xref: /freebsd/sys/kern/uipc_shm.c (revision 5ffd83dbcc34f10e07f6d3e968ae6365869615f4)
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 static fo_fallocate_t	shm_fallocate;
142 
143 /* File descriptor operations. */
144 struct fileops shm_ops = {
145 	.fo_read = shm_read,
146 	.fo_write = shm_write,
147 	.fo_truncate = shm_truncate,
148 	.fo_ioctl = shm_ioctl,
149 	.fo_poll = invfo_poll,
150 	.fo_kqfilter = invfo_kqfilter,
151 	.fo_stat = shm_stat,
152 	.fo_close = shm_close,
153 	.fo_chmod = shm_chmod,
154 	.fo_chown = shm_chown,
155 	.fo_sendfile = vn_sendfile,
156 	.fo_seek = shm_seek,
157 	.fo_fill_kinfo = shm_fill_kinfo,
158 	.fo_mmap = shm_mmap,
159 	.fo_get_seals = shm_get_seals,
160 	.fo_add_seals = shm_add_seals,
161 	.fo_fallocate = shm_fallocate,
162 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
163 };
164 
165 FEATURE(posix_shm, "POSIX shared memory");
166 
167 static int
168 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio)
169 {
170 	vm_page_t m;
171 	vm_pindex_t idx;
172 	size_t tlen;
173 	int error, offset, rv;
174 
175 	idx = OFF_TO_IDX(uio->uio_offset);
176 	offset = uio->uio_offset & PAGE_MASK;
177 	tlen = MIN(PAGE_SIZE - offset, len);
178 
179 	rv = vm_page_grab_valid_unlocked(&m, obj, idx,
180 	    VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOCREAT);
181 	if (rv == VM_PAGER_OK)
182 		goto found;
183 
184 	/*
185 	 * Read I/O without either a corresponding resident page or swap
186 	 * page: use zero_region.  This is intended to avoid instantiating
187 	 * pages on read from a sparse region.
188 	 */
189 	VM_OBJECT_WLOCK(obj);
190 	m = vm_page_lookup(obj, idx);
191 	if (uio->uio_rw == UIO_READ && m == NULL &&
192 	    !vm_pager_has_page(obj, idx, NULL, NULL)) {
193 		VM_OBJECT_WUNLOCK(obj);
194 		return (uiomove(__DECONST(void *, zero_region), tlen, uio));
195 	}
196 
197 	/*
198 	 * Although the tmpfs vnode lock is held here, it is
199 	 * nonetheless safe to sleep waiting for a free page.  The
200 	 * pageout daemon does not need to acquire the tmpfs vnode
201 	 * lock to page out tobj's pages because tobj is a OBJT_SWAP
202 	 * type object.
203 	 */
204 	rv = vm_page_grab_valid(&m, obj, idx,
205 	    VM_ALLOC_NORMAL | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY);
206 	if (rv != VM_PAGER_OK) {
207 		VM_OBJECT_WUNLOCK(obj);
208 		printf("uiomove_object: vm_obj %p idx %jd pager error %d\n",
209 		    obj, idx, rv);
210 		return (EIO);
211 	}
212 	VM_OBJECT_WUNLOCK(obj);
213 
214 found:
215 	error = uiomove_fromphys(&m, offset, tlen, uio);
216 	if (uio->uio_rw == UIO_WRITE && error == 0)
217 		vm_page_set_dirty(m);
218 	vm_page_activate(m);
219 	vm_page_sunbusy(m);
220 
221 	return (error);
222 }
223 
224 int
225 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
226 {
227 	ssize_t resid;
228 	size_t len;
229 	int error;
230 
231 	error = 0;
232 	while ((resid = uio->uio_resid) > 0) {
233 		if (obj_size <= uio->uio_offset)
234 			break;
235 		len = MIN(obj_size - uio->uio_offset, resid);
236 		if (len == 0)
237 			break;
238 		error = uiomove_object_page(obj, len, uio);
239 		if (error != 0 || resid == uio->uio_resid)
240 			break;
241 	}
242 	return (error);
243 }
244 
245 static int
246 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
247 {
248 	struct shmfd *shmfd;
249 	off_t foffset;
250 	int error;
251 
252 	shmfd = fp->f_data;
253 	foffset = foffset_lock(fp, 0);
254 	error = 0;
255 	switch (whence) {
256 	case L_INCR:
257 		if (foffset < 0 ||
258 		    (offset > 0 && foffset > OFF_MAX - offset)) {
259 			error = EOVERFLOW;
260 			break;
261 		}
262 		offset += foffset;
263 		break;
264 	case L_XTND:
265 		if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
266 			error = EOVERFLOW;
267 			break;
268 		}
269 		offset += shmfd->shm_size;
270 		break;
271 	case L_SET:
272 		break;
273 	default:
274 		error = EINVAL;
275 	}
276 	if (error == 0) {
277 		if (offset < 0 || offset > shmfd->shm_size)
278 			error = EINVAL;
279 		else
280 			td->td_uretoff.tdu_off = offset;
281 	}
282 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
283 	return (error);
284 }
285 
286 static int
287 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
288     int flags, struct thread *td)
289 {
290 	struct shmfd *shmfd;
291 	void *rl_cookie;
292 	int error;
293 
294 	shmfd = fp->f_data;
295 #ifdef MAC
296 	error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
297 	if (error)
298 		return (error);
299 #endif
300 	foffset_lock_uio(fp, uio, flags);
301 	rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset,
302 	    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
303 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
304 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
305 	foffset_unlock_uio(fp, uio, flags);
306 	return (error);
307 }
308 
309 static int
310 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
311     int flags, struct thread *td)
312 {
313 	struct shmfd *shmfd;
314 	void *rl_cookie;
315 	int error;
316 	off_t size;
317 
318 	shmfd = fp->f_data;
319 #ifdef MAC
320 	error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
321 	if (error)
322 		return (error);
323 #endif
324 	foffset_lock_uio(fp, uio, flags);
325 	if (uio->uio_resid > OFF_MAX - uio->uio_offset) {
326 		/*
327 		 * Overflow is only an error if we're supposed to expand on
328 		 * write.  Otherwise, we'll just truncate the write to the
329 		 * size of the file, which can only grow up to OFF_MAX.
330 		 */
331 		if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0) {
332 			foffset_unlock_uio(fp, uio, flags);
333 			return (EFBIG);
334 		}
335 
336 		size = shmfd->shm_size;
337 	} else {
338 		size = uio->uio_offset + uio->uio_resid;
339 	}
340 	if ((flags & FOF_OFFSET) == 0) {
341 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
342 		    &shmfd->shm_mtx);
343 	} else {
344 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset,
345 		    size, &shmfd->shm_mtx);
346 	}
347 	if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) {
348 		error = EPERM;
349 	} else {
350 		error = 0;
351 		if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0 &&
352 		    size > shmfd->shm_size) {
353 			VM_OBJECT_WLOCK(shmfd->shm_object);
354 			error = shm_dotruncate_locked(shmfd, size, rl_cookie);
355 			VM_OBJECT_WUNLOCK(shmfd->shm_object);
356 		}
357 		if (error == 0)
358 			error = uiomove_object(shmfd->shm_object,
359 			    shmfd->shm_size, uio);
360 	}
361 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
362 	foffset_unlock_uio(fp, uio, flags);
363 	return (error);
364 }
365 
366 static int
367 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
368     struct thread *td)
369 {
370 	struct shmfd *shmfd;
371 #ifdef MAC
372 	int error;
373 #endif
374 
375 	shmfd = fp->f_data;
376 #ifdef MAC
377 	error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
378 	if (error)
379 		return (error);
380 #endif
381 	return (shm_dotruncate(shmfd, length));
382 }
383 
384 int
385 shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
386     struct thread *td)
387 {
388 
389 	switch (com) {
390 	case FIONBIO:
391 	case FIOASYNC:
392 		/*
393 		 * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work,
394 		 * just like it would on an unlinked regular file
395 		 */
396 		return (0);
397 	default:
398 		return (ENOTTY);
399 	}
400 }
401 
402 static int
403 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
404     struct thread *td)
405 {
406 	struct shmfd *shmfd;
407 #ifdef MAC
408 	int error;
409 #endif
410 
411 	shmfd = fp->f_data;
412 
413 #ifdef MAC
414 	error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
415 	if (error)
416 		return (error);
417 #endif
418 
419 	/*
420 	 * Attempt to return sanish values for fstat() on a memory file
421 	 * descriptor.
422 	 */
423 	bzero(sb, sizeof(*sb));
424 	sb->st_blksize = PAGE_SIZE;
425 	sb->st_size = shmfd->shm_size;
426 	sb->st_blocks = howmany(sb->st_size, sb->st_blksize);
427 	mtx_lock(&shm_timestamp_lock);
428 	sb->st_atim = shmfd->shm_atime;
429 	sb->st_ctim = shmfd->shm_ctime;
430 	sb->st_mtim = shmfd->shm_mtime;
431 	sb->st_birthtim = shmfd->shm_birthtime;
432 	sb->st_mode = S_IFREG | shmfd->shm_mode;		/* XXX */
433 	sb->st_uid = shmfd->shm_uid;
434 	sb->st_gid = shmfd->shm_gid;
435 	mtx_unlock(&shm_timestamp_lock);
436 	sb->st_dev = shm_dev_ino;
437 	sb->st_ino = shmfd->shm_ino;
438 	sb->st_nlink = shmfd->shm_object->ref_count;
439 
440 	return (0);
441 }
442 
443 static int
444 shm_close(struct file *fp, struct thread *td)
445 {
446 	struct shmfd *shmfd;
447 
448 	shmfd = fp->f_data;
449 	fp->f_data = NULL;
450 	shm_drop(shmfd);
451 
452 	return (0);
453 }
454 
455 static int
456 shm_copyin_path(struct thread *td, const char *userpath_in, char **path_out) {
457 	int error;
458 	char *path;
459 	const char *pr_path;
460 	size_t pr_pathlen;
461 
462 	path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
463 	pr_path = td->td_ucred->cr_prison->pr_path;
464 
465 	/* Construct a full pathname for jailed callers. */
466 	pr_pathlen = strcmp(pr_path, "/") ==
467 	    0 ? 0 : strlcpy(path, pr_path, MAXPATHLEN);
468 	error = copyinstr(userpath_in, path + pr_pathlen,
469 	    MAXPATHLEN - pr_pathlen, NULL);
470 	if (error != 0)
471 		goto out;
472 
473 #ifdef KTRACE
474 	if (KTRPOINT(curthread, KTR_NAMEI))
475 		ktrnamei(path);
476 #endif
477 
478 	/* Require paths to start with a '/' character. */
479 	if (path[pr_pathlen] != '/') {
480 		error = EINVAL;
481 		goto out;
482 	}
483 
484 	*path_out = path;
485 
486 out:
487 	if (error != 0)
488 		free(path, M_SHMFD);
489 
490 	return (error);
491 }
492 
493 static int
494 shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie)
495 {
496 	vm_object_t object;
497 	vm_page_t m;
498 	vm_pindex_t idx, nobjsize;
499 	vm_ooffset_t delta;
500 	int base, rv;
501 
502 	KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
503 	object = shmfd->shm_object;
504 	VM_OBJECT_ASSERT_WLOCKED(object);
505 	rangelock_cookie_assert(rl_cookie, RA_WLOCKED);
506 	if (length == shmfd->shm_size)
507 		return (0);
508 	nobjsize = OFF_TO_IDX(length + PAGE_MASK);
509 
510 	/* Are we shrinking?  If so, trim the end. */
511 	if (length < shmfd->shm_size) {
512 		if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0)
513 			return (EPERM);
514 
515 		/*
516 		 * Disallow any requests to shrink the size if this
517 		 * object is mapped into the kernel.
518 		 */
519 		if (shmfd->shm_kmappings > 0)
520 			return (EBUSY);
521 
522 		/*
523 		 * Zero the truncated part of the last page.
524 		 */
525 		base = length & PAGE_MASK;
526 		if (base != 0) {
527 			idx = OFF_TO_IDX(length);
528 retry:
529 			m = vm_page_grab(object, idx, VM_ALLOC_NOCREAT);
530 			if (m != NULL) {
531 				MPASS(vm_page_all_valid(m));
532 			} else if (vm_pager_has_page(object, idx, NULL, NULL)) {
533 				m = vm_page_alloc(object, idx,
534 				    VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL);
535 				if (m == NULL)
536 					goto retry;
537 				vm_object_pip_add(object, 1);
538 				VM_OBJECT_WUNLOCK(object);
539 				rv = vm_pager_get_pages(object, &m, 1, NULL,
540 				    NULL);
541 				VM_OBJECT_WLOCK(object);
542 				vm_object_pip_wakeup(object);
543 				if (rv == VM_PAGER_OK) {
544 					/*
545 					 * Since the page was not resident,
546 					 * and therefore not recently
547 					 * accessed, immediately enqueue it
548 					 * for asynchronous laundering.  The
549 					 * current operation is not regarded
550 					 * as an access.
551 					 */
552 					vm_page_launder(m);
553 				} else {
554 					vm_page_free(m);
555 					VM_OBJECT_WUNLOCK(object);
556 					return (EIO);
557 				}
558 			}
559 			if (m != NULL) {
560 				pmap_zero_page_area(m, base, PAGE_SIZE - base);
561 				KASSERT(vm_page_all_valid(m),
562 				    ("shm_dotruncate: page %p is invalid", m));
563 				vm_page_set_dirty(m);
564 				vm_page_xunbusy(m);
565 			}
566 		}
567 		delta = IDX_TO_OFF(object->size - nobjsize);
568 
569 		if (nobjsize < object->size)
570 			vm_object_page_remove(object, nobjsize, object->size,
571 			    0);
572 
573 		/* Free the swap accounted for shm */
574 		swap_release_by_cred(delta, object->cred);
575 		object->charge -= delta;
576 	} else {
577 		if ((shmfd->shm_seals & F_SEAL_GROW) != 0)
578 			return (EPERM);
579 
580 		/* Try to reserve additional swap space. */
581 		delta = IDX_TO_OFF(nobjsize - object->size);
582 		if (!swap_reserve_by_cred(delta, object->cred))
583 			return (ENOMEM);
584 		object->charge += delta;
585 	}
586 	shmfd->shm_size = length;
587 	mtx_lock(&shm_timestamp_lock);
588 	vfs_timestamp(&shmfd->shm_ctime);
589 	shmfd->shm_mtime = shmfd->shm_ctime;
590 	mtx_unlock(&shm_timestamp_lock);
591 	object->size = nobjsize;
592 	return (0);
593 }
594 
595 int
596 shm_dotruncate(struct shmfd *shmfd, off_t length)
597 {
598 	void *rl_cookie;
599 	int error;
600 
601 	rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
602 	    &shmfd->shm_mtx);
603 	VM_OBJECT_WLOCK(shmfd->shm_object);
604 	error = shm_dotruncate_locked(shmfd, length, rl_cookie);
605 	VM_OBJECT_WUNLOCK(shmfd->shm_object);
606 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
607 	return (error);
608 }
609 
610 /*
611  * shmfd object management including creation and reference counting
612  * routines.
613  */
614 struct shmfd *
615 shm_alloc(struct ucred *ucred, mode_t mode)
616 {
617 	struct shmfd *shmfd;
618 
619 	shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
620 	shmfd->shm_size = 0;
621 	shmfd->shm_uid = ucred->cr_uid;
622 	shmfd->shm_gid = ucred->cr_gid;
623 	shmfd->shm_mode = mode;
624 	shmfd->shm_object = vm_pager_allocate(OBJT_SWAP, NULL,
625 	    shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
626 	KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
627 	vfs_timestamp(&shmfd->shm_birthtime);
628 	shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
629 	    shmfd->shm_birthtime;
630 	shmfd->shm_ino = alloc_unr64(&shm_ino_unr);
631 	refcount_init(&shmfd->shm_refs, 1);
632 	mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
633 	rangelock_init(&shmfd->shm_rl);
634 #ifdef MAC
635 	mac_posixshm_init(shmfd);
636 	mac_posixshm_create(ucred, shmfd);
637 #endif
638 
639 	return (shmfd);
640 }
641 
642 struct shmfd *
643 shm_hold(struct shmfd *shmfd)
644 {
645 
646 	refcount_acquire(&shmfd->shm_refs);
647 	return (shmfd);
648 }
649 
650 void
651 shm_drop(struct shmfd *shmfd)
652 {
653 
654 	if (refcount_release(&shmfd->shm_refs)) {
655 #ifdef MAC
656 		mac_posixshm_destroy(shmfd);
657 #endif
658 		rangelock_destroy(&shmfd->shm_rl);
659 		mtx_destroy(&shmfd->shm_mtx);
660 		vm_object_deallocate(shmfd->shm_object);
661 		free(shmfd, M_SHMFD);
662 	}
663 }
664 
665 /*
666  * Determine if the credentials have sufficient permissions for a
667  * specified combination of FREAD and FWRITE.
668  */
669 int
670 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
671 {
672 	accmode_t accmode;
673 	int error;
674 
675 	accmode = 0;
676 	if (flags & FREAD)
677 		accmode |= VREAD;
678 	if (flags & FWRITE)
679 		accmode |= VWRITE;
680 	mtx_lock(&shm_timestamp_lock);
681 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
682 	    accmode, ucred, NULL);
683 	mtx_unlock(&shm_timestamp_lock);
684 	return (error);
685 }
686 
687 /*
688  * Dictionary management.  We maintain an in-kernel dictionary to map
689  * paths to shmfd objects.  We use the FNV hash on the path to store
690  * the mappings in a hash table.
691  */
692 static void
693 shm_init(void *arg)
694 {
695 
696 	mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
697 	sx_init(&shm_dict_lock, "shm dictionary");
698 	shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
699 	new_unrhdr64(&shm_ino_unr, 1);
700 	shm_dev_ino = devfs_alloc_cdp_inode();
701 	KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
702 }
703 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
704 
705 static struct shmfd *
706 shm_lookup(char *path, Fnv32_t fnv)
707 {
708 	struct shm_mapping *map;
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 			return (map->sm_shmfd);
715 	}
716 
717 	return (NULL);
718 }
719 
720 static void
721 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
722 {
723 	struct shm_mapping *map;
724 
725 	map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
726 	map->sm_path = path;
727 	map->sm_fnv = fnv;
728 	map->sm_shmfd = shm_hold(shmfd);
729 	shmfd->shm_path = path;
730 	LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
731 }
732 
733 static int
734 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
735 {
736 	struct shm_mapping *map;
737 	int error;
738 
739 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
740 		if (map->sm_fnv != fnv)
741 			continue;
742 		if (strcmp(map->sm_path, path) == 0) {
743 #ifdef MAC
744 			error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
745 			if (error)
746 				return (error);
747 #endif
748 			error = shm_access(map->sm_shmfd, ucred,
749 			    FREAD | FWRITE);
750 			if (error)
751 				return (error);
752 			map->sm_shmfd->shm_path = NULL;
753 			LIST_REMOVE(map, sm_link);
754 			shm_drop(map->sm_shmfd);
755 			free(map->sm_path, M_SHMFD);
756 			free(map, M_SHMFD);
757 			return (0);
758 		}
759 	}
760 
761 	return (ENOENT);
762 }
763 
764 int
765 kern_shm_open2(struct thread *td, const char *userpath, int flags, mode_t mode,
766     int shmflags, struct filecaps *fcaps, const char *name __unused)
767 {
768 	struct filedesc *fdp;
769 	struct shmfd *shmfd;
770 	struct file *fp;
771 	char *path;
772 	void *rl_cookie;
773 	Fnv32_t fnv;
774 	mode_t cmode;
775 	int error, fd, initial_seals;
776 
777 	if ((shmflags & ~(SHM_ALLOW_SEALING | SHM_GROW_ON_WRITE)) != 0)
778 		return (EINVAL);
779 
780 	initial_seals = F_SEAL_SEAL;
781 	if ((shmflags & SHM_ALLOW_SEALING) != 0)
782 		initial_seals &= ~F_SEAL_SEAL;
783 
784 #ifdef CAPABILITY_MODE
785 	/*
786 	 * shm_open(2) is only allowed for anonymous objects.
787 	 */
788 	if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON))
789 		return (ECAPMODE);
790 #endif
791 
792 	AUDIT_ARG_FFLAGS(flags);
793 	AUDIT_ARG_MODE(mode);
794 
795 	if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
796 		return (EINVAL);
797 
798 	if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
799 		return (EINVAL);
800 
801 	/*
802 	 * Currently only F_SEAL_SEAL may be set when creating or opening shmfd.
803 	 * If the decision is made later to allow additional seals, care must be
804 	 * taken below to ensure that the seals are properly set if the shmfd
805 	 * already existed -- this currently assumes that only F_SEAL_SEAL can
806 	 * be set and doesn't take further precautions to ensure the validity of
807 	 * the seals being added with respect to current mappings.
808 	 */
809 	if ((initial_seals & ~F_SEAL_SEAL) != 0)
810 		return (EINVAL);
811 
812 	fdp = td->td_proc->p_fd;
813 	cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS;
814 
815 	/*
816 	 * shm_open(2) created shm should always have O_CLOEXEC set, as mandated
817 	 * by POSIX.  We allow it to be unset here so that an in-kernel
818 	 * interface may be written as a thin layer around shm, optionally not
819 	 * setting CLOEXEC.  For shm_open(2), O_CLOEXEC is set unconditionally
820 	 * in sys_shm_open() to keep this implementation compliant.
821 	 */
822 	error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps);
823 	if (error)
824 		return (error);
825 
826 	/* A SHM_ANON path pointer creates an anonymous object. */
827 	if (userpath == SHM_ANON) {
828 		/* A read-only anonymous object is pointless. */
829 		if ((flags & O_ACCMODE) == O_RDONLY) {
830 			fdclose(td, fp, fd);
831 			fdrop(fp, td);
832 			return (EINVAL);
833 		}
834 		shmfd = shm_alloc(td->td_ucred, cmode);
835 		shmfd->shm_seals = initial_seals;
836 	} else {
837 		error = shm_copyin_path(td, userpath, &path);
838 		if (error != 0) {
839 			fdclose(td, fp, fd);
840 			fdrop(fp, td);
841 			return (error);
842 		}
843 
844 		AUDIT_ARG_UPATH1_CANON(path);
845 		fnv = fnv_32_str(path, FNV1_32_INIT);
846 		sx_xlock(&shm_dict_lock);
847 		shmfd = shm_lookup(path, fnv);
848 		if (shmfd == NULL) {
849 			/* Object does not yet exist, create it if requested. */
850 			if (flags & O_CREAT) {
851 #ifdef MAC
852 				error = mac_posixshm_check_create(td->td_ucred,
853 				    path);
854 				if (error == 0) {
855 #endif
856 					shmfd = shm_alloc(td->td_ucred, cmode);
857 					shmfd->shm_seals = initial_seals;
858 					shm_insert(path, fnv, shmfd);
859 #ifdef MAC
860 				}
861 #endif
862 			} else {
863 				free(path, M_SHMFD);
864 				error = ENOENT;
865 			}
866 		} else {
867 			rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
868 			    &shmfd->shm_mtx);
869 
870 			/*
871 			 * kern_shm_open() likely shouldn't ever error out on
872 			 * trying to set a seal that already exists, unlike
873 			 * F_ADD_SEALS.  This would break terribly as
874 			 * shm_open(2) actually sets F_SEAL_SEAL to maintain
875 			 * historical behavior where the underlying file could
876 			 * not be sealed.
877 			 */
878 			initial_seals &= ~shmfd->shm_seals;
879 
880 			/*
881 			 * Object already exists, obtain a new
882 			 * reference if requested and permitted.
883 			 */
884 			free(path, M_SHMFD);
885 
886 			/*
887 			 * initial_seals can't set additional seals if we've
888 			 * already been set F_SEAL_SEAL.  If F_SEAL_SEAL is set,
889 			 * then we've already removed that one from
890 			 * initial_seals.  This is currently redundant as we
891 			 * only allow setting F_SEAL_SEAL at creation time, but
892 			 * it's cheap to check and decreases the effort required
893 			 * to allow additional seals.
894 			 */
895 			if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 &&
896 			    initial_seals != 0)
897 				error = EPERM;
898 			else if ((flags & (O_CREAT | O_EXCL)) ==
899 			    (O_CREAT | O_EXCL))
900 				error = EEXIST;
901 			else {
902 #ifdef MAC
903 				error = mac_posixshm_check_open(td->td_ucred,
904 				    shmfd, FFLAGS(flags & O_ACCMODE));
905 				if (error == 0)
906 #endif
907 				error = shm_access(shmfd, td->td_ucred,
908 				    FFLAGS(flags & O_ACCMODE));
909 			}
910 
911 			/*
912 			 * Truncate the file back to zero length if
913 			 * O_TRUNC was specified and the object was
914 			 * opened with read/write.
915 			 */
916 			if (error == 0 &&
917 			    (flags & (O_ACCMODE | O_TRUNC)) ==
918 			    (O_RDWR | O_TRUNC)) {
919 				VM_OBJECT_WLOCK(shmfd->shm_object);
920 #ifdef MAC
921 				error = mac_posixshm_check_truncate(
922 					td->td_ucred, fp->f_cred, shmfd);
923 				if (error == 0)
924 #endif
925 					error = shm_dotruncate_locked(shmfd, 0,
926 					    rl_cookie);
927 				VM_OBJECT_WUNLOCK(shmfd->shm_object);
928 			}
929 			if (error == 0) {
930 				/*
931 				 * Currently we only allow F_SEAL_SEAL to be
932 				 * set initially.  As noted above, this would
933 				 * need to be reworked should that change.
934 				 */
935 				shmfd->shm_seals |= initial_seals;
936 				shm_hold(shmfd);
937 			}
938 			rangelock_unlock(&shmfd->shm_rl, rl_cookie,
939 			    &shmfd->shm_mtx);
940 		}
941 		sx_xunlock(&shm_dict_lock);
942 
943 		if (error) {
944 			fdclose(td, fp, fd);
945 			fdrop(fp, td);
946 			return (error);
947 		}
948 	}
949 
950 	shmfd->shm_flags = shmflags;
951 	finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
952 
953 	td->td_retval[0] = fd;
954 	fdrop(fp, td);
955 
956 	return (0);
957 }
958 
959 /* System calls. */
960 #ifdef COMPAT_FREEBSD12
961 int
962 freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap)
963 {
964 
965 	return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC,
966 	    uap->mode, NULL));
967 }
968 #endif
969 
970 int
971 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
972 {
973 	char *path;
974 	Fnv32_t fnv;
975 	int error;
976 
977 	error = shm_copyin_path(td, uap->path, &path);
978 	if (error != 0)
979 		return (error);
980 
981 	AUDIT_ARG_UPATH1_CANON(path);
982 	fnv = fnv_32_str(path, FNV1_32_INIT);
983 	sx_xlock(&shm_dict_lock);
984 	error = shm_remove(path, fnv, td->td_ucred);
985 	sx_xunlock(&shm_dict_lock);
986 	free(path, M_SHMFD);
987 
988 	return (error);
989 }
990 
991 int
992 sys_shm_rename(struct thread *td, struct shm_rename_args *uap)
993 {
994 	char *path_from = NULL, *path_to = NULL;
995 	Fnv32_t fnv_from, fnv_to;
996 	struct shmfd *fd_from;
997 	struct shmfd *fd_to;
998 	int error;
999 	int flags;
1000 
1001 	flags = uap->flags;
1002 	AUDIT_ARG_FFLAGS(flags);
1003 
1004 	/*
1005 	 * Make sure the user passed only valid flags.
1006 	 * If you add a new flag, please add a new term here.
1007 	 */
1008 	if ((flags & ~(
1009 	    SHM_RENAME_NOREPLACE |
1010 	    SHM_RENAME_EXCHANGE
1011 	    )) != 0) {
1012 		error = EINVAL;
1013 		goto out;
1014 	}
1015 
1016 	/*
1017 	 * EXCHANGE and NOREPLACE don't quite make sense together. Let's
1018 	 * force the user to choose one or the other.
1019 	 */
1020 	if ((flags & SHM_RENAME_NOREPLACE) != 0 &&
1021 	    (flags & SHM_RENAME_EXCHANGE) != 0) {
1022 		error = EINVAL;
1023 		goto out;
1024 	}
1025 
1026 	/* Renaming to or from anonymous makes no sense */
1027 	if (uap->path_from == SHM_ANON || uap->path_to == SHM_ANON) {
1028 		error = EINVAL;
1029 		goto out;
1030 	}
1031 
1032 	error = shm_copyin_path(td, uap->path_from, &path_from);
1033 	if (error != 0)
1034 		goto out;
1035 
1036 	error = shm_copyin_path(td, uap->path_to, &path_to);
1037 	if (error != 0)
1038 		goto out;
1039 
1040 	AUDIT_ARG_UPATH1_CANON(path_from);
1041 	AUDIT_ARG_UPATH2_CANON(path_to);
1042 
1043 	/* Rename with from/to equal is a no-op */
1044 	if (strcmp(path_from, path_to) == 0)
1045 		goto out;
1046 
1047 	fnv_from = fnv_32_str(path_from, FNV1_32_INIT);
1048 	fnv_to = fnv_32_str(path_to, FNV1_32_INIT);
1049 
1050 	sx_xlock(&shm_dict_lock);
1051 
1052 	fd_from = shm_lookup(path_from, fnv_from);
1053 	if (fd_from == NULL) {
1054 		error = ENOENT;
1055 		goto out_locked;
1056 	}
1057 
1058 	fd_to = shm_lookup(path_to, fnv_to);
1059 	if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) {
1060 		error = EEXIST;
1061 		goto out_locked;
1062 	}
1063 
1064 	/*
1065 	 * Unconditionally prevents shm_remove from invalidating the 'from'
1066 	 * shm's state.
1067 	 */
1068 	shm_hold(fd_from);
1069 	error = shm_remove(path_from, fnv_from, td->td_ucred);
1070 
1071 	/*
1072 	 * One of my assumptions failed if ENOENT (e.g. locking didn't
1073 	 * protect us)
1074 	 */
1075 	KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s",
1076 	    path_from));
1077 	if (error != 0) {
1078 		shm_drop(fd_from);
1079 		goto out_locked;
1080 	}
1081 
1082 	/*
1083 	 * If we are exchanging, we need to ensure the shm_remove below
1084 	 * doesn't invalidate the dest shm's state.
1085 	 */
1086 	if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL)
1087 		shm_hold(fd_to);
1088 
1089 	/*
1090 	 * NOTE: if path_to is not already in the hash, c'est la vie;
1091 	 * it simply means we have nothing already at path_to to unlink.
1092 	 * That is the ENOENT case.
1093 	 *
1094 	 * If we somehow don't have access to unlink this guy, but
1095 	 * did for the shm at path_from, then relink the shm to path_from
1096 	 * and abort with EACCES.
1097 	 *
1098 	 * All other errors: that is weird; let's relink and abort the
1099 	 * operation.
1100 	 */
1101 	error = shm_remove(path_to, fnv_to, td->td_ucred);
1102 	if (error != 0 && error != ENOENT) {
1103 		shm_insert(path_from, fnv_from, fd_from);
1104 		shm_drop(fd_from);
1105 		/* Don't free path_from now, since the hash references it */
1106 		path_from = NULL;
1107 		goto out_locked;
1108 	}
1109 
1110 	error = 0;
1111 
1112 	shm_insert(path_to, fnv_to, fd_from);
1113 
1114 	/* Don't free path_to now, since the hash references it */
1115 	path_to = NULL;
1116 
1117 	/* We kept a ref when we removed, and incremented again in insert */
1118 	shm_drop(fd_from);
1119 	KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1120 	    fd_from->shm_refs));
1121 
1122 	if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) {
1123 		shm_insert(path_from, fnv_from, fd_to);
1124 		path_from = NULL;
1125 		shm_drop(fd_to);
1126 		KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1127 		    fd_to->shm_refs));
1128 	}
1129 
1130 out_locked:
1131 	sx_xunlock(&shm_dict_lock);
1132 
1133 out:
1134 	free(path_from, M_SHMFD);
1135 	free(path_to, M_SHMFD);
1136 	return (error);
1137 }
1138 
1139 int
1140 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
1141     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
1142     vm_ooffset_t foff, struct thread *td)
1143 {
1144 	struct shmfd *shmfd;
1145 	vm_prot_t maxprot;
1146 	int error;
1147 	bool writecnt;
1148 	void *rl_cookie;
1149 
1150 	shmfd = fp->f_data;
1151 	maxprot = VM_PROT_NONE;
1152 
1153 	rl_cookie = rangelock_rlock(&shmfd->shm_rl, 0, objsize,
1154 	    &shmfd->shm_mtx);
1155 	/* FREAD should always be set. */
1156 	if ((fp->f_flag & FREAD) != 0)
1157 		maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
1158 
1159 	/*
1160 	 * If FWRITE's set, we can allow VM_PROT_WRITE unless it's a shared
1161 	 * mapping with a write seal applied.  Private mappings are always
1162 	 * writeable.
1163 	 */
1164 	if ((flags & MAP_SHARED) == 0) {
1165 		cap_maxprot |= VM_PROT_WRITE;
1166 		maxprot |= VM_PROT_WRITE;
1167 		writecnt = false;
1168 	} else {
1169 		if ((fp->f_flag & FWRITE) != 0 &&
1170 		    (shmfd->shm_seals & F_SEAL_WRITE) == 0)
1171 			maxprot |= VM_PROT_WRITE;
1172 
1173 		/*
1174 		 * Any mappings from a writable descriptor may be upgraded to
1175 		 * VM_PROT_WRITE with mprotect(2), unless a write-seal was
1176 		 * applied between the open and subsequent mmap(2).  We want to
1177 		 * reject application of a write seal as long as any such
1178 		 * mapping exists so that the seal cannot be trivially bypassed.
1179 		 */
1180 		writecnt = (maxprot & VM_PROT_WRITE) != 0;
1181 		if (!writecnt && (prot & VM_PROT_WRITE) != 0) {
1182 			error = EACCES;
1183 			goto out;
1184 		}
1185 	}
1186 	maxprot &= cap_maxprot;
1187 
1188 	/* See comment in vn_mmap(). */
1189 	if (
1190 #ifdef _LP64
1191 	    objsize > OFF_MAX ||
1192 #endif
1193 	    foff < 0 || foff > OFF_MAX - objsize) {
1194 		error = EINVAL;
1195 		goto out;
1196 	}
1197 
1198 #ifdef MAC
1199 	error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
1200 	if (error != 0)
1201 		goto out;
1202 #endif
1203 
1204 	mtx_lock(&shm_timestamp_lock);
1205 	vfs_timestamp(&shmfd->shm_atime);
1206 	mtx_unlock(&shm_timestamp_lock);
1207 	vm_object_reference(shmfd->shm_object);
1208 
1209 	if (writecnt)
1210 		vm_pager_update_writecount(shmfd->shm_object, 0, objsize);
1211 	error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
1212 	    shmfd->shm_object, foff, writecnt, td);
1213 	if (error != 0) {
1214 		if (writecnt)
1215 			vm_pager_release_writecount(shmfd->shm_object, 0,
1216 			    objsize);
1217 		vm_object_deallocate(shmfd->shm_object);
1218 	}
1219 out:
1220 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
1221 	return (error);
1222 }
1223 
1224 static int
1225 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
1226     struct thread *td)
1227 {
1228 	struct shmfd *shmfd;
1229 	int error;
1230 
1231 	error = 0;
1232 	shmfd = fp->f_data;
1233 	mtx_lock(&shm_timestamp_lock);
1234 	/*
1235 	 * SUSv4 says that x bits of permission need not be affected.
1236 	 * Be consistent with our shm_open there.
1237 	 */
1238 #ifdef MAC
1239 	error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
1240 	if (error != 0)
1241 		goto out;
1242 #endif
1243 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid,
1244 	    shmfd->shm_gid, VADMIN, active_cred, NULL);
1245 	if (error != 0)
1246 		goto out;
1247 	shmfd->shm_mode = mode & ACCESSPERMS;
1248 out:
1249 	mtx_unlock(&shm_timestamp_lock);
1250 	return (error);
1251 }
1252 
1253 static int
1254 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1255     struct thread *td)
1256 {
1257 	struct shmfd *shmfd;
1258 	int error;
1259 
1260 	error = 0;
1261 	shmfd = fp->f_data;
1262 	mtx_lock(&shm_timestamp_lock);
1263 #ifdef MAC
1264 	error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
1265 	if (error != 0)
1266 		goto out;
1267 #endif
1268 	if (uid == (uid_t)-1)
1269 		uid = shmfd->shm_uid;
1270 	if (gid == (gid_t)-1)
1271                  gid = shmfd->shm_gid;
1272 	if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
1273 	    (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
1274 	    (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN)))
1275 		goto out;
1276 	shmfd->shm_uid = uid;
1277 	shmfd->shm_gid = gid;
1278 out:
1279 	mtx_unlock(&shm_timestamp_lock);
1280 	return (error);
1281 }
1282 
1283 /*
1284  * Helper routines to allow the backing object of a shared memory file
1285  * descriptor to be mapped in the kernel.
1286  */
1287 int
1288 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
1289 {
1290 	struct shmfd *shmfd;
1291 	vm_offset_t kva, ofs;
1292 	vm_object_t obj;
1293 	int rv;
1294 
1295 	if (fp->f_type != DTYPE_SHM)
1296 		return (EINVAL);
1297 	shmfd = fp->f_data;
1298 	obj = shmfd->shm_object;
1299 	VM_OBJECT_WLOCK(obj);
1300 	/*
1301 	 * XXXRW: This validation is probably insufficient, and subject to
1302 	 * sign errors.  It should be fixed.
1303 	 */
1304 	if (offset >= shmfd->shm_size ||
1305 	    offset + size > round_page(shmfd->shm_size)) {
1306 		VM_OBJECT_WUNLOCK(obj);
1307 		return (EINVAL);
1308 	}
1309 
1310 	shmfd->shm_kmappings++;
1311 	vm_object_reference_locked(obj);
1312 	VM_OBJECT_WUNLOCK(obj);
1313 
1314 	/* Map the object into the kernel_map and wire it. */
1315 	kva = vm_map_min(kernel_map);
1316 	ofs = offset & PAGE_MASK;
1317 	offset = trunc_page(offset);
1318 	size = round_page(size + ofs);
1319 	rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1320 	    VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1321 	    VM_PROT_READ | VM_PROT_WRITE, 0);
1322 	if (rv == KERN_SUCCESS) {
1323 		rv = vm_map_wire(kernel_map, kva, kva + size,
1324 		    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1325 		if (rv == KERN_SUCCESS) {
1326 			*memp = (void *)(kva + ofs);
1327 			return (0);
1328 		}
1329 		vm_map_remove(kernel_map, kva, kva + size);
1330 	} else
1331 		vm_object_deallocate(obj);
1332 
1333 	/* On failure, drop our mapping reference. */
1334 	VM_OBJECT_WLOCK(obj);
1335 	shmfd->shm_kmappings--;
1336 	VM_OBJECT_WUNLOCK(obj);
1337 
1338 	return (vm_mmap_to_errno(rv));
1339 }
1340 
1341 /*
1342  * We require the caller to unmap the entire entry.  This allows us to
1343  * safely decrement shm_kmappings when a mapping is removed.
1344  */
1345 int
1346 shm_unmap(struct file *fp, void *mem, size_t size)
1347 {
1348 	struct shmfd *shmfd;
1349 	vm_map_entry_t entry;
1350 	vm_offset_t kva, ofs;
1351 	vm_object_t obj;
1352 	vm_pindex_t pindex;
1353 	vm_prot_t prot;
1354 	boolean_t wired;
1355 	vm_map_t map;
1356 	int rv;
1357 
1358 	if (fp->f_type != DTYPE_SHM)
1359 		return (EINVAL);
1360 	shmfd = fp->f_data;
1361 	kva = (vm_offset_t)mem;
1362 	ofs = kva & PAGE_MASK;
1363 	kva = trunc_page(kva);
1364 	size = round_page(size + ofs);
1365 	map = kernel_map;
1366 	rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1367 	    &obj, &pindex, &prot, &wired);
1368 	if (rv != KERN_SUCCESS)
1369 		return (EINVAL);
1370 	if (entry->start != kva || entry->end != kva + size) {
1371 		vm_map_lookup_done(map, entry);
1372 		return (EINVAL);
1373 	}
1374 	vm_map_lookup_done(map, entry);
1375 	if (obj != shmfd->shm_object)
1376 		return (EINVAL);
1377 	vm_map_remove(map, kva, kva + size);
1378 	VM_OBJECT_WLOCK(obj);
1379 	KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1380 	shmfd->shm_kmappings--;
1381 	VM_OBJECT_WUNLOCK(obj);
1382 	return (0);
1383 }
1384 
1385 static int
1386 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list)
1387 {
1388 	const char *path, *pr_path;
1389 	size_t pr_pathlen;
1390 	bool visible;
1391 
1392 	sx_assert(&shm_dict_lock, SA_LOCKED);
1393 	kif->kf_type = KF_TYPE_SHM;
1394 	kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;
1395 	kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1396 	if (shmfd->shm_path != NULL) {
1397 		if (shmfd->shm_path != NULL) {
1398 			path = shmfd->shm_path;
1399 			pr_path = curthread->td_ucred->cr_prison->pr_path;
1400 			if (strcmp(pr_path, "/") != 0) {
1401 				/* Return the jail-rooted pathname. */
1402 				pr_pathlen = strlen(pr_path);
1403 				visible = strncmp(path, pr_path, pr_pathlen)
1404 				    == 0 && path[pr_pathlen] == '/';
1405 				if (list && !visible)
1406 					return (EPERM);
1407 				if (visible)
1408 					path += pr_pathlen;
1409 			}
1410 			strlcpy(kif->kf_path, path, sizeof(kif->kf_path));
1411 		}
1412 	}
1413 	return (0);
1414 }
1415 
1416 static int
1417 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif,
1418     struct filedesc *fdp __unused)
1419 {
1420 	int res;
1421 
1422 	sx_slock(&shm_dict_lock);
1423 	res = shm_fill_kinfo_locked(fp->f_data, kif, false);
1424 	sx_sunlock(&shm_dict_lock);
1425 	return (res);
1426 }
1427 
1428 static int
1429 shm_add_seals(struct file *fp, int seals)
1430 {
1431 	struct shmfd *shmfd;
1432 	void *rl_cookie;
1433 	vm_ooffset_t writemappings;
1434 	int error, nseals;
1435 
1436 	error = 0;
1437 	shmfd = fp->f_data;
1438 	rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
1439 	    &shmfd->shm_mtx);
1440 
1441 	/* Even already-set seals should result in EPERM. */
1442 	if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) {
1443 		error = EPERM;
1444 		goto out;
1445 	}
1446 	nseals = seals & ~shmfd->shm_seals;
1447 	if ((nseals & F_SEAL_WRITE) != 0) {
1448 		/*
1449 		 * The rangelock above prevents writable mappings from being
1450 		 * added after we've started applying seals.  The RLOCK here
1451 		 * is to avoid torn reads on ILP32 arches as unmapping/reducing
1452 		 * writemappings will be done without a rangelock.
1453 		 */
1454 		VM_OBJECT_RLOCK(shmfd->shm_object);
1455 		writemappings = shmfd->shm_object->un_pager.swp.writemappings;
1456 		VM_OBJECT_RUNLOCK(shmfd->shm_object);
1457 		/* kmappings are also writable */
1458 		if (writemappings > 0) {
1459 			error = EBUSY;
1460 			goto out;
1461 		}
1462 	}
1463 	shmfd->shm_seals |= nseals;
1464 out:
1465 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
1466 	return (error);
1467 }
1468 
1469 static int
1470 shm_get_seals(struct file *fp, int *seals)
1471 {
1472 	struct shmfd *shmfd;
1473 
1474 	shmfd = fp->f_data;
1475 	*seals = shmfd->shm_seals;
1476 	return (0);
1477 }
1478 
1479 static int
1480 shm_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
1481 {
1482 	void *rl_cookie;
1483 	struct shmfd *shmfd;
1484 	size_t size;
1485 	int error;
1486 
1487 	/* This assumes that the caller already checked for overflow. */
1488 	error = 0;
1489 	shmfd = fp->f_data;
1490 	size = offset + len;
1491 
1492 	/*
1493 	 * Just grab the rangelock for the range that we may be attempting to
1494 	 * grow, rather than blocking read/write for regions we won't be
1495 	 * touching while this (potential) resize is in progress.  Other
1496 	 * attempts to resize the shmfd will have to take a write lock from 0 to
1497 	 * OFF_MAX, so this being potentially beyond the current usable range of
1498 	 * the shmfd is not necessarily a concern.  If other mechanisms are
1499 	 * added to grow a shmfd, this may need to be re-evaluated.
1500 	 */
1501 	rl_cookie = rangelock_wlock(&shmfd->shm_rl, offset, size,
1502 	    &shmfd->shm_mtx);
1503 	if (size > shmfd->shm_size) {
1504 		VM_OBJECT_WLOCK(shmfd->shm_object);
1505 		error = shm_dotruncate_locked(shmfd, size, rl_cookie);
1506 		VM_OBJECT_WUNLOCK(shmfd->shm_object);
1507 	}
1508 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
1509 	/* Translate to posix_fallocate(2) return value as needed. */
1510 	if (error == ENOMEM)
1511 		error = ENOSPC;
1512 	return (error);
1513 }
1514 
1515 static int
1516 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS)
1517 {
1518 	struct shm_mapping *shmm;
1519 	struct sbuf sb;
1520 	struct kinfo_file kif;
1521 	u_long i;
1522 	ssize_t curlen;
1523 	int error, error2;
1524 
1525 	sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req);
1526 	sbuf_clear_flags(&sb, SBUF_INCLUDENUL);
1527 	curlen = 0;
1528 	error = 0;
1529 	sx_slock(&shm_dict_lock);
1530 	for (i = 0; i < shm_hash + 1; i++) {
1531 		LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) {
1532 			error = shm_fill_kinfo_locked(shmm->sm_shmfd,
1533 			    &kif, true);
1534 			if (error == EPERM)
1535 				continue;
1536 			if (error != 0)
1537 				break;
1538 			pack_kinfo(&kif);
1539 			if (req->oldptr != NULL &&
1540 			    kif.kf_structsize + curlen > req->oldlen)
1541 				break;
1542 			error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ?
1543 			    0 : ENOMEM;
1544 			if (error != 0)
1545 				break;
1546 			curlen += kif.kf_structsize;
1547 		}
1548 	}
1549 	sx_sunlock(&shm_dict_lock);
1550 	error2 = sbuf_finish(&sb);
1551 	sbuf_delete(&sb);
1552 	return (error != 0 ? error : error2);
1553 }
1554 
1555 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list,
1556     CTLFLAG_RD | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE,
1557     NULL, 0, sysctl_posix_shm_list, "",
1558     "POSIX SHM list");
1559 
1560 int
1561 kern_shm_open(struct thread *td, const char *path, int flags, mode_t mode,
1562     struct filecaps *caps)
1563 {
1564 
1565 	return (kern_shm_open2(td, path, flags, mode, 0, caps, NULL));
1566 }
1567 
1568 /*
1569  * This version of the shm_open() interface leaves CLOEXEC behavior up to the
1570  * caller, and libc will enforce it for the traditional shm_open() call.  This
1571  * allows other consumers, like memfd_create(), to opt-in for CLOEXEC.  This
1572  * interface also includes a 'name' argument that is currently unused, but could
1573  * potentially be exported later via some interface for debugging purposes.
1574  * From the kernel's perspective, it is optional.  Individual consumers like
1575  * memfd_create() may require it in order to be compatible with other systems
1576  * implementing the same function.
1577  */
1578 int
1579 sys_shm_open2(struct thread *td, struct shm_open2_args *uap)
1580 {
1581 
1582 	return (kern_shm_open2(td, uap->path, uap->flags, uap->mode,
1583 	    uap->shmflags, NULL, uap->name));
1584 }
1585