DTYPE_MISC;
1920 	fp->f_ops = fops;
1921 	fp->f_data = data;
1922 	curlwp->l_dupfd = fd;
1923 	fd_affix(curproc, fp, fd);
1924 
1925 	return EMOVEFD;
1926 }
1927 
1928 int
fnullop_fcntl(file_t * fp,u_int cmd,void * data)1929 fnullop_fcntl(file_t *fp, u_int cmd, void *data)
1930 {
1931 
1932 	if (cmd == F_SETFL)
1933 		return 0;
1934 
1935 	return EOPNOTSUPP;
1936 }
1937 
1938 int
fnullop_poll(file_t * fp,int which)1939 fnullop_poll(file_t *fp, int which)
1940 {
1941 
1942 	return 0;
1943 }
1944 
1945 int
fnullop_kqfilter(file_t * fp,struct knote * kn)1946 fnullop_kqfilter(file_t *fp, struct knote *kn)
1947 {
1948 
1949 	return EOPNOTSUPP;
1950 }
1951 
1952 void
fnullop_restart(file_t * fp)1953 fnullop_restart(file_t *fp)
1954 {
1955 
1956 }
1957 
1958 int
fbadop_read(file_t * fp,off_t * offset,struct uio * uio,kauth_cred_t cred,int flags)1959 fbadop_read(file_t *fp, off_t *offset, struct uio *uio,
1960 	    kauth_cred_t cred, int flags)
1961 {
1962 
1963 	return EOPNOTSUPP;
1964 }
1965 
1966 int
fbadop_write(file_t * fp,off_t * offset,struct uio * uio,kauth_cred_t cred,int flags)1967 fbadop_write(file_t *fp, off_t *offset, struct uio *uio,
1968 	     kauth_cred_t cred, int flags)
1969 {
1970 
1971 	return EOPNOTSUPP;
1972 }
1973 
1974 int
fbadop_ioctl(file_t * fp,u_long com,void * data)1975 fbadop_ioctl(file_t *fp, u_long com, void *data)
1976 {
1977 
1978 	return EOPNOTSUPP;
1979 }
1980 
1981 int
fbadop_stat(file_t * fp,struct stat * sb)1982 fbadop_stat(file_t *fp, struct stat *sb)
1983 {
1984 
1985 	return EOPNOTSUPP;
1986 }
1987 
1988 int
fbadop_close(file_t * fp)1989 fbadop_close(file_t *fp)
1990 {
1991 
1992 	return EOPNOTSUPP;
1993 }
1994 
1995 /*
1996  * sysctl routines pertaining to file descriptors
1997  */
1998 
1999 /* Initialized in sysctl_init() for now... */
2000 extern kmutex_t sysctl_file_marker_lock;
2001 static u_int sysctl_file_marker = 1;
2002 
2003 /*
2004  * Expects to be called with proc_lock and sysctl_file_marker_lock locked.
2005  */
2006 static void
sysctl_file_marker_reset(void)2007 sysctl_file_marker_reset(void)
2008 {
2009 	struct proc *p;
2010 
2011 	PROCLIST_FOREACH(p, &allproc) {
2012 		struct filedesc *fd = p->p_fd;
2013 		fdtab_t *dt;
2014 		u_int i;
2015 
2016 		mutex_enter(&fd->fd_lock);
2017 		dt = fd->fd_dt;
2018 		for (i = 0; i < dt->dt_nfiles; i++) {
2019 			struct file *fp;
2020 			fdfile_t *ff;
2021 
2022 			if ((ff = dt->dt_ff[i]) == NULL) {
2023 				continue;
2024 			}
2025 			if ((fp = atomic_load_consume(&ff->ff_file)) == NULL) {
2026 				continue;
2027 			}
2028 			fp->f_marker = 0;
2029 		}
2030 		mutex_exit(&fd->fd_lock);
2031 	}
2032 }
2033 
2034 /*
2035  * sysctl helper routine for kern.file pseudo-subtree.
2036  */
2037 static int
sysctl_kern_file(SYSCTLFN_ARGS)2038 sysctl_kern_file(SYSCTLFN_ARGS)
2039 {
2040 	const bool allowaddr = get_expose_address(curproc);
2041 	struct filelist flist;
2042 	int error;
2043 	size_t buflen;
2044 	struct file *fp, fbuf;
2045 	char *start, *where;
2046 	struct proc *p;
2047 
2048 	start = where = oldp;
2049 	buflen = *oldlenp;
2050 
2051 	if (where == NULL) {
2052 		/*
2053 		 * overestimate by 10 files
2054 		 */
2055 		*oldlenp = sizeof(filehead) + (nfiles + 10) *
2056 		    sizeof(struct file);
2057 		return 0;
2058 	}
2059 
2060 	/*
2061 	 * first sysctl_copyout filehead
2062 	 */
2063 	if (buflen < sizeof(filehead)) {
2064 		*oldlenp = 0;
2065 		return 0;
2066 	}
2067 	sysctl_unlock();
2068 	if (allowaddr) {
2069 		memcpy(&flist, &filehead, sizeof(flist));
2070 	} else {
2071 		memset(&flist, 0, sizeof(flist));
2072 	}
2073 	error = sysctl_copyout(l, &flist, where, sizeof(flist));
2074 	if (error) {
2075 		sysctl_relock();
2076 		return error;
2077 	}
2078 	buflen -= sizeof(flist);
2079 	where += sizeof(flist);
2080 
2081 	/*
2082 	 * followed by an array of file structures
2083 	 */
2084 	mutex_enter(&sysctl_file_marker_lock);
2085 	mutex_enter(&proc_lock);
2086 	PROCLIST_FOREACH(p, &allproc) {
2087 		struct filedesc *fd;
2088 		fdtab_t *dt;
2089 		u_int i;
2090 
2091 		if (p->p_stat == SIDL) {
2092 			/* skip embryonic processes */
2093 			continue;
2094 		}
2095 		mutex_enter(p->p_lock);
2096 		error = kauth_authorize_process(l->l_cred,
2097 		    KAUTH_PROCESS_CANSEE, p,
2098 		    KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2099 		    NULL, NULL);
2100 		mutex_exit(p->p_lock);
2101 		if (error != 0) {
2102 			/*
2103 			 * Don't leak kauth retval if we're silently
2104 			 * skipping this entry.
2105 			 */
2106 			error = 0;
2107 			continue;
2108 		}
2109 
2110 		/*
2111 		 * Grab a hold on the process.
2112 		 */
2113 		if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2114 			continue;
2115 		}
2116 		mutex_exit(&proc_lock);
2117 
2118 		fd = p->p_fd;
2119 		mutex_enter(&fd->fd_lock);
2120 		dt = fd->fd_dt;
2121 		for (i = 0; i < dt->dt_nfiles; i++) {
2122 			fdfile_t *ff;
2123 
2124 			if ((ff = dt->dt_ff[i]) == NULL) {
2125 				continue;
2126 			}
2127 			if ((fp = atomic_load_consume(&ff->ff_file)) == NULL) {
2128 				continue;
2129 			}
2130 
2131 			mutex_enter(&fp->f_lock);
2132 
2133 			if ((fp->f_count == 0) ||
2134 			    (fp->f_marker == sysctl_file_marker)) {
2135 				mutex_exit(&fp->f_lock);
2136 				continue;
2137 			}
2138 
2139 			/* Check that we have enough space. */
2140 			if (buflen < sizeof(struct file)) {
2141 				*oldlenp = where - start;
2142 				mutex_exit(&fp->f_lock);
2143 				error = ENOMEM;
2144 				break;
2145 			}
2146 
2147 			fill_file(&fbuf, fp);
2148 			mutex_exit(&fp->f_lock);
2149 			error = sysctl_copyout(l, &fbuf, where, sizeof(fbuf));
2150 			if (error) {
2151 				break;
2152 			}
2153 			buflen -= sizeof(struct file);
2154 			where += sizeof(struct file);
2155 
2156 			fp->f_marker = sysctl_file_marker;
2157 		}
2158 		mutex_exit(&fd->fd_lock);
2159 
2160 		/*
2161 		 * Release reference to process.
2162 		 */
2163 		mutex_enter(&proc_lock);
2164 		rw_exit(&p->p_reflock);
2165 
2166 		if (error)
2167 			break;
2168 	}
2169 
2170 	sysctl_file_marker++;
2171 	/* Reset all markers if wrapped. */
2172 	if (sysctl_file_marker == 0) {
2173 		sysctl_file_marker_reset();
2174 		sysctl_file_marker++;
2175 	}
2176 
2177 	mutex_exit(&proc_lock);
2178 	mutex_exit(&sysctl_file_marker_lock);
2179 
2180 	*oldlenp = where - start;
2181 	sysctl_relock();
2182 	return error;
2183 }
2184 
2185 /*
2186  * sysctl helper function for kern.file2
2187  */
2188 static int
sysctl_kern_file2(SYSCTLFN_ARGS)2189 sysctl_kern_file2(SYSCTLFN_ARGS)
2190 {
2191 	struct proc *p;
2192 	struct file *fp;
2193 	struct filedesc *fd;
2194 	struct kinfo_file kf;
2195 	char *dp;
2196 	u_int i, op;
2197 	size_t len, needed, elem_size, out_size;
2198 	int error, arg, elem_count;
2199 	fdfile_t *ff;
2200 	fdtab_t *dt;
2201 
2202 	if (namelen == 1 && name[0] == CTL_QUERY)
2203 		return sysctl_query(SYSCTLFN_CALL(rnode));
2204 
2205 	if (namelen != 4)
2206 		return EINVAL;
2207 
2208 	error = 0;
2209 	dp = oldp;
2210 	len = (oldp != NULL) ? *oldlenp : 0;
2211 	op = name[0];
2212 	arg = name[1];
2213 	elem_size = name[2];
2214 	elem_count = name[3];
2215 	out_size = MIN(sizeof(kf), elem_size);
2216 	needed = 0;
2217 
2218 	if (elem_size < 1 || elem_count < 0)
2219 		return EINVAL;
2220 
2221 	switch (op) {
2222 	case KERN_FILE_BYFILE:
2223 	case KERN_FILE_BYPID:
2224 		/*
2225 		 * We're traversing the process list in both cases; the BYFILE
2226 		 * case does additional work of keeping track of files already
2227 		 * looked at.
2228 		 */
2229 
2230 		/* doesn't use arg so it must be zero */
2231 		if ((op == KERN_FILE_BYFILE) && (arg != 0))
2232 			return EINVAL;
2233 
2234 		if ((op == KERN_FILE_BYPID) && (arg < -1))
2235 			/* -1 means all processes */
2236 			return EINVAL;
2237 
2238 		sysctl_unlock();
2239 		if (op == KERN_FILE_BYFILE)
2240 			mutex_enter(&sysctl_file_marker_lock);
2241 		mutex_enter(&proc_lock);
2242 		PROCLIST_FOREACH(p, &allproc) {
2243 			if (p->p_stat == SIDL) {
2244 				/* skip embryonic processes */
2245 				continue;
2246 			}
2247 			if (arg > 0 && p->p_pid != arg) {
2248 				/* pick only the one we want */
2249 				/* XXX want 0 to mean "kernel files" */
2250 				continue;
2251 			}
2252 			mutex_enter(p->p_lock);
2253 			error = kauth_authorize_process(l->l_cred,
2254 			    KAUTH_PROCESS_CANSEE, p,
2255 			    KAUTH_ARG(KAUTH_REQ_PROCESS_CANSEE_OPENFILES),
2256 			    NULL, NULL);
2257 			mutex_exit(p->p_lock);
2258 			if (error != 0) {
2259 				/*
2260 				 * Don't leak kauth retval if we're silently
2261 				 * skipping this entry.
2262 				 */
2263 				error = 0;
2264 				continue;
2265 			}
2266 
2267 			/*
2268 			 * Grab a hold on the process.
2269 			 */
2270 			if (!rw_tryenter(&p->p_reflock, RW_READER)) {
2271 				continue;
2272 			}
2273 			mutex_exit(&proc_lock);
2274 
2275 			fd = p->p_fd;
2276 			mutex_enter(&fd->fd_lock);
2277 			dt = fd->fd_dt;
2278 			for (i = 0; i < dt->dt_nfiles; i++) {
2279 				if ((ff = dt->dt_ff[i]) == NULL) {
2280 					continue;
2281 				}
2282 				if ((fp = atomic_load_consume(&ff->ff_file)) ==
2283 				    NULL) {
2284 					continue;
2285 				}
2286 
2287 				if ((op == KERN_FILE_BYFILE) &&
2288 				    (fp->f_marker == sysctl_file_marker)) {
2289 					continue;
2290 				}
2291 				if (len >= elem_size && elem_count > 0) {
2292 					mutex_enter(&fp->f_lock);
2293 					fill_file2(&kf, fp, ff, i, p->p_pid);
2294 					mutex_exit(&fp->f_lock);
2295 					mutex_exit(&fd->fd_lock);
2296 					error = sysctl_copyout(l,
2297 					    &kf, dp, out_size);
2298 					mutex_enter(&fd->fd_lock);
2299 					if (error)
2300 						break;
2301 					dp += elem_size;
2302 					len -= elem_size;
2303 				}
2304 				if (op == KERN_FILE_BYFILE)
2305 					fp->f_marker = sysctl_file_marker;
2306 				needed += elem_size;
2307 				if (elem_count > 0 && elem_count != INT_MAX)
2308 					elem_count--;
2309 			}
2310 			mutex_exit(&fd->fd_lock);
2311 
2312 			/*
2313 			 * Release reference to process.
2314 			 */
2315 			mutex_enter(&proc_lock);
2316 			rw_exit(&p->p_reflock);
2317 		}
2318 		if (op == KERN_FILE_BYFILE) {
2319 			sysctl_file_marker++;
2320 
2321 			/* Reset all markers if wrapped. */
2322 			if (sysctl_file_marker == 0) {
2323 				sysctl_file_marker_reset();
2324 				sysctl_file_marker++;
2325 			}
2326 		}
2327 		mutex_exit(&proc_lock);
2328 		if (op == KERN_FILE_BYFILE)
2329 			mutex_exit(&sysctl_file_marker_lock);
2330 		sysctl_relock();
2331 		break;
2332 	default:
2333 		return EINVAL;
2334 	}
2335 
2336 	if (oldp == NULL)
2337 		needed += KERN_FILESLOP * elem_size;
2338 	*oldlenp = needed;
2339 
2340 	return error;
2341 }
2342 
2343 static void
fill_file(struct file * fp,const struct file * fpsrc)2344 fill_file(struct file *fp, const struct file *fpsrc)
2345 {
2346 	const bool allowaddr = get_expose_address(curproc);
2347 
2348 	memset(fp, 0, sizeof(*fp));
2349 
2350 	fp->f_offset = fpsrc->f_offset;
2351 	COND_SET_PTR(fp->f_cred, fpsrc->f_cred, allowaddr);
2352 	COND_SET_CPTR(fp->f_ops, fpsrc->f_ops, allowaddr);
2353 	COND_SET_STRUCT(fp->f_undata, fpsrc->f_undata, allowaddr);
2354 	COND_SET_STRUCT(fp->f_list, fpsrc->f_list, allowaddr);
2355 	fp->f_flag = fpsrc->f_flag;
2356 	fp->f_marker = fpsrc->f_marker;
2357 	fp->f_type = fpsrc->f_type;
2358 	fp->f_advice = fpsrc->f_advice;
2359 	fp->f_count = fpsrc->f_count;
2360 	fp->f_msgcount = fpsrc->f_msgcount;
2361 	fp->f_unpcount = fpsrc->f_unpcount;
2362 	COND_SET_STRUCT(fp->f_unplist, fpsrc->f_unplist, allowaddr);
2363 }
2364 
2365 static void
fill_file2(struct kinfo_file * kp,const file_t * fp,const fdfile_t * ff,int i,pid_t pid)2366 fill_file2(struct kinfo_file *kp, const file_t *fp, const fdfile_t *ff,
2367 	  int i, pid_t pid)
2368 {
2369 	const bool allowaddr = get_expose_address(curproc);
2370 
2371 	memset(kp, 0, sizeof(*kp));
2372 
2373 	COND_SET_VALUE(kp->ki_fileaddr, PTRTOUINT64(fp), allowaddr);
2374 	kp->ki_flag =		fp->f_flag;
2375 	kp->ki_iflags =		0;
2376 	kp->ki_ftype =		fp->f_type;
2377 	kp->ki_count =		fp->f_count;
2378 	kp->ki_msgcount =	fp->f_msgcount;
2379 	COND_SET_VALUE(kp->ki_fucred, PTRTOUINT64(fp->f_cred), allowaddr);
2380 	kp->ki_fuid =		kauth_cred_geteuid(fp->f_cred);
2381 	kp->ki_fgid =		kauth_cred_getegid(fp->f_cred);
2382 	COND_SET_VALUE(kp->ki_fops, PTRTOUINT64(fp->f_ops), allowaddr);
2383 	kp->ki_foffset =	fp->f_offset;
2384 	COND_SET_VALUE(kp->ki_fdata, PTRTOUINT64(fp->f_data), allowaddr);
2385 
2386 	/* vnode information to glue this file to something */
2387 	if (fp->f_type == DTYPE_VNODE) {
2388 		struct vnode *vp = fp->f_vnode;
2389 
2390 		COND_SET_VALUE(kp->ki_vun, PTRTOUINT64(vp->v_un.vu_socket),
2391 		    allowaddr);
2392 		kp->ki_vsize =	vp->v_size;
2393 		kp->ki_vtype =	vp->v_type;
2394 		kp->ki_vtag =	vp->v_tag;
2395 		COND_SET_VALUE(kp->ki_vdata, PTRTOUINT64(vp->v_data),
2396 		    allowaddr);
2397 	}
2398 
2399 	/* process information when retrieved via KERN_FILE_BYPID */
2400 	if (ff != NULL) {
2401 		kp->ki_pid =		pid;
2402 		kp->ki_fd =		i;
2403 		kp->ki_ofileflags =	ff->ff_exclose;
2404 		kp->ki_usecount =	ff->ff_refcnt;
2405 	}
2406 }
2407