xref: /freebsd/sys/kern/uipc_shm.c (revision 3c134670993bf525fcd6c4dfef84a3dfc3d4ed1b)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2006, 2011, 2016-2017 Robert N. M. Watson
5  * Copyright 2020 The FreeBSD Foundation
6  * All rights reserved.
7  *
8  * Portions of this software were developed by BAE Systems, the University of
9  * Cambridge Computer Laboratory, and Memorial University under DARPA/AFRL
10  * contract FA8650-15-C-7558 ("CADETS"), as part of the DARPA Transparent
11  * Computing (TC) research program.
12  *
13  * Portions of this software were developed by Konstantin Belousov
14  * under sponsorship from the FreeBSD Foundation.
15  *
16  * Redistribution and use in source and binary forms, with or without
17  * modification, are permitted provided that the following conditions
18  * are met:
19  * 1. Redistributions of source code must retain the above copyright
20  *    notice, this list of conditions and the following disclaimer.
21  * 2. Redistributions in binary form must reproduce the above copyright
22  *    notice, this list of conditions and the following disclaimer in the
23  *    documentation and/or other materials provided with the distribution.
24  *
25  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
26  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
27  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
28  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
29  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
30  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
31  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
32  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
33  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
34  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35  * SUCH DAMAGE.
36  */
37 
38 /*
39  * Support for shared swap-backed anonymous memory objects via
40  * shm_open(2), shm_rename(2), and shm_unlink(2).
41  * While most of the implementation is here, vm_mmap.c contains
42  * mapping logic changes.
43  *
44  * posixshmcontrol(1) allows users to inspect the state of the memory
45  * objects.  Per-uid swap resource limit controls total amount of
46  * memory that user can consume for anonymous objects, including
47  * shared.
48  */
49 
50 #include <sys/cdefs.h>
51 __FBSDID("$FreeBSD$");
52 
53 #include "opt_capsicum.h"
54 #include "opt_ktrace.h"
55 
56 #include <sys/param.h>
57 #include <sys/capsicum.h>
58 #include <sys/conf.h>
59 #include <sys/fcntl.h>
60 #include <sys/file.h>
61 #include <sys/filedesc.h>
62 #include <sys/filio.h>
63 #include <sys/fnv_hash.h>
64 #include <sys/kernel.h>
65 #include <sys/limits.h>
66 #include <sys/uio.h>
67 #include <sys/signal.h>
68 #include <sys/jail.h>
69 #include <sys/ktrace.h>
70 #include <sys/lock.h>
71 #include <sys/malloc.h>
72 #include <sys/mman.h>
73 #include <sys/mutex.h>
74 #include <sys/priv.h>
75 #include <sys/proc.h>
76 #include <sys/refcount.h>
77 #include <sys/resourcevar.h>
78 #include <sys/rwlock.h>
79 #include <sys/sbuf.h>
80 #include <sys/stat.h>
81 #include <sys/syscallsubr.h>
82 #include <sys/sysctl.h>
83 #include <sys/sysproto.h>
84 #include <sys/systm.h>
85 #include <sys/sx.h>
86 #include <sys/time.h>
87 #include <sys/vmmeter.h>
88 #include <sys/vnode.h>
89 #include <sys/unistd.h>
90 #include <sys/user.h>
91 
92 #include <security/audit/audit.h>
93 #include <security/mac/mac_framework.h>
94 
95 #include <vm/vm.h>
96 #include <vm/vm_param.h>
97 #include <vm/pmap.h>
98 #include <vm/vm_extern.h>
99 #include <vm/vm_map.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_object.h>
102 #include <vm/vm_page.h>
103 #include <vm/vm_pageout.h>
104 #include <vm/vm_pager.h>
105 #include <vm/swap_pager.h>
106 
107 struct shm_mapping {
108 	char		*sm_path;
109 	Fnv32_t		sm_fnv;
110 	struct shmfd	*sm_shmfd;
111 	LIST_ENTRY(shm_mapping) sm_link;
112 };
113 
114 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor");
115 static LIST_HEAD(, shm_mapping) *shm_dictionary;
116 static struct sx shm_dict_lock;
117 static struct mtx shm_timestamp_lock;
118 static u_long shm_hash;
119 static struct unrhdr64 shm_ino_unr;
120 static dev_t shm_dev_ino;
121 
122 #define	SHM_HASH(fnv)	(&shm_dictionary[(fnv) & shm_hash])
123 
124 static void	shm_init(void *arg);
125 static void	shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd);
126 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv);
127 static int	shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred);
128 static int	shm_dotruncate_cookie(struct shmfd *shmfd, off_t length,
129     void *rl_cookie);
130 static int	shm_dotruncate_locked(struct shmfd *shmfd, off_t length,
131     void *rl_cookie);
132 static int	shm_copyin_path(struct thread *td, const char *userpath_in,
133     char **path_out);
134 
135 static fo_rdwr_t	shm_read;
136 static fo_rdwr_t	shm_write;
137 static fo_truncate_t	shm_truncate;
138 static fo_ioctl_t	shm_ioctl;
139 static fo_stat_t	shm_stat;
140 static fo_close_t	shm_close;
141 static fo_chmod_t	shm_chmod;
142 static fo_chown_t	shm_chown;
143 static fo_seek_t	shm_seek;
144 static fo_fill_kinfo_t	shm_fill_kinfo;
145 static fo_mmap_t	shm_mmap;
146 static fo_get_seals_t	shm_get_seals;
147 static fo_add_seals_t	shm_add_seals;
148 static fo_fallocate_t	shm_fallocate;
149 
150 /* File descriptor operations. */
151 struct fileops shm_ops = {
152 	.fo_read = shm_read,
153 	.fo_write = shm_write,
154 	.fo_truncate = shm_truncate,
155 	.fo_ioctl = shm_ioctl,
156 	.fo_poll = invfo_poll,
157 	.fo_kqfilter = invfo_kqfilter,
158 	.fo_stat = shm_stat,
159 	.fo_close = shm_close,
160 	.fo_chmod = shm_chmod,
161 	.fo_chown = shm_chown,
162 	.fo_sendfile = vn_sendfile,
163 	.fo_seek = shm_seek,
164 	.fo_fill_kinfo = shm_fill_kinfo,
165 	.fo_mmap = shm_mmap,
166 	.fo_get_seals = shm_get_seals,
167 	.fo_add_seals = shm_add_seals,
168 	.fo_fallocate = shm_fallocate,
169 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE,
170 };
171 
172 FEATURE(posix_shm, "POSIX shared memory");
173 
174 static SYSCTL_NODE(_vm, OID_AUTO, largepages, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
175     "");
176 
177 static int largepage_reclaim_tries = 1;
178 SYSCTL_INT(_vm_largepages, OID_AUTO, reclaim_tries,
179     CTLFLAG_RWTUN, &largepage_reclaim_tries, 0,
180     "Number of contig reclaims before giving up for default alloc policy");
181 
182 static int
183 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio)
184 {
185 	vm_page_t m;
186 	vm_pindex_t idx;
187 	size_t tlen;
188 	int error, offset, rv;
189 
190 	idx = OFF_TO_IDX(uio->uio_offset);
191 	offset = uio->uio_offset & PAGE_MASK;
192 	tlen = MIN(PAGE_SIZE - offset, len);
193 
194 	rv = vm_page_grab_valid_unlocked(&m, obj, idx,
195 	    VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOCREAT);
196 	if (rv == VM_PAGER_OK)
197 		goto found;
198 
199 	/*
200 	 * Read I/O without either a corresponding resident page or swap
201 	 * page: use zero_region.  This is intended to avoid instantiating
202 	 * pages on read from a sparse region.
203 	 */
204 	VM_OBJECT_WLOCK(obj);
205 	m = vm_page_lookup(obj, idx);
206 	if (uio->uio_rw == UIO_READ && m == NULL &&
207 	    !vm_pager_has_page(obj, idx, NULL, NULL)) {
208 		VM_OBJECT_WUNLOCK(obj);
209 		return (uiomove(__DECONST(void *, zero_region), tlen, uio));
210 	}
211 
212 	/*
213 	 * Although the tmpfs vnode lock is held here, it is
214 	 * nonetheless safe to sleep waiting for a free page.  The
215 	 * pageout daemon does not need to acquire the tmpfs vnode
216 	 * lock to page out tobj's pages because tobj is a OBJT_SWAP
217 	 * type object.
218 	 */
219 	rv = vm_page_grab_valid(&m, obj, idx,
220 	    VM_ALLOC_NORMAL | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY);
221 	if (rv != VM_PAGER_OK) {
222 		VM_OBJECT_WUNLOCK(obj);
223 		printf("uiomove_object: vm_obj %p idx %jd pager error %d\n",
224 		    obj, idx, rv);
225 		return (EIO);
226 	}
227 	VM_OBJECT_WUNLOCK(obj);
228 
229 found:
230 	error = uiomove_fromphys(&m, offset, tlen, uio);
231 	if (uio->uio_rw == UIO_WRITE && error == 0)
232 		vm_page_set_dirty(m);
233 	vm_page_activate(m);
234 	vm_page_sunbusy(m);
235 
236 	return (error);
237 }
238 
239 int
240 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
241 {
242 	ssize_t resid;
243 	size_t len;
244 	int error;
245 
246 	error = 0;
247 	while ((resid = uio->uio_resid) > 0) {
248 		if (obj_size <= uio->uio_offset)
249 			break;
250 		len = MIN(obj_size - uio->uio_offset, resid);
251 		if (len == 0)
252 			break;
253 		error = uiomove_object_page(obj, len, uio);
254 		if (error != 0 || resid == uio->uio_resid)
255 			break;
256 	}
257 	return (error);
258 }
259 
260 static u_long count_largepages[MAXPAGESIZES];
261 
262 static int
263 shm_largepage_phys_populate(vm_object_t object, vm_pindex_t pidx,
264     int fault_type, vm_prot_t max_prot, vm_pindex_t *first, vm_pindex_t *last)
265 {
266 	vm_page_t m;
267 	int psind;
268 
269 	psind = object->un_pager.phys.data_val;
270 	if (psind == 0 || pidx >= object->size)
271 		return (VM_PAGER_FAIL);
272 	*first = rounddown2(pidx, pagesizes[psind] / PAGE_SIZE);
273 
274 	/*
275 	 * We only busy the first page in the superpage run.  It is
276 	 * useless to busy whole run since we only remove full
277 	 * superpage, and it takes too long to busy e.g. 512 * 512 ==
278 	 * 262144 pages constituing 1G amd64 superage.
279 	 */
280 	m = vm_page_grab(object, *first, VM_ALLOC_NORMAL | VM_ALLOC_NOCREAT);
281 	MPASS(m != NULL);
282 
283 	*last = *first + atop(pagesizes[psind]) - 1;
284 	return (VM_PAGER_OK);
285 }
286 
287 static boolean_t
288 shm_largepage_phys_haspage(vm_object_t object, vm_pindex_t pindex,
289     int *before, int *after)
290 {
291 	int psind;
292 
293 	psind = object->un_pager.phys.data_val;
294 	if (psind == 0 || pindex >= object->size)
295 		return (FALSE);
296 	if (before != NULL) {
297 		*before = pindex - rounddown2(pindex, pagesizes[psind] /
298 		    PAGE_SIZE);
299 	}
300 	if (after != NULL) {
301 		*after = roundup2(pindex, pagesizes[psind] / PAGE_SIZE) -
302 		    pindex;
303 	}
304 	return (TRUE);
305 }
306 
307 static void
308 shm_largepage_phys_ctor(vm_object_t object, vm_prot_t prot,
309     vm_ooffset_t foff, struct ucred *cred)
310 {
311 }
312 
313 static void
314 shm_largepage_phys_dtor(vm_object_t object)
315 {
316 	int psind;
317 
318 	psind = object->un_pager.phys.data_val;
319 	if (psind != 0) {
320 		atomic_subtract_long(&count_largepages[psind],
321 		    object->size / (pagesizes[psind] / PAGE_SIZE));
322 		vm_wire_sub(object->size);
323 	} else {
324 		KASSERT(object->size == 0,
325 		    ("largepage phys obj %p not initialized bit size %#jx > 0",
326 		    object, (uintmax_t)object->size));
327 	}
328 }
329 
330 static struct phys_pager_ops shm_largepage_phys_ops = {
331 	.phys_pg_populate =	shm_largepage_phys_populate,
332 	.phys_pg_haspage =	shm_largepage_phys_haspage,
333 	.phys_pg_ctor =		shm_largepage_phys_ctor,
334 	.phys_pg_dtor =		shm_largepage_phys_dtor,
335 };
336 
337 bool
338 shm_largepage(struct shmfd *shmfd)
339 {
340 	return (shmfd->shm_object->type == OBJT_PHYS);
341 }
342 
343 static int
344 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
345 {
346 	struct shmfd *shmfd;
347 	off_t foffset;
348 	int error;
349 
350 	shmfd = fp->f_data;
351 	foffset = foffset_lock(fp, 0);
352 	error = 0;
353 	switch (whence) {
354 	case L_INCR:
355 		if (foffset < 0 ||
356 		    (offset > 0 && foffset > OFF_MAX - offset)) {
357 			error = EOVERFLOW;
358 			break;
359 		}
360 		offset += foffset;
361 		break;
362 	case L_XTND:
363 		if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
364 			error = EOVERFLOW;
365 			break;
366 		}
367 		offset += shmfd->shm_size;
368 		break;
369 	case L_SET:
370 		break;
371 	default:
372 		error = EINVAL;
373 	}
374 	if (error == 0) {
375 		if (offset < 0 || offset > shmfd->shm_size)
376 			error = EINVAL;
377 		else
378 			td->td_uretoff.tdu_off = offset;
379 	}
380 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
381 	return (error);
382 }
383 
384 static int
385 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
386     int flags, struct thread *td)
387 {
388 	struct shmfd *shmfd;
389 	void *rl_cookie;
390 	int error;
391 
392 	shmfd = fp->f_data;
393 #ifdef MAC
394 	error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
395 	if (error)
396 		return (error);
397 #endif
398 	foffset_lock_uio(fp, uio, flags);
399 	rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset,
400 	    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
401 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
402 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
403 	foffset_unlock_uio(fp, uio, flags);
404 	return (error);
405 }
406 
407 static int
408 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
409     int flags, struct thread *td)
410 {
411 	struct shmfd *shmfd;
412 	void *rl_cookie;
413 	int error;
414 	off_t size;
415 
416 	shmfd = fp->f_data;
417 #ifdef MAC
418 	error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
419 	if (error)
420 		return (error);
421 #endif
422 	if (shm_largepage(shmfd) && shmfd->shm_lp_psind == 0)
423 		return (EINVAL);
424 	foffset_lock_uio(fp, uio, flags);
425 	if (uio->uio_resid > OFF_MAX - uio->uio_offset) {
426 		/*
427 		 * Overflow is only an error if we're supposed to expand on
428 		 * write.  Otherwise, we'll just truncate the write to the
429 		 * size of the file, which can only grow up to OFF_MAX.
430 		 */
431 		if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0) {
432 			foffset_unlock_uio(fp, uio, flags);
433 			return (EFBIG);
434 		}
435 
436 		size = shmfd->shm_size;
437 	} else {
438 		size = uio->uio_offset + uio->uio_resid;
439 	}
440 	if ((flags & FOF_OFFSET) == 0) {
441 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
442 		    &shmfd->shm_mtx);
443 	} else {
444 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset,
445 		    size, &shmfd->shm_mtx);
446 	}
447 	if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) {
448 		error = EPERM;
449 	} else {
450 		error = 0;
451 		if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0 &&
452 		    size > shmfd->shm_size) {
453 			VM_OBJECT_WLOCK(shmfd->shm_object);
454 			error = shm_dotruncate_locked(shmfd, size, rl_cookie);
455 			VM_OBJECT_WUNLOCK(shmfd->shm_object);
456 		}
457 		if (error == 0)
458 			error = uiomove_object(shmfd->shm_object,
459 			    shmfd->shm_size, uio);
460 	}
461 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
462 	foffset_unlock_uio(fp, uio, flags);
463 	return (error);
464 }
465 
466 static int
467 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
468     struct thread *td)
469 {
470 	struct shmfd *shmfd;
471 #ifdef MAC
472 	int error;
473 #endif
474 
475 	shmfd = fp->f_data;
476 #ifdef MAC
477 	error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
478 	if (error)
479 		return (error);
480 #endif
481 	return (shm_dotruncate(shmfd, length));
482 }
483 
484 int
485 shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
486     struct thread *td)
487 {
488 	struct shmfd *shmfd;
489 	struct shm_largepage_conf *conf;
490 	void *rl_cookie;
491 
492 	shmfd = fp->f_data;
493 	switch (com) {
494 	case FIONBIO:
495 	case FIOASYNC:
496 		/*
497 		 * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work,
498 		 * just like it would on an unlinked regular file
499 		 */
500 		return (0);
501 	case FIOSSHMLPGCNF:
502 		if (!shm_largepage(shmfd))
503 			return (ENOTTY);
504 		conf = data;
505 		if (shmfd->shm_lp_psind != 0 &&
506 		    conf->psind != shmfd->shm_lp_psind)
507 			return (EINVAL);
508 		if (conf->psind <= 0 || conf->psind >= MAXPAGESIZES ||
509 		    pagesizes[conf->psind] == 0)
510 			return (EINVAL);
511 		if (conf->alloc_policy != SHM_LARGEPAGE_ALLOC_DEFAULT &&
512 		    conf->alloc_policy != SHM_LARGEPAGE_ALLOC_NOWAIT &&
513 		    conf->alloc_policy != SHM_LARGEPAGE_ALLOC_HARD)
514 			return (EINVAL);
515 
516 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
517 		    &shmfd->shm_mtx);
518 		shmfd->shm_lp_psind = conf->psind;
519 		shmfd->shm_lp_alloc_policy = conf->alloc_policy;
520 		shmfd->shm_object->un_pager.phys.data_val = conf->psind;
521 		rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
522 		return (0);
523 	case FIOGSHMLPGCNF:
524 		if (!shm_largepage(shmfd))
525 			return (ENOTTY);
526 		conf = data;
527 		rl_cookie = rangelock_rlock(&shmfd->shm_rl, 0, OFF_MAX,
528 		    &shmfd->shm_mtx);
529 		conf->psind = shmfd->shm_lp_psind;
530 		conf->alloc_policy = shmfd->shm_lp_alloc_policy;
531 		rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
532 		return (0);
533 	default:
534 		return (ENOTTY);
535 	}
536 }
537 
538 static int
539 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
540     struct thread *td)
541 {
542 	struct shmfd *shmfd;
543 #ifdef MAC
544 	int error;
545 #endif
546 
547 	shmfd = fp->f_data;
548 
549 #ifdef MAC
550 	error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
551 	if (error)
552 		return (error);
553 #endif
554 
555 	/*
556 	 * Attempt to return sanish values for fstat() on a memory file
557 	 * descriptor.
558 	 */
559 	bzero(sb, sizeof(*sb));
560 	sb->st_blksize = PAGE_SIZE;
561 	sb->st_size = shmfd->shm_size;
562 	sb->st_blocks = howmany(sb->st_size, sb->st_blksize);
563 	mtx_lock(&shm_timestamp_lock);
564 	sb->st_atim = shmfd->shm_atime;
565 	sb->st_ctim = shmfd->shm_ctime;
566 	sb->st_mtim = shmfd->shm_mtime;
567 	sb->st_birthtim = shmfd->shm_birthtime;
568 	sb->st_mode = S_IFREG | shmfd->shm_mode;		/* XXX */
569 	sb->st_uid = shmfd->shm_uid;
570 	sb->st_gid = shmfd->shm_gid;
571 	mtx_unlock(&shm_timestamp_lock);
572 	sb->st_dev = shm_dev_ino;
573 	sb->st_ino = shmfd->shm_ino;
574 	sb->st_nlink = shmfd->shm_object->ref_count;
575 	sb->st_blocks = shmfd->shm_object->size /
576 	    (pagesizes[shmfd->shm_lp_psind] >> PAGE_SHIFT);
577 
578 	return (0);
579 }
580 
581 static int
582 shm_close(struct file *fp, struct thread *td)
583 {
584 	struct shmfd *shmfd;
585 
586 	shmfd = fp->f_data;
587 	fp->f_data = NULL;
588 	shm_drop(shmfd);
589 
590 	return (0);
591 }
592 
593 static int
594 shm_copyin_path(struct thread *td, const char *userpath_in, char **path_out) {
595 	int error;
596 	char *path;
597 	const char *pr_path;
598 	size_t pr_pathlen;
599 
600 	path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
601 	pr_path = td->td_ucred->cr_prison->pr_path;
602 
603 	/* Construct a full pathname for jailed callers. */
604 	pr_pathlen = strcmp(pr_path, "/") ==
605 	    0 ? 0 : strlcpy(path, pr_path, MAXPATHLEN);
606 	error = copyinstr(userpath_in, path + pr_pathlen,
607 	    MAXPATHLEN - pr_pathlen, NULL);
608 	if (error != 0)
609 		goto out;
610 
611 #ifdef KTRACE
612 	if (KTRPOINT(curthread, KTR_NAMEI))
613 		ktrnamei(path);
614 #endif
615 
616 	/* Require paths to start with a '/' character. */
617 	if (path[pr_pathlen] != '/') {
618 		error = EINVAL;
619 		goto out;
620 	}
621 
622 	*path_out = path;
623 
624 out:
625 	if (error != 0)
626 		free(path, M_SHMFD);
627 
628 	return (error);
629 }
630 
631 static int
632 shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie)
633 {
634 	vm_object_t object;
635 	vm_page_t m;
636 	vm_pindex_t idx, nobjsize;
637 	vm_ooffset_t delta;
638 	int base, rv;
639 
640 	KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
641 	object = shmfd->shm_object;
642 	VM_OBJECT_ASSERT_WLOCKED(object);
643 	rangelock_cookie_assert(rl_cookie, RA_WLOCKED);
644 	if (length == shmfd->shm_size)
645 		return (0);
646 	nobjsize = OFF_TO_IDX(length + PAGE_MASK);
647 
648 	/* Are we shrinking?  If so, trim the end. */
649 	if (length < shmfd->shm_size) {
650 		if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0)
651 			return (EPERM);
652 
653 		/*
654 		 * Disallow any requests to shrink the size if this
655 		 * object is mapped into the kernel.
656 		 */
657 		if (shmfd->shm_kmappings > 0)
658 			return (EBUSY);
659 
660 		/*
661 		 * Zero the truncated part of the last page.
662 		 */
663 		base = length & PAGE_MASK;
664 		if (base != 0) {
665 			idx = OFF_TO_IDX(length);
666 retry:
667 			m = vm_page_grab(object, idx, VM_ALLOC_NOCREAT);
668 			if (m != NULL) {
669 				MPASS(vm_page_all_valid(m));
670 			} else if (vm_pager_has_page(object, idx, NULL, NULL)) {
671 				m = vm_page_alloc(object, idx,
672 				    VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL);
673 				if (m == NULL)
674 					goto retry;
675 				vm_object_pip_add(object, 1);
676 				VM_OBJECT_WUNLOCK(object);
677 				rv = vm_pager_get_pages(object, &m, 1, NULL,
678 				    NULL);
679 				VM_OBJECT_WLOCK(object);
680 				vm_object_pip_wakeup(object);
681 				if (rv == VM_PAGER_OK) {
682 					/*
683 					 * Since the page was not resident,
684 					 * and therefore not recently
685 					 * accessed, immediately enqueue it
686 					 * for asynchronous laundering.  The
687 					 * current operation is not regarded
688 					 * as an access.
689 					 */
690 					vm_page_launder(m);
691 				} else {
692 					vm_page_free(m);
693 					VM_OBJECT_WUNLOCK(object);
694 					return (EIO);
695 				}
696 			}
697 			if (m != NULL) {
698 				pmap_zero_page_area(m, base, PAGE_SIZE - base);
699 				KASSERT(vm_page_all_valid(m),
700 				    ("shm_dotruncate: page %p is invalid", m));
701 				vm_page_set_dirty(m);
702 				vm_page_xunbusy(m);
703 			}
704 		}
705 		delta = IDX_TO_OFF(object->size - nobjsize);
706 
707 		if (nobjsize < object->size)
708 			vm_object_page_remove(object, nobjsize, object->size,
709 			    0);
710 
711 		/* Free the swap accounted for shm */
712 		swap_release_by_cred(delta, object->cred);
713 		object->charge -= delta;
714 	} else {
715 		if ((shmfd->shm_seals & F_SEAL_GROW) != 0)
716 			return (EPERM);
717 
718 		/* Try to reserve additional swap space. */
719 		delta = IDX_TO_OFF(nobjsize - object->size);
720 		if (!swap_reserve_by_cred(delta, object->cred))
721 			return (ENOMEM);
722 		object->charge += delta;
723 	}
724 	shmfd->shm_size = length;
725 	mtx_lock(&shm_timestamp_lock);
726 	vfs_timestamp(&shmfd->shm_ctime);
727 	shmfd->shm_mtime = shmfd->shm_ctime;
728 	mtx_unlock(&shm_timestamp_lock);
729 	object->size = nobjsize;
730 	return (0);
731 }
732 
733 static int
734 shm_dotruncate_largepage(struct shmfd *shmfd, off_t length, void *rl_cookie)
735 {
736 	vm_object_t object;
737 	vm_page_t m;
738 	vm_pindex_t newobjsz, oldobjsz;
739 	int aflags, error, i, psind, try;
740 
741 	KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
742 	object = shmfd->shm_object;
743 	VM_OBJECT_ASSERT_WLOCKED(object);
744 	rangelock_cookie_assert(rl_cookie, RA_WLOCKED);
745 
746 	oldobjsz = object->size;
747 	newobjsz = OFF_TO_IDX(length);
748 	if (length == shmfd->shm_size)
749 		return (0);
750 	psind = shmfd->shm_lp_psind;
751 	if (psind == 0 && length != 0)
752 		return (EINVAL);
753 	if ((length & (pagesizes[psind] - 1)) != 0)
754 		return (EINVAL);
755 
756 	if (length < shmfd->shm_size) {
757 		if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0)
758 			return (EPERM);
759 		if (shmfd->shm_kmappings > 0)
760 			return (EBUSY);
761 		return (ENOTSUP);	/* Pages are unmanaged. */
762 #if 0
763 		vm_object_page_remove(object, newobjsz, oldobjsz, 0);
764 		object->size = newobjsz;
765 		shmfd->shm_size = length;
766 		return (0);
767 #endif
768 	}
769 
770 	aflags = VM_ALLOC_NORMAL | VM_ALLOC_ZERO;
771 	if (shmfd->shm_lp_alloc_policy == SHM_LARGEPAGE_ALLOC_NOWAIT)
772 		aflags |= VM_ALLOC_WAITFAIL;
773 	try = 0;
774 
775 	/*
776 	 * Extend shmfd and object, keeping all already fully
777 	 * allocated large pages intact even on error, because dropped
778 	 * object lock might allowed mapping of them.
779 	 */
780 	while (object->size < newobjsz) {
781 		m = vm_page_alloc_contig(object, object->size, aflags,
782 		    pagesizes[psind] / PAGE_SIZE, 0, ~0,
783 		    pagesizes[psind], 0,
784 		    VM_MEMATTR_DEFAULT);
785 		if (m == NULL) {
786 			VM_OBJECT_WUNLOCK(object);
787 			if (shmfd->shm_lp_alloc_policy ==
788 			    SHM_LARGEPAGE_ALLOC_NOWAIT ||
789 			    (shmfd->shm_lp_alloc_policy ==
790 			    SHM_LARGEPAGE_ALLOC_DEFAULT &&
791 			    try >= largepage_reclaim_tries)) {
792 				VM_OBJECT_WLOCK(object);
793 				return (ENOMEM);
794 			}
795 			error = vm_page_reclaim_contig(aflags,
796 			    pagesizes[psind] / PAGE_SIZE, 0, ~0,
797 			    pagesizes[psind], 0) ? 0 :
798 			    vm_wait_intr(object);
799 			if (error != 0) {
800 				VM_OBJECT_WLOCK(object);
801 				return (error);
802 			}
803 			try++;
804 			VM_OBJECT_WLOCK(object);
805 			continue;
806 		}
807 		try = 0;
808 		for (i = 0; i < pagesizes[psind] / PAGE_SIZE; i++) {
809 			if ((m[i].flags & PG_ZERO) == 0)
810 				pmap_zero_page(&m[i]);
811 			vm_page_valid(&m[i]);
812 			vm_page_xunbusy(&m[i]);
813 		}
814 		object->size += OFF_TO_IDX(pagesizes[psind]);
815 		shmfd->shm_size += pagesizes[psind];
816 		atomic_add_long(&count_largepages[psind], 1);
817 		vm_wire_add(atop(pagesizes[psind]));
818 	}
819 	return (0);
820 }
821 
822 static int
823 shm_dotruncate_cookie(struct shmfd *shmfd, off_t length, void *rl_cookie)
824 {
825 	int error;
826 
827 	VM_OBJECT_WLOCK(shmfd->shm_object);
828 	error = shm_largepage(shmfd) ? shm_dotruncate_largepage(shmfd,
829 	    length, rl_cookie) : shm_dotruncate_locked(shmfd, length,
830 	    rl_cookie);
831 	VM_OBJECT_WUNLOCK(shmfd->shm_object);
832 	return (error);
833 }
834 
835 int
836 shm_dotruncate(struct shmfd *shmfd, off_t length)
837 {
838 	void *rl_cookie;
839 	int error;
840 
841 	rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
842 	    &shmfd->shm_mtx);
843 	error = shm_dotruncate_cookie(shmfd, length, rl_cookie);
844 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
845 	return (error);
846 }
847 
848 /*
849  * shmfd object management including creation and reference counting
850  * routines.
851  */
852 struct shmfd *
853 shm_alloc(struct ucred *ucred, mode_t mode, bool largepage)
854 {
855 	struct shmfd *shmfd;
856 
857 	shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
858 	shmfd->shm_size = 0;
859 	shmfd->shm_uid = ucred->cr_uid;
860 	shmfd->shm_gid = ucred->cr_gid;
861 	shmfd->shm_mode = mode;
862 	if (largepage) {
863 		shmfd->shm_object = phys_pager_allocate(NULL,
864 		    &shm_largepage_phys_ops, NULL, shmfd->shm_size,
865 		    VM_PROT_DEFAULT, 0, ucred);
866 		shmfd->shm_lp_alloc_policy = SHM_LARGEPAGE_ALLOC_DEFAULT;
867 	} else {
868 		shmfd->shm_object = vm_pager_allocate(OBJT_SWAP, NULL,
869 		    shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
870 	}
871 	KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
872 	vfs_timestamp(&shmfd->shm_birthtime);
873 	shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
874 	    shmfd->shm_birthtime;
875 	shmfd->shm_ino = alloc_unr64(&shm_ino_unr);
876 	refcount_init(&shmfd->shm_refs, 1);
877 	mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
878 	rangelock_init(&shmfd->shm_rl);
879 #ifdef MAC
880 	mac_posixshm_init(shmfd);
881 	mac_posixshm_create(ucred, shmfd);
882 #endif
883 
884 	return (shmfd);
885 }
886 
887 struct shmfd *
888 shm_hold(struct shmfd *shmfd)
889 {
890 
891 	refcount_acquire(&shmfd->shm_refs);
892 	return (shmfd);
893 }
894 
895 void
896 shm_drop(struct shmfd *shmfd)
897 {
898 
899 	if (refcount_release(&shmfd->shm_refs)) {
900 #ifdef MAC
901 		mac_posixshm_destroy(shmfd);
902 #endif
903 		rangelock_destroy(&shmfd->shm_rl);
904 		mtx_destroy(&shmfd->shm_mtx);
905 		vm_object_deallocate(shmfd->shm_object);
906 		free(shmfd, M_SHMFD);
907 	}
908 }
909 
910 /*
911  * Determine if the credentials have sufficient permissions for a
912  * specified combination of FREAD and FWRITE.
913  */
914 int
915 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
916 {
917 	accmode_t accmode;
918 	int error;
919 
920 	accmode = 0;
921 	if (flags & FREAD)
922 		accmode |= VREAD;
923 	if (flags & FWRITE)
924 		accmode |= VWRITE;
925 	mtx_lock(&shm_timestamp_lock);
926 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
927 	    accmode, ucred);
928 	mtx_unlock(&shm_timestamp_lock);
929 	return (error);
930 }
931 
932 static void
933 shm_init(void *arg)
934 {
935 	char name[32];
936 	int i;
937 
938 	mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
939 	sx_init(&shm_dict_lock, "shm dictionary");
940 	shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
941 	new_unrhdr64(&shm_ino_unr, 1);
942 	shm_dev_ino = devfs_alloc_cdp_inode();
943 	KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
944 
945 	for (i = 1; i < MAXPAGESIZES; i++) {
946 		if (pagesizes[i] == 0)
947 			break;
948 #define	M	(1024 * 1024)
949 #define	G	(1024 * M)
950 		if (pagesizes[i] >= G)
951 			snprintf(name, sizeof(name), "%luG", pagesizes[i] / G);
952 		else if (pagesizes[i] >= M)
953 			snprintf(name, sizeof(name), "%luM", pagesizes[i] / M);
954 		else
955 			snprintf(name, sizeof(name), "%lu", pagesizes[i]);
956 #undef G
957 #undef M
958 		SYSCTL_ADD_ULONG(NULL, SYSCTL_STATIC_CHILDREN(_vm_largepages),
959 		    OID_AUTO, name, CTLFLAG_RD, &count_largepages[i],
960 		    "number of non-transient largepages allocated");
961 	}
962 }
963 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
964 
965 /*
966  * Dictionary management.  We maintain an in-kernel dictionary to map
967  * paths to shmfd objects.  We use the FNV hash on the path to store
968  * the mappings in a hash table.
969  */
970 static struct shmfd *
971 shm_lookup(char *path, Fnv32_t fnv)
972 {
973 	struct shm_mapping *map;
974 
975 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
976 		if (map->sm_fnv != fnv)
977 			continue;
978 		if (strcmp(map->sm_path, path) == 0)
979 			return (map->sm_shmfd);
980 	}
981 
982 	return (NULL);
983 }
984 
985 static void
986 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
987 {
988 	struct shm_mapping *map;
989 
990 	map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
991 	map->sm_path = path;
992 	map->sm_fnv = fnv;
993 	map->sm_shmfd = shm_hold(shmfd);
994 	shmfd->shm_path = path;
995 	LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
996 }
997 
998 static int
999 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
1000 {
1001 	struct shm_mapping *map;
1002 	int error;
1003 
1004 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
1005 		if (map->sm_fnv != fnv)
1006 			continue;
1007 		if (strcmp(map->sm_path, path) == 0) {
1008 #ifdef MAC
1009 			error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
1010 			if (error)
1011 				return (error);
1012 #endif
1013 			error = shm_access(map->sm_shmfd, ucred,
1014 			    FREAD | FWRITE);
1015 			if (error)
1016 				return (error);
1017 			map->sm_shmfd->shm_path = NULL;
1018 			LIST_REMOVE(map, sm_link);
1019 			shm_drop(map->sm_shmfd);
1020 			free(map->sm_path, M_SHMFD);
1021 			free(map, M_SHMFD);
1022 			return (0);
1023 		}
1024 	}
1025 
1026 	return (ENOENT);
1027 }
1028 
1029 int
1030 kern_shm_open2(struct thread *td, const char *userpath, int flags, mode_t mode,
1031     int shmflags, struct filecaps *fcaps, const char *name __unused)
1032 {
1033 	struct filedesc *fdp;
1034 	struct shmfd *shmfd;
1035 	struct file *fp;
1036 	char *path;
1037 	void *rl_cookie;
1038 	Fnv32_t fnv;
1039 	mode_t cmode;
1040 	int error, fd, initial_seals;
1041 	bool largepage;
1042 
1043 	if ((shmflags & ~(SHM_ALLOW_SEALING | SHM_GROW_ON_WRITE |
1044 	    SHM_LARGEPAGE)) != 0)
1045 		return (EINVAL);
1046 
1047 	initial_seals = F_SEAL_SEAL;
1048 	if ((shmflags & SHM_ALLOW_SEALING) != 0)
1049 		initial_seals &= ~F_SEAL_SEAL;
1050 
1051 #ifdef CAPABILITY_MODE
1052 	/*
1053 	 * shm_open(2) is only allowed for anonymous objects.
1054 	 */
1055 	if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON))
1056 		return (ECAPMODE);
1057 #endif
1058 
1059 	AUDIT_ARG_FFLAGS(flags);
1060 	AUDIT_ARG_MODE(mode);
1061 
1062 	if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
1063 		return (EINVAL);
1064 
1065 	if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
1066 		return (EINVAL);
1067 
1068 	largepage = (shmflags & SHM_LARGEPAGE) != 0;
1069 #if !defined(__amd64__)
1070 	if (largepage)
1071 		return (ENOTTY);
1072 #endif
1073 
1074 	/*
1075 	 * Currently only F_SEAL_SEAL may be set when creating or opening shmfd.
1076 	 * If the decision is made later to allow additional seals, care must be
1077 	 * taken below to ensure that the seals are properly set if the shmfd
1078 	 * already existed -- this currently assumes that only F_SEAL_SEAL can
1079 	 * be set and doesn't take further precautions to ensure the validity of
1080 	 * the seals being added with respect to current mappings.
1081 	 */
1082 	if ((initial_seals & ~F_SEAL_SEAL) != 0)
1083 		return (EINVAL);
1084 
1085 	fdp = td->td_proc->p_fd;
1086 	cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS;
1087 
1088 	/*
1089 	 * shm_open(2) created shm should always have O_CLOEXEC set, as mandated
1090 	 * by POSIX.  We allow it to be unset here so that an in-kernel
1091 	 * interface may be written as a thin layer around shm, optionally not
1092 	 * setting CLOEXEC.  For shm_open(2), O_CLOEXEC is set unconditionally
1093 	 * in sys_shm_open() to keep this implementation compliant.
1094 	 */
1095 	error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps);
1096 	if (error)
1097 		return (error);
1098 
1099 	/* A SHM_ANON path pointer creates an anonymous object. */
1100 	if (userpath == SHM_ANON) {
1101 		/* A read-only anonymous object is pointless. */
1102 		if ((flags & O_ACCMODE) == O_RDONLY) {
1103 			fdclose(td, fp, fd);
1104 			fdrop(fp, td);
1105 			return (EINVAL);
1106 		}
1107 		shmfd = shm_alloc(td->td_ucred, cmode, largepage);
1108 		shmfd->shm_seals = initial_seals;
1109 		shmfd->shm_flags = shmflags;
1110 	} else {
1111 		error = shm_copyin_path(td, userpath, &path);
1112 		if (error != 0) {
1113 			fdclose(td, fp, fd);
1114 			fdrop(fp, td);
1115 			return (error);
1116 		}
1117 
1118 		AUDIT_ARG_UPATH1_CANON(path);
1119 		fnv = fnv_32_str(path, FNV1_32_INIT);
1120 		sx_xlock(&shm_dict_lock);
1121 		shmfd = shm_lookup(path, fnv);
1122 		if (shmfd == NULL) {
1123 			/* Object does not yet exist, create it if requested. */
1124 			if (flags & O_CREAT) {
1125 #ifdef MAC
1126 				error = mac_posixshm_check_create(td->td_ucred,
1127 				    path);
1128 				if (error == 0) {
1129 #endif
1130 					shmfd = shm_alloc(td->td_ucred, cmode,
1131 					    largepage);
1132 					shmfd->shm_seals = initial_seals;
1133 					shmfd->shm_flags = shmflags;
1134 					shm_insert(path, fnv, shmfd);
1135 #ifdef MAC
1136 				}
1137 #endif
1138 			} else {
1139 				free(path, M_SHMFD);
1140 				error = ENOENT;
1141 			}
1142 		} else {
1143 			rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
1144 			    &shmfd->shm_mtx);
1145 
1146 			/*
1147 			 * kern_shm_open() likely shouldn't ever error out on
1148 			 * trying to set a seal that already exists, unlike
1149 			 * F_ADD_SEALS.  This would break terribly as
1150 			 * shm_open(2) actually sets F_SEAL_SEAL to maintain
1151 			 * historical behavior where the underlying file could
1152 			 * not be sealed.
1153 			 */
1154 			initial_seals &= ~shmfd->shm_seals;
1155 
1156 			/*
1157 			 * Object already exists, obtain a new
1158 			 * reference if requested and permitted.
1159 			 */
1160 			free(path, M_SHMFD);
1161 
1162 			/*
1163 			 * initial_seals can't set additional seals if we've
1164 			 * already been set F_SEAL_SEAL.  If F_SEAL_SEAL is set,
1165 			 * then we've already removed that one from
1166 			 * initial_seals.  This is currently redundant as we
1167 			 * only allow setting F_SEAL_SEAL at creation time, but
1168 			 * it's cheap to check and decreases the effort required
1169 			 * to allow additional seals.
1170 			 */
1171 			if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 &&
1172 			    initial_seals != 0)
1173 				error = EPERM;
1174 			else if ((flags & (O_CREAT | O_EXCL)) ==
1175 			    (O_CREAT | O_EXCL))
1176 				error = EEXIST;
1177 			else if (shmflags != 0 && shmflags != shmfd->shm_flags)
1178 				error = EINVAL;
1179 			else {
1180 #ifdef MAC
1181 				error = mac_posixshm_check_open(td->td_ucred,
1182 				    shmfd, FFLAGS(flags & O_ACCMODE));
1183 				if (error == 0)
1184 #endif
1185 				error = shm_access(shmfd, td->td_ucred,
1186 				    FFLAGS(flags & O_ACCMODE));
1187 			}
1188 
1189 			/*
1190 			 * Truncate the file back to zero length if
1191 			 * O_TRUNC was specified and the object was
1192 			 * opened with read/write.
1193 			 */
1194 			if (error == 0 &&
1195 			    (flags & (O_ACCMODE | O_TRUNC)) ==
1196 			    (O_RDWR | O_TRUNC)) {
1197 				VM_OBJECT_WLOCK(shmfd->shm_object);
1198 #ifdef MAC
1199 				error = mac_posixshm_check_truncate(
1200 					td->td_ucred, fp->f_cred, shmfd);
1201 				if (error == 0)
1202 #endif
1203 					error = shm_dotruncate_locked(shmfd, 0,
1204 					    rl_cookie);
1205 				VM_OBJECT_WUNLOCK(shmfd->shm_object);
1206 			}
1207 			if (error == 0) {
1208 				/*
1209 				 * Currently we only allow F_SEAL_SEAL to be
1210 				 * set initially.  As noted above, this would
1211 				 * need to be reworked should that change.
1212 				 */
1213 				shmfd->shm_seals |= initial_seals;
1214 				shm_hold(shmfd);
1215 			}
1216 			rangelock_unlock(&shmfd->shm_rl, rl_cookie,
1217 			    &shmfd->shm_mtx);
1218 		}
1219 		sx_xunlock(&shm_dict_lock);
1220 
1221 		if (error) {
1222 			fdclose(td, fp, fd);
1223 			fdrop(fp, td);
1224 			return (error);
1225 		}
1226 	}
1227 
1228 	finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
1229 
1230 	td->td_retval[0] = fd;
1231 	fdrop(fp, td);
1232 
1233 	return (0);
1234 }
1235 
1236 /* System calls. */
1237 #ifdef COMPAT_FREEBSD12
1238 int
1239 freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap)
1240 {
1241 
1242 	return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC,
1243 	    uap->mode, NULL));
1244 }
1245 #endif
1246 
1247 int
1248 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
1249 {
1250 	char *path;
1251 	Fnv32_t fnv;
1252 	int error;
1253 
1254 	error = shm_copyin_path(td, uap->path, &path);
1255 	if (error != 0)
1256 		return (error);
1257 
1258 	AUDIT_ARG_UPATH1_CANON(path);
1259 	fnv = fnv_32_str(path, FNV1_32_INIT);
1260 	sx_xlock(&shm_dict_lock);
1261 	error = shm_remove(path, fnv, td->td_ucred);
1262 	sx_xunlock(&shm_dict_lock);
1263 	free(path, M_SHMFD);
1264 
1265 	return (error);
1266 }
1267 
1268 int
1269 sys_shm_rename(struct thread *td, struct shm_rename_args *uap)
1270 {
1271 	char *path_from = NULL, *path_to = NULL;
1272 	Fnv32_t fnv_from, fnv_to;
1273 	struct shmfd *fd_from;
1274 	struct shmfd *fd_to;
1275 	int error;
1276 	int flags;
1277 
1278 	flags = uap->flags;
1279 	AUDIT_ARG_FFLAGS(flags);
1280 
1281 	/*
1282 	 * Make sure the user passed only valid flags.
1283 	 * If you add a new flag, please add a new term here.
1284 	 */
1285 	if ((flags & ~(
1286 	    SHM_RENAME_NOREPLACE |
1287 	    SHM_RENAME_EXCHANGE
1288 	    )) != 0) {
1289 		error = EINVAL;
1290 		goto out;
1291 	}
1292 
1293 	/*
1294 	 * EXCHANGE and NOREPLACE don't quite make sense together. Let's
1295 	 * force the user to choose one or the other.
1296 	 */
1297 	if ((flags & SHM_RENAME_NOREPLACE) != 0 &&
1298 	    (flags & SHM_RENAME_EXCHANGE) != 0) {
1299 		error = EINVAL;
1300 		goto out;
1301 	}
1302 
1303 	/* Renaming to or from anonymous makes no sense */
1304 	if (uap->path_from == SHM_ANON || uap->path_to == SHM_ANON) {
1305 		error = EINVAL;
1306 		goto out;
1307 	}
1308 
1309 	error = shm_copyin_path(td, uap->path_from, &path_from);
1310 	if (error != 0)
1311 		goto out;
1312 
1313 	error = shm_copyin_path(td, uap->path_to, &path_to);
1314 	if (error != 0)
1315 		goto out;
1316 
1317 	AUDIT_ARG_UPATH1_CANON(path_from);
1318 	AUDIT_ARG_UPATH2_CANON(path_to);
1319 
1320 	/* Rename with from/to equal is a no-op */
1321 	if (strcmp(path_from, path_to) == 0)
1322 		goto out;
1323 
1324 	fnv_from = fnv_32_str(path_from, FNV1_32_INIT);
1325 	fnv_to = fnv_32_str(path_to, FNV1_32_INIT);
1326 
1327 	sx_xlock(&shm_dict_lock);
1328 
1329 	fd_from = shm_lookup(path_from, fnv_from);
1330 	if (fd_from == NULL) {
1331 		error = ENOENT;
1332 		goto out_locked;
1333 	}
1334 
1335 	fd_to = shm_lookup(path_to, fnv_to);
1336 	if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) {
1337 		error = EEXIST;
1338 		goto out_locked;
1339 	}
1340 
1341 	/*
1342 	 * Unconditionally prevents shm_remove from invalidating the 'from'
1343 	 * shm's state.
1344 	 */
1345 	shm_hold(fd_from);
1346 	error = shm_remove(path_from, fnv_from, td->td_ucred);
1347 
1348 	/*
1349 	 * One of my assumptions failed if ENOENT (e.g. locking didn't
1350 	 * protect us)
1351 	 */
1352 	KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s",
1353 	    path_from));
1354 	if (error != 0) {
1355 		shm_drop(fd_from);
1356 		goto out_locked;
1357 	}
1358 
1359 	/*
1360 	 * If we are exchanging, we need to ensure the shm_remove below
1361 	 * doesn't invalidate the dest shm's state.
1362 	 */
1363 	if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL)
1364 		shm_hold(fd_to);
1365 
1366 	/*
1367 	 * NOTE: if path_to is not already in the hash, c'est la vie;
1368 	 * it simply means we have nothing already at path_to to unlink.
1369 	 * That is the ENOENT case.
1370 	 *
1371 	 * If we somehow don't have access to unlink this guy, but
1372 	 * did for the shm at path_from, then relink the shm to path_from
1373 	 * and abort with EACCES.
1374 	 *
1375 	 * All other errors: that is weird; let's relink and abort the
1376 	 * operation.
1377 	 */
1378 	error = shm_remove(path_to, fnv_to, td->td_ucred);
1379 	if (error != 0 && error != ENOENT) {
1380 		shm_insert(path_from, fnv_from, fd_from);
1381 		shm_drop(fd_from);
1382 		/* Don't free path_from now, since the hash references it */
1383 		path_from = NULL;
1384 		goto out_locked;
1385 	}
1386 
1387 	error = 0;
1388 
1389 	shm_insert(path_to, fnv_to, fd_from);
1390 
1391 	/* Don't free path_to now, since the hash references it */
1392 	path_to = NULL;
1393 
1394 	/* We kept a ref when we removed, and incremented again in insert */
1395 	shm_drop(fd_from);
1396 	KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1397 	    fd_from->shm_refs));
1398 
1399 	if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) {
1400 		shm_insert(path_from, fnv_from, fd_to);
1401 		path_from = NULL;
1402 		shm_drop(fd_to);
1403 		KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1404 		    fd_to->shm_refs));
1405 	}
1406 
1407 out_locked:
1408 	sx_xunlock(&shm_dict_lock);
1409 
1410 out:
1411 	free(path_from, M_SHMFD);
1412 	free(path_to, M_SHMFD);
1413 	return (error);
1414 }
1415 
1416 static int
1417 shm_mmap_large(struct shmfd *shmfd, vm_map_t map, vm_offset_t *addr,
1418     vm_size_t size, vm_prot_t prot, vm_prot_t max_prot, int flags,
1419     vm_ooffset_t foff, bool writecounted, struct thread *td)
1420 {
1421 	struct vmspace *vms;
1422 	vm_map_entry_t next_entry, prev_entry;
1423 	vm_offset_t align, mask, maxaddr;
1424 	int docow, error, rv, try;
1425 	bool curmap;
1426 
1427 	if (shmfd->shm_lp_psind == 0)
1428 		return (EINVAL);
1429 
1430 	/* MAP_PRIVATE is disabled */
1431 	if ((flags & ~(MAP_SHARED | MAP_FIXED | MAP_EXCL |
1432 	    MAP_NOCORE |
1433 #ifdef MAP_32BIT
1434 	    MAP_32BIT |
1435 #endif
1436 	    MAP_ALIGNMENT_MASK)) != 0)
1437 		return (EINVAL);
1438 
1439 	vms = td->td_proc->p_vmspace;
1440 	curmap = map == &vms->vm_map;
1441 	if (curmap) {
1442 		error = kern_mmap_racct_check(td, map, size);
1443 		if (error != 0)
1444 			return (error);
1445 	}
1446 
1447 	docow = shmfd->shm_lp_psind << MAP_SPLIT_BOUNDARY_SHIFT;
1448 	docow |= MAP_INHERIT_SHARE;
1449 	if ((flags & MAP_NOCORE) != 0)
1450 		docow |= MAP_DISABLE_COREDUMP;
1451 	if (writecounted)
1452 		docow |= MAP_WRITECOUNT;
1453 
1454 	mask = pagesizes[shmfd->shm_lp_psind] - 1;
1455 	if ((foff & mask) != 0)
1456 		return (EINVAL);
1457 	maxaddr = vm_map_max(map);
1458 #ifdef MAP_32BIT
1459 	if ((flags & MAP_32BIT) != 0 && maxaddr > MAP_32BIT_MAX_ADDR)
1460 		maxaddr = MAP_32BIT_MAX_ADDR;
1461 #endif
1462 	if (size == 0 || (size & mask) != 0 ||
1463 	    (*addr != 0 && ((*addr & mask) != 0 ||
1464 	    *addr + size < *addr || *addr + size > maxaddr)))
1465 		return (EINVAL);
1466 
1467 	align = flags & MAP_ALIGNMENT_MASK;
1468 	if (align == 0) {
1469 		align = pagesizes[shmfd->shm_lp_psind];
1470 	} else if (align == MAP_ALIGNED_SUPER) {
1471 		if (shmfd->shm_lp_psind != 1)
1472 			return (EINVAL);
1473 		align = pagesizes[1];
1474 	} else {
1475 		align >>= MAP_ALIGNMENT_SHIFT;
1476 		align = 1ULL << align;
1477 		/* Also handles overflow. */
1478 		if (align < pagesizes[shmfd->shm_lp_psind])
1479 			return (EINVAL);
1480 	}
1481 
1482 	vm_map_lock(map);
1483 	if ((flags & MAP_FIXED) == 0) {
1484 		try = 1;
1485 		if (curmap && (*addr == 0 ||
1486 		    (*addr >= round_page((vm_offset_t)vms->vm_taddr) &&
1487 		    *addr < round_page((vm_offset_t)vms->vm_daddr +
1488 		    lim_max(td, RLIMIT_DATA))))) {
1489 			*addr = roundup2((vm_offset_t)vms->vm_daddr +
1490 			    lim_max(td, RLIMIT_DATA),
1491 			    pagesizes[shmfd->shm_lp_psind]);
1492 		}
1493 again:
1494 		rv = vm_map_find_aligned(map, addr, size, maxaddr, align);
1495 		if (rv != KERN_SUCCESS) {
1496 			if (try == 1) {
1497 				try = 2;
1498 				*addr = vm_map_min(map);
1499 				if ((*addr & mask) != 0)
1500 					*addr = (*addr + mask) & mask;
1501 				goto again;
1502 			}
1503 			goto fail1;
1504 		}
1505 	} else if ((flags & MAP_EXCL) == 0) {
1506 		rv = vm_map_delete(map, *addr, *addr + size);
1507 		if (rv != KERN_SUCCESS)
1508 			goto fail1;
1509 	} else {
1510 		error = ENOSPC;
1511 		if (vm_map_lookup_entry(map, *addr, &prev_entry))
1512 			goto fail;
1513 		next_entry = vm_map_entry_succ(prev_entry);
1514 		if (next_entry->start < *addr + size)
1515 			goto fail;
1516 	}
1517 
1518 	rv = vm_map_insert(map, shmfd->shm_object, foff, *addr, *addr + size,
1519 	    prot, max_prot, docow);
1520 fail1:
1521 	error = vm_mmap_to_errno(rv);
1522 fail:
1523 	vm_map_unlock(map);
1524 	return (error);
1525 }
1526 
1527 static int
1528 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
1529     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
1530     vm_ooffset_t foff, struct thread *td)
1531 {
1532 	struct shmfd *shmfd;
1533 	vm_prot_t maxprot;
1534 	int error;
1535 	bool writecnt;
1536 	void *rl_cookie;
1537 
1538 	shmfd = fp->f_data;
1539 	maxprot = VM_PROT_NONE;
1540 
1541 	rl_cookie = rangelock_rlock(&shmfd->shm_rl, 0, objsize,
1542 	    &shmfd->shm_mtx);
1543 	/* FREAD should always be set. */
1544 	if ((fp->f_flag & FREAD) != 0)
1545 		maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
1546 
1547 	/*
1548 	 * If FWRITE's set, we can allow VM_PROT_WRITE unless it's a shared
1549 	 * mapping with a write seal applied.  Private mappings are always
1550 	 * writeable.
1551 	 */
1552 	if ((flags & MAP_SHARED) == 0) {
1553 		cap_maxprot |= VM_PROT_WRITE;
1554 		maxprot |= VM_PROT_WRITE;
1555 		writecnt = false;
1556 	} else {
1557 		if ((fp->f_flag & FWRITE) != 0 &&
1558 		    (shmfd->shm_seals & F_SEAL_WRITE) == 0)
1559 			maxprot |= VM_PROT_WRITE;
1560 
1561 		/*
1562 		 * Any mappings from a writable descriptor may be upgraded to
1563 		 * VM_PROT_WRITE with mprotect(2), unless a write-seal was
1564 		 * applied between the open and subsequent mmap(2).  We want to
1565 		 * reject application of a write seal as long as any such
1566 		 * mapping exists so that the seal cannot be trivially bypassed.
1567 		 */
1568 		writecnt = (maxprot & VM_PROT_WRITE) != 0;
1569 		if (!writecnt && (prot & VM_PROT_WRITE) != 0) {
1570 			error = EACCES;
1571 			goto out;
1572 		}
1573 	}
1574 	maxprot &= cap_maxprot;
1575 
1576 	/* See comment in vn_mmap(). */
1577 	if (
1578 #ifdef _LP64
1579 	    objsize > OFF_MAX ||
1580 #endif
1581 	    foff < 0 || foff > OFF_MAX - objsize) {
1582 		error = EINVAL;
1583 		goto out;
1584 	}
1585 
1586 #ifdef MAC
1587 	error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
1588 	if (error != 0)
1589 		goto out;
1590 #endif
1591 
1592 	mtx_lock(&shm_timestamp_lock);
1593 	vfs_timestamp(&shmfd->shm_atime);
1594 	mtx_unlock(&shm_timestamp_lock);
1595 	vm_object_reference(shmfd->shm_object);
1596 
1597 	if (writecnt)
1598 		vm_pager_update_writecount(shmfd->shm_object, 0, objsize);
1599 	if (shm_largepage(shmfd)) {
1600 		error = shm_mmap_large(shmfd, map, addr, objsize, prot,
1601 		    maxprot, flags, foff, writecnt, td);
1602 	} else {
1603 		error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
1604 		    shmfd->shm_object, foff, writecnt, td);
1605 	}
1606 	if (error != 0) {
1607 		if (writecnt)
1608 			vm_pager_release_writecount(shmfd->shm_object, 0,
1609 			    objsize);
1610 		vm_object_deallocate(shmfd->shm_object);
1611 	}
1612 out:
1613 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
1614 	return (error);
1615 }
1616 
1617 static int
1618 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
1619     struct thread *td)
1620 {
1621 	struct shmfd *shmfd;
1622 	int error;
1623 
1624 	error = 0;
1625 	shmfd = fp->f_data;
1626 	mtx_lock(&shm_timestamp_lock);
1627 	/*
1628 	 * SUSv4 says that x bits of permission need not be affected.
1629 	 * Be consistent with our shm_open there.
1630 	 */
1631 #ifdef MAC
1632 	error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
1633 	if (error != 0)
1634 		goto out;
1635 #endif
1636 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
1637 	    VADMIN, active_cred);
1638 	if (error != 0)
1639 		goto out;
1640 	shmfd->shm_mode = mode & ACCESSPERMS;
1641 out:
1642 	mtx_unlock(&shm_timestamp_lock);
1643 	return (error);
1644 }
1645 
1646 static int
1647 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1648     struct thread *td)
1649 {
1650 	struct shmfd *shmfd;
1651 	int error;
1652 
1653 	error = 0;
1654 	shmfd = fp->f_data;
1655 	mtx_lock(&shm_timestamp_lock);
1656 #ifdef MAC
1657 	error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
1658 	if (error != 0)
1659 		goto out;
1660 #endif
1661 	if (uid == (uid_t)-1)
1662 		uid = shmfd->shm_uid;
1663 	if (gid == (gid_t)-1)
1664                  gid = shmfd->shm_gid;
1665 	if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
1666 	    (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
1667 	    (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN)))
1668 		goto out;
1669 	shmfd->shm_uid = uid;
1670 	shmfd->shm_gid = gid;
1671 out:
1672 	mtx_unlock(&shm_timestamp_lock);
1673 	return (error);
1674 }
1675 
1676 /*
1677  * Helper routines to allow the backing object of a shared memory file
1678  * descriptor to be mapped in the kernel.
1679  */
1680 int
1681 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
1682 {
1683 	struct shmfd *shmfd;
1684 	vm_offset_t kva, ofs;
1685 	vm_object_t obj;
1686 	int rv;
1687 
1688 	if (fp->f_type != DTYPE_SHM)
1689 		return (EINVAL);
1690 	shmfd = fp->f_data;
1691 	obj = shmfd->shm_object;
1692 	VM_OBJECT_WLOCK(obj);
1693 	/*
1694 	 * XXXRW: This validation is probably insufficient, and subject to
1695 	 * sign errors.  It should be fixed.
1696 	 */
1697 	if (offset >= shmfd->shm_size ||
1698 	    offset + size > round_page(shmfd->shm_size)) {
1699 		VM_OBJECT_WUNLOCK(obj);
1700 		return (EINVAL);
1701 	}
1702 
1703 	shmfd->shm_kmappings++;
1704 	vm_object_reference_locked(obj);
1705 	VM_OBJECT_WUNLOCK(obj);
1706 
1707 	/* Map the object into the kernel_map and wire it. */
1708 	kva = vm_map_min(kernel_map);
1709 	ofs = offset & PAGE_MASK;
1710 	offset = trunc_page(offset);
1711 	size = round_page(size + ofs);
1712 	rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1713 	    VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1714 	    VM_PROT_READ | VM_PROT_WRITE, 0);
1715 	if (rv == KERN_SUCCESS) {
1716 		rv = vm_map_wire(kernel_map, kva, kva + size,
1717 		    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1718 		if (rv == KERN_SUCCESS) {
1719 			*memp = (void *)(kva + ofs);
1720 			return (0);
1721 		}
1722 		vm_map_remove(kernel_map, kva, kva + size);
1723 	} else
1724 		vm_object_deallocate(obj);
1725 
1726 	/* On failure, drop our mapping reference. */
1727 	VM_OBJECT_WLOCK(obj);
1728 	shmfd->shm_kmappings--;
1729 	VM_OBJECT_WUNLOCK(obj);
1730 
1731 	return (vm_mmap_to_errno(rv));
1732 }
1733 
1734 /*
1735  * We require the caller to unmap the entire entry.  This allows us to
1736  * safely decrement shm_kmappings when a mapping is removed.
1737  */
1738 int
1739 shm_unmap(struct file *fp, void *mem, size_t size)
1740 {
1741 	struct shmfd *shmfd;
1742 	vm_map_entry_t entry;
1743 	vm_offset_t kva, ofs;
1744 	vm_object_t obj;
1745 	vm_pindex_t pindex;
1746 	vm_prot_t prot;
1747 	boolean_t wired;
1748 	vm_map_t map;
1749 	int rv;
1750 
1751 	if (fp->f_type != DTYPE_SHM)
1752 		return (EINVAL);
1753 	shmfd = fp->f_data;
1754 	kva = (vm_offset_t)mem;
1755 	ofs = kva & PAGE_MASK;
1756 	kva = trunc_page(kva);
1757 	size = round_page(size + ofs);
1758 	map = kernel_map;
1759 	rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1760 	    &obj, &pindex, &prot, &wired);
1761 	if (rv != KERN_SUCCESS)
1762 		return (EINVAL);
1763 	if (entry->start != kva || entry->end != kva + size) {
1764 		vm_map_lookup_done(map, entry);
1765 		return (EINVAL);
1766 	}
1767 	vm_map_lookup_done(map, entry);
1768 	if (obj != shmfd->shm_object)
1769 		return (EINVAL);
1770 	vm_map_remove(map, kva, kva + size);
1771 	VM_OBJECT_WLOCK(obj);
1772 	KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1773 	shmfd->shm_kmappings--;
1774 	VM_OBJECT_WUNLOCK(obj);
1775 	return (0);
1776 }
1777 
1778 static int
1779 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list)
1780 {
1781 	const char *path, *pr_path;
1782 	size_t pr_pathlen;
1783 	bool visible;
1784 
1785 	sx_assert(&shm_dict_lock, SA_LOCKED);
1786 	kif->kf_type = KF_TYPE_SHM;
1787 	kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;
1788 	kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1789 	if (shmfd->shm_path != NULL) {
1790 		if (shmfd->shm_path != NULL) {
1791 			path = shmfd->shm_path;
1792 			pr_path = curthread->td_ucred->cr_prison->pr_path;
1793 			if (strcmp(pr_path, "/") != 0) {
1794 				/* Return the jail-rooted pathname. */
1795 				pr_pathlen = strlen(pr_path);
1796 				visible = strncmp(path, pr_path, pr_pathlen)
1797 				    == 0 && path[pr_pathlen] == '/';
1798 				if (list && !visible)
1799 					return (EPERM);
1800 				if (visible)
1801 					path += pr_pathlen;
1802 			}
1803 			strlcpy(kif->kf_path, path, sizeof(kif->kf_path));
1804 		}
1805 	}
1806 	return (0);
1807 }
1808 
1809 static int
1810 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif,
1811     struct filedesc *fdp __unused)
1812 {
1813 	int res;
1814 
1815 	sx_slock(&shm_dict_lock);
1816 	res = shm_fill_kinfo_locked(fp->f_data, kif, false);
1817 	sx_sunlock(&shm_dict_lock);
1818 	return (res);
1819 }
1820 
1821 static int
1822 shm_add_seals(struct file *fp, int seals)
1823 {
1824 	struct shmfd *shmfd;
1825 	void *rl_cookie;
1826 	vm_ooffset_t writemappings;
1827 	int error, nseals;
1828 
1829 	error = 0;
1830 	shmfd = fp->f_data;
1831 	rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
1832 	    &shmfd->shm_mtx);
1833 
1834 	/* Even already-set seals should result in EPERM. */
1835 	if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) {
1836 		error = EPERM;
1837 		goto out;
1838 	}
1839 	nseals = seals & ~shmfd->shm_seals;
1840 	if ((nseals & F_SEAL_WRITE) != 0) {
1841 		/*
1842 		 * The rangelock above prevents writable mappings from being
1843 		 * added after we've started applying seals.  The RLOCK here
1844 		 * is to avoid torn reads on ILP32 arches as unmapping/reducing
1845 		 * writemappings will be done without a rangelock.
1846 		 */
1847 		VM_OBJECT_RLOCK(shmfd->shm_object);
1848 		writemappings = shmfd->shm_object->un_pager.swp.writemappings;
1849 		VM_OBJECT_RUNLOCK(shmfd->shm_object);
1850 		/* kmappings are also writable */
1851 		if (writemappings > 0) {
1852 			error = EBUSY;
1853 			goto out;
1854 		}
1855 	}
1856 	shmfd->shm_seals |= nseals;
1857 out:
1858 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
1859 	return (error);
1860 }
1861 
1862 static int
1863 shm_get_seals(struct file *fp, int *seals)
1864 {
1865 	struct shmfd *shmfd;
1866 
1867 	shmfd = fp->f_data;
1868 	*seals = shmfd->shm_seals;
1869 	return (0);
1870 }
1871 
1872 static int
1873 shm_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
1874 {
1875 	void *rl_cookie;
1876 	struct shmfd *shmfd;
1877 	size_t size;
1878 	int error;
1879 
1880 	/* This assumes that the caller already checked for overflow. */
1881 	error = 0;
1882 	shmfd = fp->f_data;
1883 	size = offset + len;
1884 
1885 	/*
1886 	 * Just grab the rangelock for the range that we may be attempting to
1887 	 * grow, rather than blocking read/write for regions we won't be
1888 	 * touching while this (potential) resize is in progress.  Other
1889 	 * attempts to resize the shmfd will have to take a write lock from 0 to
1890 	 * OFF_MAX, so this being potentially beyond the current usable range of
1891 	 * the shmfd is not necessarily a concern.  If other mechanisms are
1892 	 * added to grow a shmfd, this may need to be re-evaluated.
1893 	 */
1894 	rl_cookie = rangelock_wlock(&shmfd->shm_rl, offset, size,
1895 	    &shmfd->shm_mtx);
1896 	if (size > shmfd->shm_size)
1897 		error = shm_dotruncate_cookie(shmfd, size, rl_cookie);
1898 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
1899 	/* Translate to posix_fallocate(2) return value as needed. */
1900 	if (error == ENOMEM)
1901 		error = ENOSPC;
1902 	return (error);
1903 }
1904 
1905 static int
1906 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS)
1907 {
1908 	struct shm_mapping *shmm;
1909 	struct sbuf sb;
1910 	struct kinfo_file kif;
1911 	u_long i;
1912 	ssize_t curlen;
1913 	int error, error2;
1914 
1915 	sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req);
1916 	sbuf_clear_flags(&sb, SBUF_INCLUDENUL);
1917 	curlen = 0;
1918 	error = 0;
1919 	sx_slock(&shm_dict_lock);
1920 	for (i = 0; i < shm_hash + 1; i++) {
1921 		LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) {
1922 			error = shm_fill_kinfo_locked(shmm->sm_shmfd,
1923 			    &kif, true);
1924 			if (error == EPERM)
1925 				continue;
1926 			if (error != 0)
1927 				break;
1928 			pack_kinfo(&kif);
1929 			if (req->oldptr != NULL &&
1930 			    kif.kf_structsize + curlen > req->oldlen)
1931 				break;
1932 			error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ?
1933 			    0 : ENOMEM;
1934 			if (error != 0)
1935 				break;
1936 			curlen += kif.kf_structsize;
1937 		}
1938 	}
1939 	sx_sunlock(&shm_dict_lock);
1940 	error2 = sbuf_finish(&sb);
1941 	sbuf_delete(&sb);
1942 	return (error != 0 ? error : error2);
1943 }
1944 
1945 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list,
1946     CTLFLAG_RD | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE,
1947     NULL, 0, sysctl_posix_shm_list, "",
1948     "POSIX SHM list");
1949 
1950 int
1951 kern_shm_open(struct thread *td, const char *path, int flags, mode_t mode,
1952     struct filecaps *caps)
1953 {
1954 
1955 	return (kern_shm_open2(td, path, flags, mode, 0, caps, NULL));
1956 }
1957 
1958 /*
1959  * This version of the shm_open() interface leaves CLOEXEC behavior up to the
1960  * caller, and libc will enforce it for the traditional shm_open() call.  This
1961  * allows other consumers, like memfd_create(), to opt-in for CLOEXEC.  This
1962  * interface also includes a 'name' argument that is currently unused, but could
1963  * potentially be exported later via some interface for debugging purposes.
1964  * From the kernel's perspective, it is optional.  Individual consumers like
1965  * memfd_create() may require it in order to be compatible with other systems
1966  * implementing the same function.
1967  */
1968 int
1969 sys_shm_open2(struct thread *td, struct shm_open2_args *uap)
1970 {
1971 
1972 	return (kern_shm_open2(td, uap->path, uap->flags, uap->mode,
1973 	    uap->shmflags, NULL, uap->name));
1974 }
1975