xref: /freebsd/sys/kern/uipc_shm.c (revision f157ca4696f5922275d5d451736005b9332eb136)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2006, 2011, 2016-2017 Robert N. M. Watson
5  * All rights reserved.
6  *
7  * Portions of this software were developed by BAE Systems, the University of
8  * Cambridge Computer Laboratory, and Memorial University under DARPA/AFRL
9  * contract FA8650-15-C-7558 ("CADETS"), as part of the DARPA Transparent
10  * Computing (TC) research program.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 /*
35  * Support for shared swap-backed anonymous memory objects via
36  * shm_open(2), shm_rename(2), and shm_unlink(2).
37  * While most of the implementation is here, vm_mmap.c contains
38  * mapping logic changes.
39  *
40  * posixshmcontrol(1) allows users to inspect the state of the memory
41  * objects.  Per-uid swap resource limit controls total amount of
42  * memory that user can consume for anonymous objects, including
43  * shared.
44  */
45 
46 #include <sys/cdefs.h>
47 __FBSDID("$FreeBSD$");
48 
49 #include "opt_capsicum.h"
50 #include "opt_ktrace.h"
51 
52 #include <sys/param.h>
53 #include <sys/capsicum.h>
54 #include <sys/conf.h>
55 #include <sys/fcntl.h>
56 #include <sys/file.h>
57 #include <sys/filedesc.h>
58 #include <sys/filio.h>
59 #include <sys/fnv_hash.h>
60 #include <sys/kernel.h>
61 #include <sys/limits.h>
62 #include <sys/uio.h>
63 #include <sys/signal.h>
64 #include <sys/jail.h>
65 #include <sys/ktrace.h>
66 #include <sys/lock.h>
67 #include <sys/malloc.h>
68 #include <sys/mman.h>
69 #include <sys/mutex.h>
70 #include <sys/priv.h>
71 #include <sys/proc.h>
72 #include <sys/refcount.h>
73 #include <sys/resourcevar.h>
74 #include <sys/rwlock.h>
75 #include <sys/sbuf.h>
76 #include <sys/stat.h>
77 #include <sys/syscallsubr.h>
78 #include <sys/sysctl.h>
79 #include <sys/sysproto.h>
80 #include <sys/systm.h>
81 #include <sys/sx.h>
82 #include <sys/time.h>
83 #include <sys/vnode.h>
84 #include <sys/unistd.h>
85 #include <sys/user.h>
86 
87 #include <security/audit/audit.h>
88 #include <security/mac/mac_framework.h>
89 
90 #include <vm/vm.h>
91 #include <vm/vm_param.h>
92 #include <vm/pmap.h>
93 #include <vm/vm_extern.h>
94 #include <vm/vm_map.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_pageout.h>
99 #include <vm/vm_pager.h>
100 #include <vm/swap_pager.h>
101 
102 struct shm_mapping {
103 	char		*sm_path;
104 	Fnv32_t		sm_fnv;
105 	struct shmfd	*sm_shmfd;
106 	LIST_ENTRY(shm_mapping) sm_link;
107 };
108 
109 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor");
110 static LIST_HEAD(, shm_mapping) *shm_dictionary;
111 static struct sx shm_dict_lock;
112 static struct mtx shm_timestamp_lock;
113 static u_long shm_hash;
114 static struct unrhdr64 shm_ino_unr;
115 static dev_t shm_dev_ino;
116 
117 #define	SHM_HASH(fnv)	(&shm_dictionary[(fnv) & shm_hash])
118 
119 static void	shm_init(void *arg);
120 static void	shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd);
121 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv);
122 static int	shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred);
123 static int	shm_dotruncate_locked(struct shmfd *shmfd, off_t length,
124     void *rl_cookie);
125 static int	shm_copyin_path(struct thread *td, const char *userpath_in,
126     char **path_out);
127 
128 static fo_rdwr_t	shm_read;
129 static fo_rdwr_t	shm_write;
130 static fo_truncate_t	shm_truncate;
131 static fo_ioctl_t	shm_ioctl;
132 static fo_stat_t	shm_stat;
133 static fo_close_t	shm_close;
134 static fo_chmod_t	shm_chmod;
135 static fo_chown_t	shm_chown;
136 static fo_seek_t	shm_seek;
137 static fo_fill_kinfo_t	shm_fill_kinfo;
138 static fo_mmap_t	shm_mmap;
139 static fo_get_seals_t	shm_get_seals;
140 static fo_add_seals_t	shm_add_seals;
141 
142 /* File descriptor operations. */
143 struct fileops shm_ops = {
144 	.fo_read = shm_read,
145 	.fo_write = shm_write,
146 	.fo_truncate = shm_truncate,
147 	.fo_ioctl = shm_ioctl,
148 	.fo_poll = invfo_poll,
149 	.fo_kqfilter = invfo_kqfilter,
150 	.fo_stat = shm_stat,
151 	.fo_close = shm_close,
152 	.fo_chmod = shm_chmod,
153 	.fo_chown = shm_chown,
154 	.fo_sendfile = vn_sendfile,
155 	.fo_seek = shm_seek,
156 	.fo_fill_kinfo = shm_fill_kinfo,
157 	.fo_mmap = shm_mmap,
158 	.fo_get_seals = shm_get_seals,
159 	.fo_add_seals = shm_add_seals,
160 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
161 };
162 
163 FEATURE(posix_shm, "POSIX shared memory");
164 
165 static int
166 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio)
167 {
168 	vm_page_t m;
169 	vm_pindex_t idx;
170 	size_t tlen;
171 	int error, offset, rv;
172 
173 	idx = OFF_TO_IDX(uio->uio_offset);
174 	offset = uio->uio_offset & PAGE_MASK;
175 	tlen = MIN(PAGE_SIZE - offset, len);
176 
177 	VM_OBJECT_WLOCK(obj);
178 
179 	/*
180 	 * Read I/O without either a corresponding resident page or swap
181 	 * page: use zero_region.  This is intended to avoid instantiating
182 	 * pages on read from a sparse region.
183 	 */
184 	if (uio->uio_rw == UIO_READ && vm_page_lookup(obj, idx) == NULL &&
185 	    !vm_pager_has_page(obj, idx, NULL, NULL)) {
186 		VM_OBJECT_WUNLOCK(obj);
187 		return (uiomove(__DECONST(void *, zero_region), tlen, uio));
188 	}
189 
190 	/*
191 	 * Parallel reads of the page content from disk are prevented
192 	 * by exclusive busy.
193 	 *
194 	 * Although the tmpfs vnode lock is held here, it is
195 	 * nonetheless safe to sleep waiting for a free page.  The
196 	 * pageout daemon does not need to acquire the tmpfs vnode
197 	 * lock to page out tobj's pages because tobj is a OBJT_SWAP
198 	 * type object.
199 	 */
200 	rv = vm_page_grab_valid(&m, obj, idx,
201 	    VM_ALLOC_NORMAL | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY);
202 	if (rv != VM_PAGER_OK) {
203 		VM_OBJECT_WUNLOCK(obj);
204 		printf("uiomove_object: vm_obj %p idx %jd pager error %d\n",
205 		    obj, idx, rv);
206 		return (EIO);
207 	}
208 	VM_OBJECT_WUNLOCK(obj);
209 	error = uiomove_fromphys(&m, offset, tlen, uio);
210 	if (uio->uio_rw == UIO_WRITE && error == 0)
211 		vm_page_set_dirty(m);
212 	vm_page_lock(m);
213 	vm_page_activate(m);
214 	vm_page_unlock(m);
215 	vm_page_sunbusy(m);
216 
217 	return (error);
218 }
219 
220 int
221 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
222 {
223 	ssize_t resid;
224 	size_t len;
225 	int error;
226 
227 	error = 0;
228 	while ((resid = uio->uio_resid) > 0) {
229 		if (obj_size <= uio->uio_offset)
230 			break;
231 		len = MIN(obj_size - uio->uio_offset, resid);
232 		if (len == 0)
233 			break;
234 		error = uiomove_object_page(obj, len, uio);
235 		if (error != 0 || resid == uio->uio_resid)
236 			break;
237 	}
238 	return (error);
239 }
240 
241 static int
242 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
243 {
244 	struct shmfd *shmfd;
245 	off_t foffset;
246 	int error;
247 
248 	shmfd = fp->f_data;
249 	foffset = foffset_lock(fp, 0);
250 	error = 0;
251 	switch (whence) {
252 	case L_INCR:
253 		if (foffset < 0 ||
254 		    (offset > 0 && foffset > OFF_MAX - offset)) {
255 			error = EOVERFLOW;
256 			break;
257 		}
258 		offset += foffset;
259 		break;
260 	case L_XTND:
261 		if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
262 			error = EOVERFLOW;
263 			break;
264 		}
265 		offset += shmfd->shm_size;
266 		break;
267 	case L_SET:
268 		break;
269 	default:
270 		error = EINVAL;
271 	}
272 	if (error == 0) {
273 		if (offset < 0 || offset > shmfd->shm_size)
274 			error = EINVAL;
275 		else
276 			td->td_uretoff.tdu_off = offset;
277 	}
278 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
279 	return (error);
280 }
281 
282 static int
283 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
284     int flags, struct thread *td)
285 {
286 	struct shmfd *shmfd;
287 	void *rl_cookie;
288 	int error;
289 
290 	shmfd = fp->f_data;
291 #ifdef MAC
292 	error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
293 	if (error)
294 		return (error);
295 #endif
296 	foffset_lock_uio(fp, uio, flags);
297 	rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset,
298 	    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
299 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
300 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
301 	foffset_unlock_uio(fp, uio, flags);
302 	return (error);
303 }
304 
305 static int
306 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
307     int flags, struct thread *td)
308 {
309 	struct shmfd *shmfd;
310 	void *rl_cookie;
311 	int error;
312 
313 	shmfd = fp->f_data;
314 #ifdef MAC
315 	error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
316 	if (error)
317 		return (error);
318 #endif
319 	foffset_lock_uio(fp, uio, flags);
320 	if ((flags & FOF_OFFSET) == 0) {
321 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
322 		    &shmfd->shm_mtx);
323 	} else {
324 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset,
325 		    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
326 	}
327 	if ((shmfd->shm_seals & F_SEAL_WRITE) != 0)
328 		error = EPERM;
329 	else
330 		error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
331 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
332 	foffset_unlock_uio(fp, uio, flags);
333 	return (error);
334 }
335 
336 static int
337 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
338     struct thread *td)
339 {
340 	struct shmfd *shmfd;
341 #ifdef MAC
342 	int error;
343 #endif
344 
345 	shmfd = fp->f_data;
346 #ifdef MAC
347 	error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
348 	if (error)
349 		return (error);
350 #endif
351 	return (shm_dotruncate(shmfd, length));
352 }
353 
354 int
355 shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
356     struct thread *td)
357 {
358 
359 	switch (com) {
360 	case FIONBIO:
361 	case FIOASYNC:
362 		/*
363 		 * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work,
364 		 * just like it would on an unlinked regular file
365 		 */
366 		return (0);
367 	default:
368 		return (ENOTTY);
369 	}
370 }
371 
372 static int
373 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
374     struct thread *td)
375 {
376 	struct shmfd *shmfd;
377 #ifdef MAC
378 	int error;
379 #endif
380 
381 	shmfd = fp->f_data;
382 
383 #ifdef MAC
384 	error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
385 	if (error)
386 		return (error);
387 #endif
388 
389 	/*
390 	 * Attempt to return sanish values for fstat() on a memory file
391 	 * descriptor.
392 	 */
393 	bzero(sb, sizeof(*sb));
394 	sb->st_blksize = PAGE_SIZE;
395 	sb->st_size = shmfd->shm_size;
396 	sb->st_blocks = howmany(sb->st_size, sb->st_blksize);
397 	mtx_lock(&shm_timestamp_lock);
398 	sb->st_atim = shmfd->shm_atime;
399 	sb->st_ctim = shmfd->shm_ctime;
400 	sb->st_mtim = shmfd->shm_mtime;
401 	sb->st_birthtim = shmfd->shm_birthtime;
402 	sb->st_mode = S_IFREG | shmfd->shm_mode;		/* XXX */
403 	sb->st_uid = shmfd->shm_uid;
404 	sb->st_gid = shmfd->shm_gid;
405 	mtx_unlock(&shm_timestamp_lock);
406 	sb->st_dev = shm_dev_ino;
407 	sb->st_ino = shmfd->shm_ino;
408 	sb->st_nlink = shmfd->shm_object->ref_count;
409 
410 	return (0);
411 }
412 
413 static int
414 shm_close(struct file *fp, struct thread *td)
415 {
416 	struct shmfd *shmfd;
417 
418 	shmfd = fp->f_data;
419 	fp->f_data = NULL;
420 	shm_drop(shmfd);
421 
422 	return (0);
423 }
424 
425 static int
426 shm_copyin_path(struct thread *td, const char *userpath_in, char **path_out) {
427 	int error;
428 	char *path;
429 	const char *pr_path;
430 	size_t pr_pathlen;
431 
432 	path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
433 	pr_path = td->td_ucred->cr_prison->pr_path;
434 
435 	/* Construct a full pathname for jailed callers. */
436 	pr_pathlen = strcmp(pr_path, "/") ==
437 	    0 ? 0 : strlcpy(path, pr_path, MAXPATHLEN);
438 	error = copyinstr(userpath_in, path + pr_pathlen,
439 	    MAXPATHLEN - pr_pathlen, NULL);
440 	if (error != 0)
441 		goto out;
442 
443 #ifdef KTRACE
444 	if (KTRPOINT(curthread, KTR_NAMEI))
445 		ktrnamei(path);
446 #endif
447 
448 	/* Require paths to start with a '/' character. */
449 	if (path[pr_pathlen] != '/') {
450 		error = EINVAL;
451 		goto out;
452 	}
453 
454 	*path_out = path;
455 
456 out:
457 	if (error != 0)
458 		free(path, M_SHMFD);
459 
460 	return (error);
461 }
462 
463 static int
464 shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie)
465 {
466 	vm_object_t object;
467 	vm_page_t m;
468 	vm_pindex_t idx, nobjsize;
469 	vm_ooffset_t delta;
470 	int base, rv;
471 
472 	KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
473 	object = shmfd->shm_object;
474 	VM_OBJECT_ASSERT_WLOCKED(object);
475 	rangelock_cookie_assert(rl_cookie, RA_WLOCKED);
476 	if (length == shmfd->shm_size)
477 		return (0);
478 	nobjsize = OFF_TO_IDX(length + PAGE_MASK);
479 
480 	/* Are we shrinking?  If so, trim the end. */
481 	if (length < shmfd->shm_size) {
482 		if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0)
483 			return (EPERM);
484 
485 		/*
486 		 * Disallow any requests to shrink the size if this
487 		 * object is mapped into the kernel.
488 		 */
489 		if (shmfd->shm_kmappings > 0)
490 			return (EBUSY);
491 
492 		/*
493 		 * Zero the truncated part of the last page.
494 		 */
495 		base = length & PAGE_MASK;
496 		if (base != 0) {
497 			idx = OFF_TO_IDX(length);
498 retry:
499 			m = vm_page_grab(object, idx, VM_ALLOC_NOCREAT);
500 			if (m != NULL) {
501 				MPASS(vm_page_all_valid(m));
502 			} else if (vm_pager_has_page(object, idx, NULL, NULL)) {
503 				m = vm_page_alloc(object, idx,
504 				    VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL);
505 				if (m == NULL)
506 					goto retry;
507 				rv = vm_pager_get_pages(object, &m, 1, NULL,
508 				    NULL);
509 				if (rv == VM_PAGER_OK) {
510 					/*
511 					 * Since the page was not resident,
512 					 * and therefore not recently
513 					 * accessed, immediately enqueue it
514 					 * for asynchronous laundering.  The
515 					 * current operation is not regarded
516 					 * as an access.
517 					 */
518 					vm_page_launder(m);
519 				} else {
520 					vm_page_free(m);
521 					VM_OBJECT_WUNLOCK(object);
522 					return (EIO);
523 				}
524 			}
525 			if (m != NULL) {
526 				pmap_zero_page_area(m, base, PAGE_SIZE - base);
527 				KASSERT(vm_page_all_valid(m),
528 				    ("shm_dotruncate: page %p is invalid", m));
529 				vm_page_set_dirty(m);
530 				vm_page_xunbusy(m);
531 			}
532 		}
533 		delta = IDX_TO_OFF(object->size - nobjsize);
534 
535 		/* Toss in memory pages. */
536 		if (nobjsize < object->size)
537 			vm_object_page_remove(object, nobjsize, object->size,
538 			    0);
539 
540 		/* Toss pages from swap. */
541 		if (object->type == OBJT_SWAP)
542 			swap_pager_freespace(object, nobjsize, delta);
543 
544 		/* Free the swap accounted for shm */
545 		swap_release_by_cred(delta, object->cred);
546 		object->charge -= delta;
547 	} else {
548 		if ((shmfd->shm_seals & F_SEAL_GROW) != 0)
549 			return (EPERM);
550 
551 		/* Try to reserve additional swap space. */
552 		delta = IDX_TO_OFF(nobjsize - object->size);
553 		if (!swap_reserve_by_cred(delta, object->cred))
554 			return (ENOMEM);
555 		object->charge += delta;
556 	}
557 	shmfd->shm_size = length;
558 	mtx_lock(&shm_timestamp_lock);
559 	vfs_timestamp(&shmfd->shm_ctime);
560 	shmfd->shm_mtime = shmfd->shm_ctime;
561 	mtx_unlock(&shm_timestamp_lock);
562 	object->size = nobjsize;
563 	return (0);
564 }
565 
566 int
567 shm_dotruncate(struct shmfd *shmfd, off_t length)
568 {
569 	void *rl_cookie;
570 	int error;
571 
572 	rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
573 	    &shmfd->shm_mtx);
574 	VM_OBJECT_WLOCK(shmfd->shm_object);
575 	error = shm_dotruncate_locked(shmfd, length, rl_cookie);
576 	VM_OBJECT_WUNLOCK(shmfd->shm_object);
577 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
578 	return (error);
579 }
580 
581 /*
582  * shmfd object management including creation and reference counting
583  * routines.
584  */
585 struct shmfd *
586 shm_alloc(struct ucred *ucred, mode_t mode)
587 {
588 	struct shmfd *shmfd;
589 
590 	shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
591 	shmfd->shm_size = 0;
592 	shmfd->shm_uid = ucred->cr_uid;
593 	shmfd->shm_gid = ucred->cr_gid;
594 	shmfd->shm_mode = mode;
595 	shmfd->shm_object = vm_pager_allocate(OBJT_SWAP, NULL,
596 	    shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
597 	KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
598 	vfs_timestamp(&shmfd->shm_birthtime);
599 	shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
600 	    shmfd->shm_birthtime;
601 	shmfd->shm_ino = alloc_unr64(&shm_ino_unr);
602 	refcount_init(&shmfd->shm_refs, 1);
603 	mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
604 	rangelock_init(&shmfd->shm_rl);
605 #ifdef MAC
606 	mac_posixshm_init(shmfd);
607 	mac_posixshm_create(ucred, shmfd);
608 #endif
609 
610 	return (shmfd);
611 }
612 
613 struct shmfd *
614 shm_hold(struct shmfd *shmfd)
615 {
616 
617 	refcount_acquire(&shmfd->shm_refs);
618 	return (shmfd);
619 }
620 
621 void
622 shm_drop(struct shmfd *shmfd)
623 {
624 
625 	if (refcount_release(&shmfd->shm_refs)) {
626 #ifdef MAC
627 		mac_posixshm_destroy(shmfd);
628 #endif
629 		rangelock_destroy(&shmfd->shm_rl);
630 		mtx_destroy(&shmfd->shm_mtx);
631 		vm_object_deallocate(shmfd->shm_object);
632 		free(shmfd, M_SHMFD);
633 	}
634 }
635 
636 /*
637  * Determine if the credentials have sufficient permissions for a
638  * specified combination of FREAD and FWRITE.
639  */
640 int
641 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
642 {
643 	accmode_t accmode;
644 	int error;
645 
646 	accmode = 0;
647 	if (flags & FREAD)
648 		accmode |= VREAD;
649 	if (flags & FWRITE)
650 		accmode |= VWRITE;
651 	mtx_lock(&shm_timestamp_lock);
652 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
653 	    accmode, ucred, NULL);
654 	mtx_unlock(&shm_timestamp_lock);
655 	return (error);
656 }
657 
658 /*
659  * Dictionary management.  We maintain an in-kernel dictionary to map
660  * paths to shmfd objects.  We use the FNV hash on the path to store
661  * the mappings in a hash table.
662  */
663 static void
664 shm_init(void *arg)
665 {
666 
667 	mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
668 	sx_init(&shm_dict_lock, "shm dictionary");
669 	shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
670 	new_unrhdr64(&shm_ino_unr, 1);
671 	shm_dev_ino = devfs_alloc_cdp_inode();
672 	KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
673 }
674 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
675 
676 static struct shmfd *
677 shm_lookup(char *path, Fnv32_t fnv)
678 {
679 	struct shm_mapping *map;
680 
681 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
682 		if (map->sm_fnv != fnv)
683 			continue;
684 		if (strcmp(map->sm_path, path) == 0)
685 			return (map->sm_shmfd);
686 	}
687 
688 	return (NULL);
689 }
690 
691 static void
692 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
693 {
694 	struct shm_mapping *map;
695 
696 	map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
697 	map->sm_path = path;
698 	map->sm_fnv = fnv;
699 	map->sm_shmfd = shm_hold(shmfd);
700 	shmfd->shm_path = path;
701 	LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
702 }
703 
704 static int
705 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
706 {
707 	struct shm_mapping *map;
708 	int error;
709 
710 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
711 		if (map->sm_fnv != fnv)
712 			continue;
713 		if (strcmp(map->sm_path, path) == 0) {
714 #ifdef MAC
715 			error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
716 			if (error)
717 				return (error);
718 #endif
719 			error = shm_access(map->sm_shmfd, ucred,
720 			    FREAD | FWRITE);
721 			if (error)
722 				return (error);
723 			map->sm_shmfd->shm_path = NULL;
724 			LIST_REMOVE(map, sm_link);
725 			shm_drop(map->sm_shmfd);
726 			free(map->sm_path, M_SHMFD);
727 			free(map, M_SHMFD);
728 			return (0);
729 		}
730 	}
731 
732 	return (ENOENT);
733 }
734 
735 int
736 kern_shm_open(struct thread *td, const char *userpath, int flags, mode_t mode,
737     struct filecaps *fcaps, int initial_seals)
738 {
739 	struct filedesc *fdp;
740 	struct shmfd *shmfd;
741 	struct file *fp;
742 	char *path;
743 	void *rl_cookie;
744 	Fnv32_t fnv;
745 	mode_t cmode;
746 	int fd, error;
747 
748 #ifdef CAPABILITY_MODE
749 	/*
750 	 * shm_open(2) is only allowed for anonymous objects.
751 	 */
752 	if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON))
753 		return (ECAPMODE);
754 #endif
755 
756 	AUDIT_ARG_FFLAGS(flags);
757 	AUDIT_ARG_MODE(mode);
758 
759 	if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
760 		return (EINVAL);
761 
762 	if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
763 		return (EINVAL);
764 
765 	/*
766 	 * Currently only F_SEAL_SEAL may be set when creating or opening shmfd.
767 	 * If the decision is made later to allow additional seals, care must be
768 	 * taken below to ensure that the seals are properly set if the shmfd
769 	 * already existed -- this currently assumes that only F_SEAL_SEAL can
770 	 * be set and doesn't take further precautions to ensure the validity of
771 	 * the seals being added with respect to current mappings.
772 	 */
773 	if ((initial_seals & ~F_SEAL_SEAL) != 0)
774 		return (EINVAL);
775 
776 	fdp = td->td_proc->p_fd;
777 	cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS;
778 
779 	/*
780 	 * shm_open(2) created shm should always have O_CLOEXEC set, as mandated
781 	 * by POSIX.  We allow it to be unset here so that an in-kernel
782 	 * interface may be written as a thin layer around shm, optionally not
783 	 * setting CLOEXEC.  For shm_open(2), O_CLOEXEC is set unconditionally
784 	 * in sys_shm_open() to keep this implementation compliant.
785 	 */
786 	error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps);
787 	if (error)
788 		return (error);
789 
790 	/* A SHM_ANON path pointer creates an anonymous object. */
791 	if (userpath == SHM_ANON) {
792 		/* A read-only anonymous object is pointless. */
793 		if ((flags & O_ACCMODE) == O_RDONLY) {
794 			fdclose(td, fp, fd);
795 			fdrop(fp, td);
796 			return (EINVAL);
797 		}
798 		shmfd = shm_alloc(td->td_ucred, cmode);
799 		shmfd->shm_seals = initial_seals;
800 	} else {
801 		error = shm_copyin_path(td, userpath, &path);
802 		if (error != 0) {
803 			fdclose(td, fp, fd);
804 			fdrop(fp, td);
805 			return (error);
806 		}
807 
808 		AUDIT_ARG_UPATH1_CANON(path);
809 		fnv = fnv_32_str(path, FNV1_32_INIT);
810 		sx_xlock(&shm_dict_lock);
811 		shmfd = shm_lookup(path, fnv);
812 		if (shmfd == NULL) {
813 			/* Object does not yet exist, create it if requested. */
814 			if (flags & O_CREAT) {
815 #ifdef MAC
816 				error = mac_posixshm_check_create(td->td_ucred,
817 				    path);
818 				if (error == 0) {
819 #endif
820 					shmfd = shm_alloc(td->td_ucred, cmode);
821 					shmfd->shm_seals = initial_seals;
822 					shm_insert(path, fnv, shmfd);
823 #ifdef MAC
824 				}
825 #endif
826 			} else {
827 				free(path, M_SHMFD);
828 				error = ENOENT;
829 			}
830 		} else {
831 			rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
832 			    &shmfd->shm_mtx);
833 
834 			/*
835 			 * kern_shm_open() likely shouldn't ever error out on
836 			 * trying to set a seal that already exists, unlike
837 			 * F_ADD_SEALS.  This would break terribly as
838 			 * shm_open(2) actually sets F_SEAL_SEAL to maintain
839 			 * historical behavior where the underlying file could
840 			 * not be sealed.
841 			 */
842 			initial_seals &= ~shmfd->shm_seals;
843 
844 			/*
845 			 * Object already exists, obtain a new
846 			 * reference if requested and permitted.
847 			 */
848 			free(path, M_SHMFD);
849 
850 			/*
851 			 * initial_seals can't set additional seals if we've
852 			 * already been set F_SEAL_SEAL.  If F_SEAL_SEAL is set,
853 			 * then we've already removed that one from
854 			 * initial_seals.  This is currently redundant as we
855 			 * only allow setting F_SEAL_SEAL at creation time, but
856 			 * it's cheap to check and decreases the effort required
857 			 * to allow additional seals.
858 			 */
859 			if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 &&
860 			    initial_seals != 0)
861 				error = EPERM;
862 			else if ((flags & (O_CREAT | O_EXCL)) ==
863 			    (O_CREAT | O_EXCL))
864 				error = EEXIST;
865 			else {
866 #ifdef MAC
867 				error = mac_posixshm_check_open(td->td_ucred,
868 				    shmfd, FFLAGS(flags & O_ACCMODE));
869 				if (error == 0)
870 #endif
871 				error = shm_access(shmfd, td->td_ucred,
872 				    FFLAGS(flags & O_ACCMODE));
873 			}
874 
875 			/*
876 			 * Truncate the file back to zero length if
877 			 * O_TRUNC was specified and the object was
878 			 * opened with read/write.
879 			 */
880 			if (error == 0 &&
881 			    (flags & (O_ACCMODE | O_TRUNC)) ==
882 			    (O_RDWR | O_TRUNC)) {
883 				VM_OBJECT_WLOCK(shmfd->shm_object);
884 #ifdef MAC
885 				error = mac_posixshm_check_truncate(
886 					td->td_ucred, fp->f_cred, shmfd);
887 				if (error == 0)
888 #endif
889 					error = shm_dotruncate_locked(shmfd, 0,
890 					    rl_cookie);
891 				VM_OBJECT_WUNLOCK(shmfd->shm_object);
892 			}
893 			if (error == 0) {
894 				/*
895 				 * Currently we only allow F_SEAL_SEAL to be
896 				 * set initially.  As noted above, this would
897 				 * need to be reworked should that change.
898 				 */
899 				shmfd->shm_seals |= initial_seals;
900 				shm_hold(shmfd);
901 			}
902 			rangelock_unlock(&shmfd->shm_rl, rl_cookie,
903 			    &shmfd->shm_mtx);
904 		}
905 		sx_xunlock(&shm_dict_lock);
906 
907 		if (error) {
908 			fdclose(td, fp, fd);
909 			fdrop(fp, td);
910 			return (error);
911 		}
912 	}
913 
914 	finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
915 
916 	td->td_retval[0] = fd;
917 	fdrop(fp, td);
918 
919 	return (0);
920 }
921 
922 /* System calls. */
923 #ifdef COMPAT_FREEBSD12
924 int
925 freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap)
926 {
927 
928 	return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC, uap->mode,
929 	    NULL, F_SEAL_SEAL));
930 }
931 #endif
932 
933 int
934 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
935 {
936 	char *path;
937 	Fnv32_t fnv;
938 	int error;
939 
940 	error = shm_copyin_path(td, uap->path, &path);
941 	if (error != 0)
942 		return (error);
943 
944 	AUDIT_ARG_UPATH1_CANON(path);
945 	fnv = fnv_32_str(path, FNV1_32_INIT);
946 	sx_xlock(&shm_dict_lock);
947 	error = shm_remove(path, fnv, td->td_ucred);
948 	sx_xunlock(&shm_dict_lock);
949 	free(path, M_TEMP);
950 
951 	return (error);
952 }
953 
954 int
955 sys_shm_rename(struct thread *td, struct shm_rename_args *uap)
956 {
957 	char *path_from = NULL, *path_to = NULL;
958 	Fnv32_t fnv_from, fnv_to;
959 	struct shmfd *fd_from;
960 	struct shmfd *fd_to;
961 	int error;
962 	int flags;
963 
964 	flags = uap->flags;
965 	AUDIT_ARG_FFLAGS(flags);
966 
967 	/*
968 	 * Make sure the user passed only valid flags.
969 	 * If you add a new flag, please add a new term here.
970 	 */
971 	if ((flags & ~(
972 	    SHM_RENAME_NOREPLACE |
973 	    SHM_RENAME_EXCHANGE
974 	    )) != 0) {
975 		error = EINVAL;
976 		goto out;
977 	}
978 
979 	/*
980 	 * EXCHANGE and NOREPLACE don't quite make sense together. Let's
981 	 * force the user to choose one or the other.
982 	 */
983 	if ((flags & SHM_RENAME_NOREPLACE) != 0 &&
984 	    (flags & SHM_RENAME_EXCHANGE) != 0) {
985 		error = EINVAL;
986 		goto out;
987 	}
988 
989 	/* Renaming to or from anonymous makes no sense */
990 	if (uap->path_from == SHM_ANON || uap->path_to == SHM_ANON) {
991 		error = EINVAL;
992 		goto out;
993 	}
994 
995 	error = shm_copyin_path(td, uap->path_from, &path_from);
996 	if (error != 0)
997 		goto out;
998 
999 	error = shm_copyin_path(td, uap->path_to, &path_to);
1000 	if (error != 0)
1001 		goto out;
1002 
1003 	AUDIT_ARG_UPATH1_CANON(path_from);
1004 	AUDIT_ARG_UPATH2_CANON(path_to);
1005 
1006 	/* Rename with from/to equal is a no-op */
1007 	if (strcmp(path_from, path_to) == 0)
1008 		goto out;
1009 
1010 	fnv_from = fnv_32_str(path_from, FNV1_32_INIT);
1011 	fnv_to = fnv_32_str(path_to, FNV1_32_INIT);
1012 
1013 	sx_xlock(&shm_dict_lock);
1014 
1015 	fd_from = shm_lookup(path_from, fnv_from);
1016 	if (fd_from == NULL) {
1017 		error = ENOENT;
1018 		goto out_locked;
1019 	}
1020 
1021 	fd_to = shm_lookup(path_to, fnv_to);
1022 	if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) {
1023 		error = EEXIST;
1024 		goto out_locked;
1025 	}
1026 
1027 	/*
1028 	 * Unconditionally prevents shm_remove from invalidating the 'from'
1029 	 * shm's state.
1030 	 */
1031 	shm_hold(fd_from);
1032 	error = shm_remove(path_from, fnv_from, td->td_ucred);
1033 
1034 	/*
1035 	 * One of my assumptions failed if ENOENT (e.g. locking didn't
1036 	 * protect us)
1037 	 */
1038 	KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s",
1039 	    path_from));
1040 	if (error != 0) {
1041 		shm_drop(fd_from);
1042 		goto out_locked;
1043 	}
1044 
1045 	/*
1046 	 * If we are exchanging, we need to ensure the shm_remove below
1047 	 * doesn't invalidate the dest shm's state.
1048 	 */
1049 	if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL)
1050 		shm_hold(fd_to);
1051 
1052 	/*
1053 	 * NOTE: if path_to is not already in the hash, c'est la vie;
1054 	 * it simply means we have nothing already at path_to to unlink.
1055 	 * That is the ENOENT case.
1056 	 *
1057 	 * If we somehow don't have access to unlink this guy, but
1058 	 * did for the shm at path_from, then relink the shm to path_from
1059 	 * and abort with EACCES.
1060 	 *
1061 	 * All other errors: that is weird; let's relink and abort the
1062 	 * operation.
1063 	 */
1064 	error = shm_remove(path_to, fnv_to, td->td_ucred);
1065 	if (error != 0 && error != ENOENT) {
1066 		shm_insert(path_from, fnv_from, fd_from);
1067 		shm_drop(fd_from);
1068 		/* Don't free path_from now, since the hash references it */
1069 		path_from = NULL;
1070 		goto out_locked;
1071 	}
1072 
1073 	error = 0;
1074 
1075 	shm_insert(path_to, fnv_to, fd_from);
1076 
1077 	/* Don't free path_to now, since the hash references it */
1078 	path_to = NULL;
1079 
1080 	/* We kept a ref when we removed, and incremented again in insert */
1081 	shm_drop(fd_from);
1082 	KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1083 	    fd_from->shm_refs));
1084 
1085 	if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) {
1086 		shm_insert(path_from, fnv_from, fd_to);
1087 		path_from = NULL;
1088 		shm_drop(fd_to);
1089 		KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1090 		    fd_to->shm_refs));
1091 	}
1092 
1093 out_locked:
1094 	sx_xunlock(&shm_dict_lock);
1095 
1096 out:
1097 	free(path_from, M_SHMFD);
1098 	free(path_to, M_SHMFD);
1099 	return (error);
1100 }
1101 
1102 int
1103 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
1104     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
1105     vm_ooffset_t foff, struct thread *td)
1106 {
1107 	struct shmfd *shmfd;
1108 	vm_prot_t maxprot;
1109 	int error;
1110 	bool writecnt;
1111 	void *rl_cookie;
1112 
1113 	shmfd = fp->f_data;
1114 	maxprot = VM_PROT_NONE;
1115 
1116 	rl_cookie = rangelock_rlock(&shmfd->shm_rl, 0, objsize,
1117 	    &shmfd->shm_mtx);
1118 	/* FREAD should always be set. */
1119 	if ((fp->f_flag & FREAD) != 0)
1120 		maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
1121 	if ((fp->f_flag & FWRITE) != 0)
1122 		maxprot |= VM_PROT_WRITE;
1123 
1124 	writecnt = (flags & MAP_SHARED) != 0 && (prot & VM_PROT_WRITE) != 0;
1125 
1126 	if (writecnt && (shmfd->shm_seals & F_SEAL_WRITE) != 0) {
1127 		error = EPERM;
1128 		goto out;
1129 	}
1130 
1131 	/* Don't permit shared writable mappings on read-only descriptors. */
1132 	if (writecnt && (maxprot & VM_PROT_WRITE) == 0) {
1133 		error = EACCES;
1134 		goto out;
1135 	}
1136 	maxprot &= cap_maxprot;
1137 
1138 	/* See comment in vn_mmap(). */
1139 	if (
1140 #ifdef _LP64
1141 	    objsize > OFF_MAX ||
1142 #endif
1143 	    foff < 0 || foff > OFF_MAX - objsize) {
1144 		error = EINVAL;
1145 		goto out;
1146 	}
1147 
1148 #ifdef MAC
1149 	error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
1150 	if (error != 0)
1151 		goto out;
1152 #endif
1153 
1154 	mtx_lock(&shm_timestamp_lock);
1155 	vfs_timestamp(&shmfd->shm_atime);
1156 	mtx_unlock(&shm_timestamp_lock);
1157 	vm_object_reference(shmfd->shm_object);
1158 
1159 	if (writecnt)
1160 		vm_pager_update_writecount(shmfd->shm_object, 0, objsize);
1161 	error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
1162 	    shmfd->shm_object, foff, writecnt, td);
1163 	if (error != 0) {
1164 		if (writecnt)
1165 			vm_pager_release_writecount(shmfd->shm_object, 0,
1166 			    objsize);
1167 		vm_object_deallocate(shmfd->shm_object);
1168 	}
1169 out:
1170 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
1171 	return (error);
1172 }
1173 
1174 static int
1175 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
1176     struct thread *td)
1177 {
1178 	struct shmfd *shmfd;
1179 	int error;
1180 
1181 	error = 0;
1182 	shmfd = fp->f_data;
1183 	mtx_lock(&shm_timestamp_lock);
1184 	/*
1185 	 * SUSv4 says that x bits of permission need not be affected.
1186 	 * Be consistent with our shm_open there.
1187 	 */
1188 #ifdef MAC
1189 	error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
1190 	if (error != 0)
1191 		goto out;
1192 #endif
1193 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid,
1194 	    shmfd->shm_gid, VADMIN, active_cred, NULL);
1195 	if (error != 0)
1196 		goto out;
1197 	shmfd->shm_mode = mode & ACCESSPERMS;
1198 out:
1199 	mtx_unlock(&shm_timestamp_lock);
1200 	return (error);
1201 }
1202 
1203 static int
1204 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1205     struct thread *td)
1206 {
1207 	struct shmfd *shmfd;
1208 	int error;
1209 
1210 	error = 0;
1211 	shmfd = fp->f_data;
1212 	mtx_lock(&shm_timestamp_lock);
1213 #ifdef MAC
1214 	error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
1215 	if (error != 0)
1216 		goto out;
1217 #endif
1218 	if (uid == (uid_t)-1)
1219 		uid = shmfd->shm_uid;
1220 	if (gid == (gid_t)-1)
1221                  gid = shmfd->shm_gid;
1222 	if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
1223 	    (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
1224 	    (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN)))
1225 		goto out;
1226 	shmfd->shm_uid = uid;
1227 	shmfd->shm_gid = gid;
1228 out:
1229 	mtx_unlock(&shm_timestamp_lock);
1230 	return (error);
1231 }
1232 
1233 /*
1234  * Helper routines to allow the backing object of a shared memory file
1235  * descriptor to be mapped in the kernel.
1236  */
1237 int
1238 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
1239 {
1240 	struct shmfd *shmfd;
1241 	vm_offset_t kva, ofs;
1242 	vm_object_t obj;
1243 	int rv;
1244 
1245 	if (fp->f_type != DTYPE_SHM)
1246 		return (EINVAL);
1247 	shmfd = fp->f_data;
1248 	obj = shmfd->shm_object;
1249 	VM_OBJECT_WLOCK(obj);
1250 	/*
1251 	 * XXXRW: This validation is probably insufficient, and subject to
1252 	 * sign errors.  It should be fixed.
1253 	 */
1254 	if (offset >= shmfd->shm_size ||
1255 	    offset + size > round_page(shmfd->shm_size)) {
1256 		VM_OBJECT_WUNLOCK(obj);
1257 		return (EINVAL);
1258 	}
1259 
1260 	shmfd->shm_kmappings++;
1261 	vm_object_reference_locked(obj);
1262 	VM_OBJECT_WUNLOCK(obj);
1263 
1264 	/* Map the object into the kernel_map and wire it. */
1265 	kva = vm_map_min(kernel_map);
1266 	ofs = offset & PAGE_MASK;
1267 	offset = trunc_page(offset);
1268 	size = round_page(size + ofs);
1269 	rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1270 	    VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1271 	    VM_PROT_READ | VM_PROT_WRITE, 0);
1272 	if (rv == KERN_SUCCESS) {
1273 		rv = vm_map_wire(kernel_map, kva, kva + size,
1274 		    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1275 		if (rv == KERN_SUCCESS) {
1276 			*memp = (void *)(kva + ofs);
1277 			return (0);
1278 		}
1279 		vm_map_remove(kernel_map, kva, kva + size);
1280 	} else
1281 		vm_object_deallocate(obj);
1282 
1283 	/* On failure, drop our mapping reference. */
1284 	VM_OBJECT_WLOCK(obj);
1285 	shmfd->shm_kmappings--;
1286 	VM_OBJECT_WUNLOCK(obj);
1287 
1288 	return (vm_mmap_to_errno(rv));
1289 }
1290 
1291 /*
1292  * We require the caller to unmap the entire entry.  This allows us to
1293  * safely decrement shm_kmappings when a mapping is removed.
1294  */
1295 int
1296 shm_unmap(struct file *fp, void *mem, size_t size)
1297 {
1298 	struct shmfd *shmfd;
1299 	vm_map_entry_t entry;
1300 	vm_offset_t kva, ofs;
1301 	vm_object_t obj;
1302 	vm_pindex_t pindex;
1303 	vm_prot_t prot;
1304 	boolean_t wired;
1305 	vm_map_t map;
1306 	int rv;
1307 
1308 	if (fp->f_type != DTYPE_SHM)
1309 		return (EINVAL);
1310 	shmfd = fp->f_data;
1311 	kva = (vm_offset_t)mem;
1312 	ofs = kva & PAGE_MASK;
1313 	kva = trunc_page(kva);
1314 	size = round_page(size + ofs);
1315 	map = kernel_map;
1316 	rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1317 	    &obj, &pindex, &prot, &wired);
1318 	if (rv != KERN_SUCCESS)
1319 		return (EINVAL);
1320 	if (entry->start != kva || entry->end != kva + size) {
1321 		vm_map_lookup_done(map, entry);
1322 		return (EINVAL);
1323 	}
1324 	vm_map_lookup_done(map, entry);
1325 	if (obj != shmfd->shm_object)
1326 		return (EINVAL);
1327 	vm_map_remove(map, kva, kva + size);
1328 	VM_OBJECT_WLOCK(obj);
1329 	KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1330 	shmfd->shm_kmappings--;
1331 	VM_OBJECT_WUNLOCK(obj);
1332 	return (0);
1333 }
1334 
1335 static int
1336 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list)
1337 {
1338 	const char *path, *pr_path;
1339 	size_t pr_pathlen;
1340 	bool visible;
1341 
1342 	sx_assert(&shm_dict_lock, SA_LOCKED);
1343 	kif->kf_type = KF_TYPE_SHM;
1344 	kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;
1345 	kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1346 	if (shmfd->shm_path != NULL) {
1347 		if (shmfd->shm_path != NULL) {
1348 			path = shmfd->shm_path;
1349 			pr_path = curthread->td_ucred->cr_prison->pr_path;
1350 			if (strcmp(pr_path, "/") != 0) {
1351 				/* Return the jail-rooted pathname. */
1352 				pr_pathlen = strlen(pr_path);
1353 				visible = strncmp(path, pr_path, pr_pathlen)
1354 				    == 0 && path[pr_pathlen] == '/';
1355 				if (list && !visible)
1356 					return (EPERM);
1357 				if (visible)
1358 					path += pr_pathlen;
1359 			}
1360 			strlcpy(kif->kf_path, path, sizeof(kif->kf_path));
1361 		}
1362 	}
1363 	return (0);
1364 }
1365 
1366 static int
1367 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif,
1368     struct filedesc *fdp __unused)
1369 {
1370 	int res;
1371 
1372 	sx_slock(&shm_dict_lock);
1373 	res = shm_fill_kinfo_locked(fp->f_data, kif, false);
1374 	sx_sunlock(&shm_dict_lock);
1375 	return (res);
1376 }
1377 
1378 static int
1379 shm_add_seals(struct file *fp, int seals)
1380 {
1381 	struct shmfd *shmfd;
1382 	void *rl_cookie;
1383 	vm_ooffset_t writemappings;
1384 	int error, nseals;
1385 
1386 	error = 0;
1387 	shmfd = fp->f_data;
1388 	rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
1389 	    &shmfd->shm_mtx);
1390 
1391 	/* Even already-set seals should result in EPERM. */
1392 	if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) {
1393 		error = EPERM;
1394 		goto out;
1395 	}
1396 	nseals = seals & ~shmfd->shm_seals;
1397 	if ((nseals & F_SEAL_WRITE) != 0) {
1398 		/*
1399 		 * The rangelock above prevents writable mappings from being
1400 		 * added after we've started applying seals.  The RLOCK here
1401 		 * is to avoid torn reads on ILP32 arches as unmapping/reducing
1402 		 * writemappings will be done without a rangelock.
1403 		 */
1404 		VM_OBJECT_RLOCK(shmfd->shm_object);
1405 		writemappings = shmfd->shm_object->un_pager.swp.writemappings;
1406 		VM_OBJECT_RUNLOCK(shmfd->shm_object);
1407 		/* kmappings are also writable */
1408 		if (writemappings > 0) {
1409 			error = EBUSY;
1410 			goto out;
1411 		}
1412 	}
1413 	shmfd->shm_seals |= nseals;
1414 out:
1415 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
1416 	return (error);
1417 }
1418 
1419 static int
1420 shm_get_seals(struct file *fp, int *seals)
1421 {
1422 	struct shmfd *shmfd;
1423 
1424 	shmfd = fp->f_data;
1425 	*seals = shmfd->shm_seals;
1426 	return (0);
1427 }
1428 
1429 static int
1430 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS)
1431 {
1432 	struct shm_mapping *shmm;
1433 	struct sbuf sb;
1434 	struct kinfo_file kif;
1435 	u_long i;
1436 	ssize_t curlen;
1437 	int error, error2;
1438 
1439 	sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req);
1440 	sbuf_clear_flags(&sb, SBUF_INCLUDENUL);
1441 	curlen = 0;
1442 	error = 0;
1443 	sx_slock(&shm_dict_lock);
1444 	for (i = 0; i < shm_hash + 1; i++) {
1445 		LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) {
1446 			error = shm_fill_kinfo_locked(shmm->sm_shmfd,
1447 			    &kif, true);
1448 			if (error == EPERM)
1449 				continue;
1450 			if (error != 0)
1451 				break;
1452 			pack_kinfo(&kif);
1453 			if (req->oldptr != NULL &&
1454 			    kif.kf_structsize + curlen > req->oldlen)
1455 				break;
1456 			error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ?
1457 			    0 : ENOMEM;
1458 			if (error != 0)
1459 				break;
1460 			curlen += kif.kf_structsize;
1461 		}
1462 	}
1463 	sx_sunlock(&shm_dict_lock);
1464 	error2 = sbuf_finish(&sb);
1465 	sbuf_delete(&sb);
1466 	return (error != 0 ? error : error2);
1467 }
1468 
1469 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list,
1470     CTLFLAG_RD | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE,
1471     NULL, 0, sysctl_posix_shm_list, "",
1472     "POSIX SHM list");
1473 
1474 int
1475 kern_shm_open2(struct thread *td, const char *path, int flags, mode_t mode,
1476     int shmflags, const char *name __unused)
1477 {
1478 	int initial_seals;
1479 
1480 	if ((shmflags & ~SHM_ALLOW_SEALING) != 0)
1481 		return (EINVAL);
1482 
1483 	initial_seals = F_SEAL_SEAL;
1484 	if ((shmflags & SHM_ALLOW_SEALING) != 0)
1485 		initial_seals &= ~F_SEAL_SEAL;
1486 	return (kern_shm_open(td, path, flags, mode, NULL, initial_seals));
1487 }
1488 
1489 /*
1490  * This version of the shm_open() interface leaves CLOEXEC behavior up to the
1491  * caller, and libc will enforce it for the traditional shm_open() call.  This
1492  * allows other consumers, like memfd_create(), to opt-in for CLOEXEC.  This
1493  * interface also includes a 'name' argument that is currently unused, but could
1494  * potentially be exported later via some interface for debugging purposes.
1495  * From the kernel's perspective, it is optional.  Individual consumers like
1496  * memfd_create() may require it in order to be compatible with other systems
1497  * implementing the same function.
1498  */
1499 int
1500 sys_shm_open2(struct thread *td, struct shm_open2_args *uap)
1501 {
1502 
1503 	return (kern_shm_open2(td, uap->path, uap->flags, uap->mode,
1504 	    uap->shmflags, uap->name));
1505 }
1506