xref: /freebsd/sys/kern/uipc_shm.c (revision a33774adb3e7e902ca1742227ee070b93d602fc8)
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_activate(m);
213 	vm_page_sunbusy(m);
214 
215 	return (error);
216 }
217 
218 int
219 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
220 {
221 	ssize_t resid;
222 	size_t len;
223 	int error;
224 
225 	error = 0;
226 	while ((resid = uio->uio_resid) > 0) {
227 		if (obj_size <= uio->uio_offset)
228 			break;
229 		len = MIN(obj_size - uio->uio_offset, resid);
230 		if (len == 0)
231 			break;
232 		error = uiomove_object_page(obj, len, uio);
233 		if (error != 0 || resid == uio->uio_resid)
234 			break;
235 	}
236 	return (error);
237 }
238 
239 static int
240 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
241 {
242 	struct shmfd *shmfd;
243 	off_t foffset;
244 	int error;
245 
246 	shmfd = fp->f_data;
247 	foffset = foffset_lock(fp, 0);
248 	error = 0;
249 	switch (whence) {
250 	case L_INCR:
251 		if (foffset < 0 ||
252 		    (offset > 0 && foffset > OFF_MAX - offset)) {
253 			error = EOVERFLOW;
254 			break;
255 		}
256 		offset += foffset;
257 		break;
258 	case L_XTND:
259 		if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
260 			error = EOVERFLOW;
261 			break;
262 		}
263 		offset += shmfd->shm_size;
264 		break;
265 	case L_SET:
266 		break;
267 	default:
268 		error = EINVAL;
269 	}
270 	if (error == 0) {
271 		if (offset < 0 || offset > shmfd->shm_size)
272 			error = EINVAL;
273 		else
274 			td->td_uretoff.tdu_off = offset;
275 	}
276 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
277 	return (error);
278 }
279 
280 static int
281 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
282     int flags, struct thread *td)
283 {
284 	struct shmfd *shmfd;
285 	void *rl_cookie;
286 	int error;
287 
288 	shmfd = fp->f_data;
289 #ifdef MAC
290 	error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
291 	if (error)
292 		return (error);
293 #endif
294 	foffset_lock_uio(fp, uio, flags);
295 	rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset,
296 	    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
297 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
298 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
299 	foffset_unlock_uio(fp, uio, flags);
300 	return (error);
301 }
302 
303 static int
304 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
305     int flags, struct thread *td)
306 {
307 	struct shmfd *shmfd;
308 	void *rl_cookie;
309 	int error;
310 
311 	shmfd = fp->f_data;
312 #ifdef MAC
313 	error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
314 	if (error)
315 		return (error);
316 #endif
317 	foffset_lock_uio(fp, uio, flags);
318 	if ((flags & FOF_OFFSET) == 0) {
319 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
320 		    &shmfd->shm_mtx);
321 	} else {
322 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset,
323 		    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
324 	}
325 	if ((shmfd->shm_seals & F_SEAL_WRITE) != 0)
326 		error = EPERM;
327 	else
328 		error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
329 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
330 	foffset_unlock_uio(fp, uio, flags);
331 	return (error);
332 }
333 
334 static int
335 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
336     struct thread *td)
337 {
338 	struct shmfd *shmfd;
339 #ifdef MAC
340 	int error;
341 #endif
342 
343 	shmfd = fp->f_data;
344 #ifdef MAC
345 	error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
346 	if (error)
347 		return (error);
348 #endif
349 	return (shm_dotruncate(shmfd, length));
350 }
351 
352 int
353 shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
354     struct thread *td)
355 {
356 
357 	switch (com) {
358 	case FIONBIO:
359 	case FIOASYNC:
360 		/*
361 		 * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work,
362 		 * just like it would on an unlinked regular file
363 		 */
364 		return (0);
365 	default:
366 		return (ENOTTY);
367 	}
368 }
369 
370 static int
371 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
372     struct thread *td)
373 {
374 	struct shmfd *shmfd;
375 #ifdef MAC
376 	int error;
377 #endif
378 
379 	shmfd = fp->f_data;
380 
381 #ifdef MAC
382 	error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
383 	if (error)
384 		return (error);
385 #endif
386 
387 	/*
388 	 * Attempt to return sanish values for fstat() on a memory file
389 	 * descriptor.
390 	 */
391 	bzero(sb, sizeof(*sb));
392 	sb->st_blksize = PAGE_SIZE;
393 	sb->st_size = shmfd->shm_size;
394 	sb->st_blocks = howmany(sb->st_size, sb->st_blksize);
395 	mtx_lock(&shm_timestamp_lock);
396 	sb->st_atim = shmfd->shm_atime;
397 	sb->st_ctim = shmfd->shm_ctime;
398 	sb->st_mtim = shmfd->shm_mtime;
399 	sb->st_birthtim = shmfd->shm_birthtime;
400 	sb->st_mode = S_IFREG | shmfd->shm_mode;		/* XXX */
401 	sb->st_uid = shmfd->shm_uid;
402 	sb->st_gid = shmfd->shm_gid;
403 	mtx_unlock(&shm_timestamp_lock);
404 	sb->st_dev = shm_dev_ino;
405 	sb->st_ino = shmfd->shm_ino;
406 	sb->st_nlink = shmfd->shm_object->ref_count;
407 
408 	return (0);
409 }
410 
411 static int
412 shm_close(struct file *fp, struct thread *td)
413 {
414 	struct shmfd *shmfd;
415 
416 	shmfd = fp->f_data;
417 	fp->f_data = NULL;
418 	shm_drop(shmfd);
419 
420 	return (0);
421 }
422 
423 static int
424 shm_copyin_path(struct thread *td, const char *userpath_in, char **path_out) {
425 	int error;
426 	char *path;
427 	const char *pr_path;
428 	size_t pr_pathlen;
429 
430 	path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
431 	pr_path = td->td_ucred->cr_prison->pr_path;
432 
433 	/* Construct a full pathname for jailed callers. */
434 	pr_pathlen = strcmp(pr_path, "/") ==
435 	    0 ? 0 : strlcpy(path, pr_path, MAXPATHLEN);
436 	error = copyinstr(userpath_in, path + pr_pathlen,
437 	    MAXPATHLEN - pr_pathlen, NULL);
438 	if (error != 0)
439 		goto out;
440 
441 #ifdef KTRACE
442 	if (KTRPOINT(curthread, KTR_NAMEI))
443 		ktrnamei(path);
444 #endif
445 
446 	/* Require paths to start with a '/' character. */
447 	if (path[pr_pathlen] != '/') {
448 		error = EINVAL;
449 		goto out;
450 	}
451 
452 	*path_out = path;
453 
454 out:
455 	if (error != 0)
456 		free(path, M_SHMFD);
457 
458 	return (error);
459 }
460 
461 static int
462 shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie)
463 {
464 	vm_object_t object;
465 	vm_page_t m;
466 	vm_pindex_t idx, nobjsize;
467 	vm_ooffset_t delta;
468 	int base, rv;
469 
470 	KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
471 	object = shmfd->shm_object;
472 	VM_OBJECT_ASSERT_WLOCKED(object);
473 	rangelock_cookie_assert(rl_cookie, RA_WLOCKED);
474 	if (length == shmfd->shm_size)
475 		return (0);
476 	nobjsize = OFF_TO_IDX(length + PAGE_MASK);
477 
478 	/* Are we shrinking?  If so, trim the end. */
479 	if (length < shmfd->shm_size) {
480 		if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0)
481 			return (EPERM);
482 
483 		/*
484 		 * Disallow any requests to shrink the size if this
485 		 * object is mapped into the kernel.
486 		 */
487 		if (shmfd->shm_kmappings > 0)
488 			return (EBUSY);
489 
490 		/*
491 		 * Zero the truncated part of the last page.
492 		 */
493 		base = length & PAGE_MASK;
494 		if (base != 0) {
495 			idx = OFF_TO_IDX(length);
496 retry:
497 			m = vm_page_grab(object, idx, VM_ALLOC_NOCREAT);
498 			if (m != NULL) {
499 				MPASS(vm_page_all_valid(m));
500 			} else if (vm_pager_has_page(object, idx, NULL, NULL)) {
501 				m = vm_page_alloc(object, idx,
502 				    VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL);
503 				if (m == NULL)
504 					goto retry;
505 				rv = vm_pager_get_pages(object, &m, 1, NULL,
506 				    NULL);
507 				if (rv == VM_PAGER_OK) {
508 					/*
509 					 * Since the page was not resident,
510 					 * and therefore not recently
511 					 * accessed, immediately enqueue it
512 					 * for asynchronous laundering.  The
513 					 * current operation is not regarded
514 					 * as an access.
515 					 */
516 					vm_page_launder(m);
517 				} else {
518 					vm_page_free(m);
519 					VM_OBJECT_WUNLOCK(object);
520 					return (EIO);
521 				}
522 			}
523 			if (m != NULL) {
524 				pmap_zero_page_area(m, base, PAGE_SIZE - base);
525 				KASSERT(vm_page_all_valid(m),
526 				    ("shm_dotruncate: page %p is invalid", m));
527 				vm_page_set_dirty(m);
528 				vm_page_xunbusy(m);
529 			}
530 		}
531 		delta = IDX_TO_OFF(object->size - nobjsize);
532 
533 		/* Toss in memory pages. */
534 		if (nobjsize < object->size)
535 			vm_object_page_remove(object, nobjsize, object->size,
536 			    0);
537 
538 		/* Toss pages from swap. */
539 		if (object->type == OBJT_SWAP)
540 			swap_pager_freespace(object, nobjsize, delta);
541 
542 		/* Free the swap accounted for shm */
543 		swap_release_by_cred(delta, object->cred);
544 		object->charge -= delta;
545 	} else {
546 		if ((shmfd->shm_seals & F_SEAL_GROW) != 0)
547 			return (EPERM);
548 
549 		/* Try to reserve additional swap space. */
550 		delta = IDX_TO_OFF(nobjsize - object->size);
551 		if (!swap_reserve_by_cred(delta, object->cred))
552 			return (ENOMEM);
553 		object->charge += delta;
554 	}
555 	shmfd->shm_size = length;
556 	mtx_lock(&shm_timestamp_lock);
557 	vfs_timestamp(&shmfd->shm_ctime);
558 	shmfd->shm_mtime = shmfd->shm_ctime;
559 	mtx_unlock(&shm_timestamp_lock);
560 	object->size = nobjsize;
561 	return (0);
562 }
563 
564 int
565 shm_dotruncate(struct shmfd *shmfd, off_t length)
566 {
567 	void *rl_cookie;
568 	int error;
569 
570 	rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
571 	    &shmfd->shm_mtx);
572 	VM_OBJECT_WLOCK(shmfd->shm_object);
573 	error = shm_dotruncate_locked(shmfd, length, rl_cookie);
574 	VM_OBJECT_WUNLOCK(shmfd->shm_object);
575 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
576 	return (error);
577 }
578 
579 /*
580  * shmfd object management including creation and reference counting
581  * routines.
582  */
583 struct shmfd *
584 shm_alloc(struct ucred *ucred, mode_t mode)
585 {
586 	struct shmfd *shmfd;
587 
588 	shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
589 	shmfd->shm_size = 0;
590 	shmfd->shm_uid = ucred->cr_uid;
591 	shmfd->shm_gid = ucred->cr_gid;
592 	shmfd->shm_mode = mode;
593 	shmfd->shm_object = vm_pager_allocate(OBJT_SWAP, NULL,
594 	    shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
595 	KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
596 	vfs_timestamp(&shmfd->shm_birthtime);
597 	shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
598 	    shmfd->shm_birthtime;
599 	shmfd->shm_ino = alloc_unr64(&shm_ino_unr);
600 	refcount_init(&shmfd->shm_refs, 1);
601 	mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
602 	rangelock_init(&shmfd->shm_rl);
603 #ifdef MAC
604 	mac_posixshm_init(shmfd);
605 	mac_posixshm_create(ucred, shmfd);
606 #endif
607 
608 	return (shmfd);
609 }
610 
611 struct shmfd *
612 shm_hold(struct shmfd *shmfd)
613 {
614 
615 	refcount_acquire(&shmfd->shm_refs);
616 	return (shmfd);
617 }
618 
619 void
620 shm_drop(struct shmfd *shmfd)
621 {
622 
623 	if (refcount_release(&shmfd->shm_refs)) {
624 #ifdef MAC
625 		mac_posixshm_destroy(shmfd);
626 #endif
627 		rangelock_destroy(&shmfd->shm_rl);
628 		mtx_destroy(&shmfd->shm_mtx);
629 		vm_object_deallocate(shmfd->shm_object);
630 		free(shmfd, M_SHMFD);
631 	}
632 }
633 
634 /*
635  * Determine if the credentials have sufficient permissions for a
636  * specified combination of FREAD and FWRITE.
637  */
638 int
639 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
640 {
641 	accmode_t accmode;
642 	int error;
643 
644 	accmode = 0;
645 	if (flags & FREAD)
646 		accmode |= VREAD;
647 	if (flags & FWRITE)
648 		accmode |= VWRITE;
649 	mtx_lock(&shm_timestamp_lock);
650 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
651 	    accmode, ucred, NULL);
652 	mtx_unlock(&shm_timestamp_lock);
653 	return (error);
654 }
655 
656 /*
657  * Dictionary management.  We maintain an in-kernel dictionary to map
658  * paths to shmfd objects.  We use the FNV hash on the path to store
659  * the mappings in a hash table.
660  */
661 static void
662 shm_init(void *arg)
663 {
664 
665 	mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
666 	sx_init(&shm_dict_lock, "shm dictionary");
667 	shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
668 	new_unrhdr64(&shm_ino_unr, 1);
669 	shm_dev_ino = devfs_alloc_cdp_inode();
670 	KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
671 }
672 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
673 
674 static struct shmfd *
675 shm_lookup(char *path, Fnv32_t fnv)
676 {
677 	struct shm_mapping *map;
678 
679 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
680 		if (map->sm_fnv != fnv)
681 			continue;
682 		if (strcmp(map->sm_path, path) == 0)
683 			return (map->sm_shmfd);
684 	}
685 
686 	return (NULL);
687 }
688 
689 static void
690 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
691 {
692 	struct shm_mapping *map;
693 
694 	map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
695 	map->sm_path = path;
696 	map->sm_fnv = fnv;
697 	map->sm_shmfd = shm_hold(shmfd);
698 	shmfd->shm_path = path;
699 	LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
700 }
701 
702 static int
703 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
704 {
705 	struct shm_mapping *map;
706 	int error;
707 
708 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
709 		if (map->sm_fnv != fnv)
710 			continue;
711 		if (strcmp(map->sm_path, path) == 0) {
712 #ifdef MAC
713 			error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
714 			if (error)
715 				return (error);
716 #endif
717 			error = shm_access(map->sm_shmfd, ucred,
718 			    FREAD | FWRITE);
719 			if (error)
720 				return (error);
721 			map->sm_shmfd->shm_path = NULL;
722 			LIST_REMOVE(map, sm_link);
723 			shm_drop(map->sm_shmfd);
724 			free(map->sm_path, M_SHMFD);
725 			free(map, M_SHMFD);
726 			return (0);
727 		}
728 	}
729 
730 	return (ENOENT);
731 }
732 
733 int
734 kern_shm_open2(struct thread *td, const char *userpath, int flags, mode_t mode,
735     int shmflags, struct filecaps *fcaps, const char *name __unused)
736 {
737 	struct filedesc *fdp;
738 	struct shmfd *shmfd;
739 	struct file *fp;
740 	char *path;
741 	void *rl_cookie;
742 	Fnv32_t fnv;
743 	mode_t cmode;
744 	int error, fd, initial_seals;
745 
746 	if ((shmflags & ~SHM_ALLOW_SEALING) != 0)
747 		return (EINVAL);
748 
749 	initial_seals = F_SEAL_SEAL;
750 	if ((shmflags & SHM_ALLOW_SEALING) != 0)
751 		initial_seals &= ~F_SEAL_SEAL;
752 
753 #ifdef CAPABILITY_MODE
754 	/*
755 	 * shm_open(2) is only allowed for anonymous objects.
756 	 */
757 	if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON))
758 		return (ECAPMODE);
759 #endif
760 
761 	AUDIT_ARG_FFLAGS(flags);
762 	AUDIT_ARG_MODE(mode);
763 
764 	if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
765 		return (EINVAL);
766 
767 	if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
768 		return (EINVAL);
769 
770 	/*
771 	 * Currently only F_SEAL_SEAL may be set when creating or opening shmfd.
772 	 * If the decision is made later to allow additional seals, care must be
773 	 * taken below to ensure that the seals are properly set if the shmfd
774 	 * already existed -- this currently assumes that only F_SEAL_SEAL can
775 	 * be set and doesn't take further precautions to ensure the validity of
776 	 * the seals being added with respect to current mappings.
777 	 */
778 	if ((initial_seals & ~F_SEAL_SEAL) != 0)
779 		return (EINVAL);
780 
781 	fdp = td->td_proc->p_fd;
782 	cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS;
783 
784 	/*
785 	 * shm_open(2) created shm should always have O_CLOEXEC set, as mandated
786 	 * by POSIX.  We allow it to be unset here so that an in-kernel
787 	 * interface may be written as a thin layer around shm, optionally not
788 	 * setting CLOEXEC.  For shm_open(2), O_CLOEXEC is set unconditionally
789 	 * in sys_shm_open() to keep this implementation compliant.
790 	 */
791 	error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps);
792 	if (error)
793 		return (error);
794 
795 	/* A SHM_ANON path pointer creates an anonymous object. */
796 	if (userpath == SHM_ANON) {
797 		/* A read-only anonymous object is pointless. */
798 		if ((flags & O_ACCMODE) == O_RDONLY) {
799 			fdclose(td, fp, fd);
800 			fdrop(fp, td);
801 			return (EINVAL);
802 		}
803 		shmfd = shm_alloc(td->td_ucred, cmode);
804 		shmfd->shm_seals = initial_seals;
805 	} else {
806 		error = shm_copyin_path(td, userpath, &path);
807 		if (error != 0) {
808 			fdclose(td, fp, fd);
809 			fdrop(fp, td);
810 			return (error);
811 		}
812 
813 		AUDIT_ARG_UPATH1_CANON(path);
814 		fnv = fnv_32_str(path, FNV1_32_INIT);
815 		sx_xlock(&shm_dict_lock);
816 		shmfd = shm_lookup(path, fnv);
817 		if (shmfd == NULL) {
818 			/* Object does not yet exist, create it if requested. */
819 			if (flags & O_CREAT) {
820 #ifdef MAC
821 				error = mac_posixshm_check_create(td->td_ucred,
822 				    path);
823 				if (error == 0) {
824 #endif
825 					shmfd = shm_alloc(td->td_ucred, cmode);
826 					shmfd->shm_seals = initial_seals;
827 					shm_insert(path, fnv, shmfd);
828 #ifdef MAC
829 				}
830 #endif
831 			} else {
832 				free(path, M_SHMFD);
833 				error = ENOENT;
834 			}
835 		} else {
836 			rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
837 			    &shmfd->shm_mtx);
838 
839 			/*
840 			 * kern_shm_open() likely shouldn't ever error out on
841 			 * trying to set a seal that already exists, unlike
842 			 * F_ADD_SEALS.  This would break terribly as
843 			 * shm_open(2) actually sets F_SEAL_SEAL to maintain
844 			 * historical behavior where the underlying file could
845 			 * not be sealed.
846 			 */
847 			initial_seals &= ~shmfd->shm_seals;
848 
849 			/*
850 			 * Object already exists, obtain a new
851 			 * reference if requested and permitted.
852 			 */
853 			free(path, M_SHMFD);
854 
855 			/*
856 			 * initial_seals can't set additional seals if we've
857 			 * already been set F_SEAL_SEAL.  If F_SEAL_SEAL is set,
858 			 * then we've already removed that one from
859 			 * initial_seals.  This is currently redundant as we
860 			 * only allow setting F_SEAL_SEAL at creation time, but
861 			 * it's cheap to check and decreases the effort required
862 			 * to allow additional seals.
863 			 */
864 			if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 &&
865 			    initial_seals != 0)
866 				error = EPERM;
867 			else if ((flags & (O_CREAT | O_EXCL)) ==
868 			    (O_CREAT | O_EXCL))
869 				error = EEXIST;
870 			else {
871 #ifdef MAC
872 				error = mac_posixshm_check_open(td->td_ucred,
873 				    shmfd, FFLAGS(flags & O_ACCMODE));
874 				if (error == 0)
875 #endif
876 				error = shm_access(shmfd, td->td_ucred,
877 				    FFLAGS(flags & O_ACCMODE));
878 			}
879 
880 			/*
881 			 * Truncate the file back to zero length if
882 			 * O_TRUNC was specified and the object was
883 			 * opened with read/write.
884 			 */
885 			if (error == 0 &&
886 			    (flags & (O_ACCMODE | O_TRUNC)) ==
887 			    (O_RDWR | O_TRUNC)) {
888 				VM_OBJECT_WLOCK(shmfd->shm_object);
889 #ifdef MAC
890 				error = mac_posixshm_check_truncate(
891 					td->td_ucred, fp->f_cred, shmfd);
892 				if (error == 0)
893 #endif
894 					error = shm_dotruncate_locked(shmfd, 0,
895 					    rl_cookie);
896 				VM_OBJECT_WUNLOCK(shmfd->shm_object);
897 			}
898 			if (error == 0) {
899 				/*
900 				 * Currently we only allow F_SEAL_SEAL to be
901 				 * set initially.  As noted above, this would
902 				 * need to be reworked should that change.
903 				 */
904 				shmfd->shm_seals |= initial_seals;
905 				shm_hold(shmfd);
906 			}
907 			rangelock_unlock(&shmfd->shm_rl, rl_cookie,
908 			    &shmfd->shm_mtx);
909 		}
910 		sx_xunlock(&shm_dict_lock);
911 
912 		if (error) {
913 			fdclose(td, fp, fd);
914 			fdrop(fp, td);
915 			return (error);
916 		}
917 	}
918 
919 	finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
920 
921 	td->td_retval[0] = fd;
922 	fdrop(fp, td);
923 
924 	return (0);
925 }
926 
927 /* System calls. */
928 #ifdef COMPAT_FREEBSD12
929 int
930 freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap)
931 {
932 
933 	return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC,
934 	    uap->mode, NULL));
935 }
936 #endif
937 
938 int
939 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
940 {
941 	char *path;
942 	Fnv32_t fnv;
943 	int error;
944 
945 	error = shm_copyin_path(td, uap->path, &path);
946 	if (error != 0)
947 		return (error);
948 
949 	AUDIT_ARG_UPATH1_CANON(path);
950 	fnv = fnv_32_str(path, FNV1_32_INIT);
951 	sx_xlock(&shm_dict_lock);
952 	error = shm_remove(path, fnv, td->td_ucred);
953 	sx_xunlock(&shm_dict_lock);
954 	free(path, M_TEMP);
955 
956 	return (error);
957 }
958 
959 int
960 sys_shm_rename(struct thread *td, struct shm_rename_args *uap)
961 {
962 	char *path_from = NULL, *path_to = NULL;
963 	Fnv32_t fnv_from, fnv_to;
964 	struct shmfd *fd_from;
965 	struct shmfd *fd_to;
966 	int error;
967 	int flags;
968 
969 	flags = uap->flags;
970 	AUDIT_ARG_FFLAGS(flags);
971 
972 	/*
973 	 * Make sure the user passed only valid flags.
974 	 * If you add a new flag, please add a new term here.
975 	 */
976 	if ((flags & ~(
977 	    SHM_RENAME_NOREPLACE |
978 	    SHM_RENAME_EXCHANGE
979 	    )) != 0) {
980 		error = EINVAL;
981 		goto out;
982 	}
983 
984 	/*
985 	 * EXCHANGE and NOREPLACE don't quite make sense together. Let's
986 	 * force the user to choose one or the other.
987 	 */
988 	if ((flags & SHM_RENAME_NOREPLACE) != 0 &&
989 	    (flags & SHM_RENAME_EXCHANGE) != 0) {
990 		error = EINVAL;
991 		goto out;
992 	}
993 
994 	/* Renaming to or from anonymous makes no sense */
995 	if (uap->path_from == SHM_ANON || uap->path_to == SHM_ANON) {
996 		error = EINVAL;
997 		goto out;
998 	}
999 
1000 	error = shm_copyin_path(td, uap->path_from, &path_from);
1001 	if (error != 0)
1002 		goto out;
1003 
1004 	error = shm_copyin_path(td, uap->path_to, &path_to);
1005 	if (error != 0)
1006 		goto out;
1007 
1008 	AUDIT_ARG_UPATH1_CANON(path_from);
1009 	AUDIT_ARG_UPATH2_CANON(path_to);
1010 
1011 	/* Rename with from/to equal is a no-op */
1012 	if (strcmp(path_from, path_to) == 0)
1013 		goto out;
1014 
1015 	fnv_from = fnv_32_str(path_from, FNV1_32_INIT);
1016 	fnv_to = fnv_32_str(path_to, FNV1_32_INIT);
1017 
1018 	sx_xlock(&shm_dict_lock);
1019 
1020 	fd_from = shm_lookup(path_from, fnv_from);
1021 	if (fd_from == NULL) {
1022 		error = ENOENT;
1023 		goto out_locked;
1024 	}
1025 
1026 	fd_to = shm_lookup(path_to, fnv_to);
1027 	if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) {
1028 		error = EEXIST;
1029 		goto out_locked;
1030 	}
1031 
1032 	/*
1033 	 * Unconditionally prevents shm_remove from invalidating the 'from'
1034 	 * shm's state.
1035 	 */
1036 	shm_hold(fd_from);
1037 	error = shm_remove(path_from, fnv_from, td->td_ucred);
1038 
1039 	/*
1040 	 * One of my assumptions failed if ENOENT (e.g. locking didn't
1041 	 * protect us)
1042 	 */
1043 	KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s",
1044 	    path_from));
1045 	if (error != 0) {
1046 		shm_drop(fd_from);
1047 		goto out_locked;
1048 	}
1049 
1050 	/*
1051 	 * If we are exchanging, we need to ensure the shm_remove below
1052 	 * doesn't invalidate the dest shm's state.
1053 	 */
1054 	if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL)
1055 		shm_hold(fd_to);
1056 
1057 	/*
1058 	 * NOTE: if path_to is not already in the hash, c'est la vie;
1059 	 * it simply means we have nothing already at path_to to unlink.
1060 	 * That is the ENOENT case.
1061 	 *
1062 	 * If we somehow don't have access to unlink this guy, but
1063 	 * did for the shm at path_from, then relink the shm to path_from
1064 	 * and abort with EACCES.
1065 	 *
1066 	 * All other errors: that is weird; let's relink and abort the
1067 	 * operation.
1068 	 */
1069 	error = shm_remove(path_to, fnv_to, td->td_ucred);
1070 	if (error != 0 && error != ENOENT) {
1071 		shm_insert(path_from, fnv_from, fd_from);
1072 		shm_drop(fd_from);
1073 		/* Don't free path_from now, since the hash references it */
1074 		path_from = NULL;
1075 		goto out_locked;
1076 	}
1077 
1078 	error = 0;
1079 
1080 	shm_insert(path_to, fnv_to, fd_from);
1081 
1082 	/* Don't free path_to now, since the hash references it */
1083 	path_to = NULL;
1084 
1085 	/* We kept a ref when we removed, and incremented again in insert */
1086 	shm_drop(fd_from);
1087 	KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1088 	    fd_from->shm_refs));
1089 
1090 	if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) {
1091 		shm_insert(path_from, fnv_from, fd_to);
1092 		path_from = NULL;
1093 		shm_drop(fd_to);
1094 		KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1095 		    fd_to->shm_refs));
1096 	}
1097 
1098 out_locked:
1099 	sx_xunlock(&shm_dict_lock);
1100 
1101 out:
1102 	free(path_from, M_SHMFD);
1103 	free(path_to, M_SHMFD);
1104 	return (error);
1105 }
1106 
1107 int
1108 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
1109     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
1110     vm_ooffset_t foff, struct thread *td)
1111 {
1112 	struct shmfd *shmfd;
1113 	vm_prot_t maxprot;
1114 	int error;
1115 	bool writecnt;
1116 	void *rl_cookie;
1117 
1118 	shmfd = fp->f_data;
1119 	maxprot = VM_PROT_NONE;
1120 
1121 	rl_cookie = rangelock_rlock(&shmfd->shm_rl, 0, objsize,
1122 	    &shmfd->shm_mtx);
1123 	/* FREAD should always be set. */
1124 	if ((fp->f_flag & FREAD) != 0)
1125 		maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
1126 
1127 	/*
1128 	 * If FWRITE's set, we can allow VM_PROT_WRITE unless it's a shared
1129 	 * mapping with a write seal applied.
1130 	 */
1131 	if ((fp->f_flag & FWRITE) != 0 && ((flags & MAP_SHARED) == 0 ||
1132 	    (shmfd->shm_seals & F_SEAL_WRITE) == 0))
1133 		maxprot |= VM_PROT_WRITE;
1134 
1135 	writecnt = (flags & MAP_SHARED) != 0 && (prot & VM_PROT_WRITE) != 0;
1136 
1137 	if (writecnt && (shmfd->shm_seals & F_SEAL_WRITE) != 0) {
1138 		error = EPERM;
1139 		goto out;
1140 	}
1141 
1142 	/* Don't permit shared writable mappings on read-only descriptors. */
1143 	if (writecnt && (maxprot & VM_PROT_WRITE) == 0) {
1144 		error = EACCES;
1145 		goto out;
1146 	}
1147 	maxprot &= cap_maxprot;
1148 
1149 	/* See comment in vn_mmap(). */
1150 	if (
1151 #ifdef _LP64
1152 	    objsize > OFF_MAX ||
1153 #endif
1154 	    foff < 0 || foff > OFF_MAX - objsize) {
1155 		error = EINVAL;
1156 		goto out;
1157 	}
1158 
1159 #ifdef MAC
1160 	error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
1161 	if (error != 0)
1162 		goto out;
1163 #endif
1164 
1165 	mtx_lock(&shm_timestamp_lock);
1166 	vfs_timestamp(&shmfd->shm_atime);
1167 	mtx_unlock(&shm_timestamp_lock);
1168 	vm_object_reference(shmfd->shm_object);
1169 
1170 	if (writecnt)
1171 		vm_pager_update_writecount(shmfd->shm_object, 0, objsize);
1172 	error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
1173 	    shmfd->shm_object, foff, writecnt, td);
1174 	if (error != 0) {
1175 		if (writecnt)
1176 			vm_pager_release_writecount(shmfd->shm_object, 0,
1177 			    objsize);
1178 		vm_object_deallocate(shmfd->shm_object);
1179 	}
1180 out:
1181 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
1182 	return (error);
1183 }
1184 
1185 static int
1186 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
1187     struct thread *td)
1188 {
1189 	struct shmfd *shmfd;
1190 	int error;
1191 
1192 	error = 0;
1193 	shmfd = fp->f_data;
1194 	mtx_lock(&shm_timestamp_lock);
1195 	/*
1196 	 * SUSv4 says that x bits of permission need not be affected.
1197 	 * Be consistent with our shm_open there.
1198 	 */
1199 #ifdef MAC
1200 	error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
1201 	if (error != 0)
1202 		goto out;
1203 #endif
1204 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid,
1205 	    shmfd->shm_gid, VADMIN, active_cred, NULL);
1206 	if (error != 0)
1207 		goto out;
1208 	shmfd->shm_mode = mode & ACCESSPERMS;
1209 out:
1210 	mtx_unlock(&shm_timestamp_lock);
1211 	return (error);
1212 }
1213 
1214 static int
1215 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1216     struct thread *td)
1217 {
1218 	struct shmfd *shmfd;
1219 	int error;
1220 
1221 	error = 0;
1222 	shmfd = fp->f_data;
1223 	mtx_lock(&shm_timestamp_lock);
1224 #ifdef MAC
1225 	error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
1226 	if (error != 0)
1227 		goto out;
1228 #endif
1229 	if (uid == (uid_t)-1)
1230 		uid = shmfd->shm_uid;
1231 	if (gid == (gid_t)-1)
1232                  gid = shmfd->shm_gid;
1233 	if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
1234 	    (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
1235 	    (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN)))
1236 		goto out;
1237 	shmfd->shm_uid = uid;
1238 	shmfd->shm_gid = gid;
1239 out:
1240 	mtx_unlock(&shm_timestamp_lock);
1241 	return (error);
1242 }
1243 
1244 /*
1245  * Helper routines to allow the backing object of a shared memory file
1246  * descriptor to be mapped in the kernel.
1247  */
1248 int
1249 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
1250 {
1251 	struct shmfd *shmfd;
1252 	vm_offset_t kva, ofs;
1253 	vm_object_t obj;
1254 	int rv;
1255 
1256 	if (fp->f_type != DTYPE_SHM)
1257 		return (EINVAL);
1258 	shmfd = fp->f_data;
1259 	obj = shmfd->shm_object;
1260 	VM_OBJECT_WLOCK(obj);
1261 	/*
1262 	 * XXXRW: This validation is probably insufficient, and subject to
1263 	 * sign errors.  It should be fixed.
1264 	 */
1265 	if (offset >= shmfd->shm_size ||
1266 	    offset + size > round_page(shmfd->shm_size)) {
1267 		VM_OBJECT_WUNLOCK(obj);
1268 		return (EINVAL);
1269 	}
1270 
1271 	shmfd->shm_kmappings++;
1272 	vm_object_reference_locked(obj);
1273 	VM_OBJECT_WUNLOCK(obj);
1274 
1275 	/* Map the object into the kernel_map and wire it. */
1276 	kva = vm_map_min(kernel_map);
1277 	ofs = offset & PAGE_MASK;
1278 	offset = trunc_page(offset);
1279 	size = round_page(size + ofs);
1280 	rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1281 	    VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1282 	    VM_PROT_READ | VM_PROT_WRITE, 0);
1283 	if (rv == KERN_SUCCESS) {
1284 		rv = vm_map_wire(kernel_map, kva, kva + size,
1285 		    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1286 		if (rv == KERN_SUCCESS) {
1287 			*memp = (void *)(kva + ofs);
1288 			return (0);
1289 		}
1290 		vm_map_remove(kernel_map, kva, kva + size);
1291 	} else
1292 		vm_object_deallocate(obj);
1293 
1294 	/* On failure, drop our mapping reference. */
1295 	VM_OBJECT_WLOCK(obj);
1296 	shmfd->shm_kmappings--;
1297 	VM_OBJECT_WUNLOCK(obj);
1298 
1299 	return (vm_mmap_to_errno(rv));
1300 }
1301 
1302 /*
1303  * We require the caller to unmap the entire entry.  This allows us to
1304  * safely decrement shm_kmappings when a mapping is removed.
1305  */
1306 int
1307 shm_unmap(struct file *fp, void *mem, size_t size)
1308 {
1309 	struct shmfd *shmfd;
1310 	vm_map_entry_t entry;
1311 	vm_offset_t kva, ofs;
1312 	vm_object_t obj;
1313 	vm_pindex_t pindex;
1314 	vm_prot_t prot;
1315 	boolean_t wired;
1316 	vm_map_t map;
1317 	int rv;
1318 
1319 	if (fp->f_type != DTYPE_SHM)
1320 		return (EINVAL);
1321 	shmfd = fp->f_data;
1322 	kva = (vm_offset_t)mem;
1323 	ofs = kva & PAGE_MASK;
1324 	kva = trunc_page(kva);
1325 	size = round_page(size + ofs);
1326 	map = kernel_map;
1327 	rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1328 	    &obj, &pindex, &prot, &wired);
1329 	if (rv != KERN_SUCCESS)
1330 		return (EINVAL);
1331 	if (entry->start != kva || entry->end != kva + size) {
1332 		vm_map_lookup_done(map, entry);
1333 		return (EINVAL);
1334 	}
1335 	vm_map_lookup_done(map, entry);
1336 	if (obj != shmfd->shm_object)
1337 		return (EINVAL);
1338 	vm_map_remove(map, kva, kva + size);
1339 	VM_OBJECT_WLOCK(obj);
1340 	KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1341 	shmfd->shm_kmappings--;
1342 	VM_OBJECT_WUNLOCK(obj);
1343 	return (0);
1344 }
1345 
1346 static int
1347 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list)
1348 {
1349 	const char *path, *pr_path;
1350 	size_t pr_pathlen;
1351 	bool visible;
1352 
1353 	sx_assert(&shm_dict_lock, SA_LOCKED);
1354 	kif->kf_type = KF_TYPE_SHM;
1355 	kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;
1356 	kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1357 	if (shmfd->shm_path != NULL) {
1358 		if (shmfd->shm_path != NULL) {
1359 			path = shmfd->shm_path;
1360 			pr_path = curthread->td_ucred->cr_prison->pr_path;
1361 			if (strcmp(pr_path, "/") != 0) {
1362 				/* Return the jail-rooted pathname. */
1363 				pr_pathlen = strlen(pr_path);
1364 				visible = strncmp(path, pr_path, pr_pathlen)
1365 				    == 0 && path[pr_pathlen] == '/';
1366 				if (list && !visible)
1367 					return (EPERM);
1368 				if (visible)
1369 					path += pr_pathlen;
1370 			}
1371 			strlcpy(kif->kf_path, path, sizeof(kif->kf_path));
1372 		}
1373 	}
1374 	return (0);
1375 }
1376 
1377 static int
1378 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif,
1379     struct filedesc *fdp __unused)
1380 {
1381 	int res;
1382 
1383 	sx_slock(&shm_dict_lock);
1384 	res = shm_fill_kinfo_locked(fp->f_data, kif, false);
1385 	sx_sunlock(&shm_dict_lock);
1386 	return (res);
1387 }
1388 
1389 static int
1390 shm_add_seals(struct file *fp, int seals)
1391 {
1392 	struct shmfd *shmfd;
1393 	void *rl_cookie;
1394 	vm_ooffset_t writemappings;
1395 	int error, nseals;
1396 
1397 	error = 0;
1398 	shmfd = fp->f_data;
1399 	rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
1400 	    &shmfd->shm_mtx);
1401 
1402 	/* Even already-set seals should result in EPERM. */
1403 	if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) {
1404 		error = EPERM;
1405 		goto out;
1406 	}
1407 	nseals = seals & ~shmfd->shm_seals;
1408 	if ((nseals & F_SEAL_WRITE) != 0) {
1409 		/*
1410 		 * The rangelock above prevents writable mappings from being
1411 		 * added after we've started applying seals.  The RLOCK here
1412 		 * is to avoid torn reads on ILP32 arches as unmapping/reducing
1413 		 * writemappings will be done without a rangelock.
1414 		 */
1415 		VM_OBJECT_RLOCK(shmfd->shm_object);
1416 		writemappings = shmfd->shm_object->un_pager.swp.writemappings;
1417 		VM_OBJECT_RUNLOCK(shmfd->shm_object);
1418 		/* kmappings are also writable */
1419 		if (writemappings > 0) {
1420 			error = EBUSY;
1421 			goto out;
1422 		}
1423 	}
1424 	shmfd->shm_seals |= nseals;
1425 out:
1426 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
1427 	return (error);
1428 }
1429 
1430 static int
1431 shm_get_seals(struct file *fp, int *seals)
1432 {
1433 	struct shmfd *shmfd;
1434 
1435 	shmfd = fp->f_data;
1436 	*seals = shmfd->shm_seals;
1437 	return (0);
1438 }
1439 
1440 static int
1441 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS)
1442 {
1443 	struct shm_mapping *shmm;
1444 	struct sbuf sb;
1445 	struct kinfo_file kif;
1446 	u_long i;
1447 	ssize_t curlen;
1448 	int error, error2;
1449 
1450 	sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req);
1451 	sbuf_clear_flags(&sb, SBUF_INCLUDENUL);
1452 	curlen = 0;
1453 	error = 0;
1454 	sx_slock(&shm_dict_lock);
1455 	for (i = 0; i < shm_hash + 1; i++) {
1456 		LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) {
1457 			error = shm_fill_kinfo_locked(shmm->sm_shmfd,
1458 			    &kif, true);
1459 			if (error == EPERM)
1460 				continue;
1461 			if (error != 0)
1462 				break;
1463 			pack_kinfo(&kif);
1464 			if (req->oldptr != NULL &&
1465 			    kif.kf_structsize + curlen > req->oldlen)
1466 				break;
1467 			error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ?
1468 			    0 : ENOMEM;
1469 			if (error != 0)
1470 				break;
1471 			curlen += kif.kf_structsize;
1472 		}
1473 	}
1474 	sx_sunlock(&shm_dict_lock);
1475 	error2 = sbuf_finish(&sb);
1476 	sbuf_delete(&sb);
1477 	return (error != 0 ? error : error2);
1478 }
1479 
1480 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list,
1481     CTLFLAG_RD | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE,
1482     NULL, 0, sysctl_posix_shm_list, "",
1483     "POSIX SHM list");
1484 
1485 int
1486 kern_shm_open(struct thread *td, const char *path, int flags, mode_t mode,
1487     struct filecaps *caps)
1488 {
1489 
1490 	return (kern_shm_open2(td, path, flags, mode, 0, caps, NULL));
1491 }
1492 
1493 /*
1494  * This version of the shm_open() interface leaves CLOEXEC behavior up to the
1495  * caller, and libc will enforce it for the traditional shm_open() call.  This
1496  * allows other consumers, like memfd_create(), to opt-in for CLOEXEC.  This
1497  * interface also includes a 'name' argument that is currently unused, but could
1498  * potentially be exported later via some interface for debugging purposes.
1499  * From the kernel's perspective, it is optional.  Individual consumers like
1500  * memfd_create() may require it in order to be compatible with other systems
1501  * implementing the same function.
1502  */
1503 int
1504 sys_shm_open2(struct thread *td, struct shm_open2_args *uap)
1505 {
1506 
1507 	return (kern_shm_open2(td, uap->path, uap->flags, uap->mode,
1508 	    uap->shmflags, NULL, uap->name));
1509 }
1510