xref: /freebsd/sys/kern/uipc_shm.c (revision 3f0efe05432b1633991114ca4ca330102a561959)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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 #include "opt_capsicum.h"
52 #include "opt_ktrace.h"
53 
54 #include <sys/param.h>
55 #include <sys/capsicum.h>
56 #include <sys/conf.h>
57 #include <sys/fcntl.h>
58 #include <sys/file.h>
59 #include <sys/filedesc.h>
60 #include <sys/filio.h>
61 #include <sys/fnv_hash.h>
62 #include <sys/kernel.h>
63 #include <sys/limits.h>
64 #include <sys/uio.h>
65 #include <sys/signal.h>
66 #include <sys/jail.h>
67 #include <sys/ktrace.h>
68 #include <sys/lock.h>
69 #include <sys/malloc.h>
70 #include <sys/mman.h>
71 #include <sys/mutex.h>
72 #include <sys/priv.h>
73 #include <sys/proc.h>
74 #include <sys/refcount.h>
75 #include <sys/resourcevar.h>
76 #include <sys/rwlock.h>
77 #include <sys/sbuf.h>
78 #include <sys/stat.h>
79 #include <sys/syscallsubr.h>
80 #include <sys/sysctl.h>
81 #include <sys/sysproto.h>
82 #include <sys/systm.h>
83 #include <sys/sx.h>
84 #include <sys/time.h>
85 #include <sys/vmmeter.h>
86 #include <sys/vnode.h>
87 #include <sys/unistd.h>
88 #include <sys/user.h>
89 
90 #include <security/audit/audit.h>
91 #include <security/mac/mac_framework.h>
92 
93 #include <vm/vm.h>
94 #include <vm/vm_param.h>
95 #include <vm/pmap.h>
96 #include <vm/vm_extern.h>
97 #include <vm/vm_map.h>
98 #include <vm/vm_kern.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_page.h>
101 #include <vm/vm_pageout.h>
102 #include <vm/vm_pager.h>
103 #include <vm/swap_pager.h>
104 
105 struct shm_mapping {
106 	char		*sm_path;
107 	Fnv32_t		sm_fnv;
108 	struct shmfd	*sm_shmfd;
109 	LIST_ENTRY(shm_mapping) sm_link;
110 };
111 
112 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor");
113 static LIST_HEAD(, shm_mapping) *shm_dictionary;
114 static struct sx shm_dict_lock;
115 static struct mtx shm_timestamp_lock;
116 static u_long shm_hash;
117 static struct unrhdr64 shm_ino_unr;
118 static dev_t shm_dev_ino;
119 
120 #define	SHM_HASH(fnv)	(&shm_dictionary[(fnv) & shm_hash])
121 
122 static void	shm_init(void *arg);
123 static void	shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd);
124 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv);
125 static int	shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred);
126 static void	shm_doremove(struct shm_mapping *map);
127 static int	shm_dotruncate_cookie(struct shmfd *shmfd, off_t length,
128     void *rl_cookie);
129 static int	shm_dotruncate_locked(struct shmfd *shmfd, off_t length,
130     void *rl_cookie);
131 static int	shm_copyin_path(struct thread *td, const char *userpath_in,
132     char **path_out);
133 static int	shm_deallocate(struct shmfd *shmfd, off_t *offset,
134     off_t *length, int flags);
135 
136 static fo_rdwr_t	shm_read;
137 static fo_rdwr_t	shm_write;
138 static fo_truncate_t	shm_truncate;
139 static fo_ioctl_t	shm_ioctl;
140 static fo_stat_t	shm_stat;
141 static fo_close_t	shm_close;
142 static fo_chmod_t	shm_chmod;
143 static fo_chown_t	shm_chown;
144 static fo_seek_t	shm_seek;
145 static fo_fill_kinfo_t	shm_fill_kinfo;
146 static fo_mmap_t	shm_mmap;
147 static fo_get_seals_t	shm_get_seals;
148 static fo_add_seals_t	shm_add_seals;
149 static fo_fallocate_t	shm_fallocate;
150 static fo_fspacectl_t	shm_fspacectl;
151 
152 /* File descriptor operations. */
153 struct fileops shm_ops = {
154 	.fo_read = shm_read,
155 	.fo_write = shm_write,
156 	.fo_truncate = shm_truncate,
157 	.fo_ioctl = shm_ioctl,
158 	.fo_poll = invfo_poll,
159 	.fo_kqfilter = invfo_kqfilter,
160 	.fo_stat = shm_stat,
161 	.fo_close = shm_close,
162 	.fo_chmod = shm_chmod,
163 	.fo_chown = shm_chown,
164 	.fo_sendfile = vn_sendfile,
165 	.fo_seek = shm_seek,
166 	.fo_fill_kinfo = shm_fill_kinfo,
167 	.fo_mmap = shm_mmap,
168 	.fo_get_seals = shm_get_seals,
169 	.fo_add_seals = shm_add_seals,
170 	.fo_fallocate = shm_fallocate,
171 	.fo_fspacectl = shm_fspacectl,
172 	.fo_cmp = file_kcmp_generic,
173 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE,
174 };
175 
176 FEATURE(posix_shm, "POSIX shared memory");
177 
178 static SYSCTL_NODE(_vm, OID_AUTO, largepages, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
179     "");
180 
181 static int largepage_reclaim_tries = 1;
182 SYSCTL_INT(_vm_largepages, OID_AUTO, reclaim_tries,
183     CTLFLAG_RWTUN, &largepage_reclaim_tries, 0,
184     "Number of contig reclaims before giving up for default alloc policy");
185 
186 #define	shm_rangelock_unlock(shmfd, cookie)				\
187 	rangelock_unlock(&(shmfd)->shm_rl, (cookie))
188 #define	shm_rangelock_rlock(shmfd, start, end)				\
189 	rangelock_rlock(&(shmfd)->shm_rl, (start), (end))
190 #define	shm_rangelock_tryrlock(shmfd, start, end)			\
191 	rangelock_tryrlock(&(shmfd)->shm_rl, (start), (end))
192 #define	shm_rangelock_wlock(shmfd, start, end)				\
193 	rangelock_wlock(&(shmfd)->shm_rl, (start), (end))
194 
195 static int
196 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio)
197 {
198 	vm_page_t m;
199 	vm_pindex_t idx;
200 	size_t tlen;
201 	int error, offset, rv;
202 
203 	idx = OFF_TO_IDX(uio->uio_offset);
204 	offset = uio->uio_offset & PAGE_MASK;
205 	tlen = MIN(PAGE_SIZE - offset, len);
206 
207 	rv = vm_page_grab_valid_unlocked(&m, obj, idx,
208 	    VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOCREAT);
209 	if (rv == VM_PAGER_OK)
210 		goto found;
211 
212 	/*
213 	 * Read I/O without either a corresponding resident page or swap
214 	 * page: use zero_region.  This is intended to avoid instantiating
215 	 * pages on read from a sparse region.
216 	 */
217 	VM_OBJECT_WLOCK(obj);
218 	m = vm_page_lookup(obj, idx);
219 	if (uio->uio_rw == UIO_READ && m == NULL &&
220 	    !vm_pager_has_page(obj, idx, NULL, NULL)) {
221 		VM_OBJECT_WUNLOCK(obj);
222 		return (uiomove(__DECONST(void *, zero_region), tlen, uio));
223 	}
224 
225 	/*
226 	 * Although the tmpfs vnode lock is held here, it is
227 	 * nonetheless safe to sleep waiting for a free page.  The
228 	 * pageout daemon does not need to acquire the tmpfs vnode
229 	 * lock to page out tobj's pages because tobj is a OBJT_SWAP
230 	 * type object.
231 	 */
232 	rv = vm_page_grab_valid(&m, obj, idx,
233 	    VM_ALLOC_NORMAL | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY);
234 	if (rv != VM_PAGER_OK) {
235 		VM_OBJECT_WUNLOCK(obj);
236 		if (bootverbose) {
237 			printf("uiomove_object: vm_obj %p idx %jd "
238 			    "pager error %d\n", obj, idx, rv);
239 		}
240 		return (rv == VM_PAGER_AGAIN ? ENOSPC : EIO);
241 	}
242 	VM_OBJECT_WUNLOCK(obj);
243 
244 found:
245 	error = uiomove_fromphys(&m, offset, tlen, uio);
246 	if (uio->uio_rw == UIO_WRITE && error == 0)
247 		vm_page_set_dirty(m);
248 	vm_page_activate(m);
249 	vm_page_sunbusy(m);
250 
251 	return (error);
252 }
253 
254 int
255 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
256 {
257 	ssize_t resid;
258 	size_t len;
259 	int error;
260 
261 	error = 0;
262 	while ((resid = uio->uio_resid) > 0) {
263 		if (obj_size <= uio->uio_offset)
264 			break;
265 		len = MIN(obj_size - uio->uio_offset, resid);
266 		if (len == 0)
267 			break;
268 		error = uiomove_object_page(obj, len, uio);
269 		if (error != 0 || resid == uio->uio_resid)
270 			break;
271 	}
272 	return (error);
273 }
274 
275 static u_long count_largepages[MAXPAGESIZES];
276 
277 static int
278 shm_largepage_phys_populate(vm_object_t object, vm_pindex_t pidx,
279     int fault_type, vm_prot_t max_prot, vm_pindex_t *first, vm_pindex_t *last)
280 {
281 	vm_page_t m __diagused;
282 	int psind;
283 
284 	psind = object->un_pager.phys.data_val;
285 	if (psind == 0 || pidx >= object->size)
286 		return (VM_PAGER_FAIL);
287 	*first = rounddown2(pidx, pagesizes[psind] / PAGE_SIZE);
288 
289 	/*
290 	 * We only busy the first page in the superpage run.  It is
291 	 * useless to busy whole run since we only remove full
292 	 * superpage, and it takes too long to busy e.g. 512 * 512 ==
293 	 * 262144 pages constituing 1G amd64 superage.
294 	 */
295 	m = vm_page_grab(object, *first, VM_ALLOC_NORMAL | VM_ALLOC_NOCREAT);
296 	MPASS(m != NULL);
297 
298 	*last = *first + atop(pagesizes[psind]) - 1;
299 	return (VM_PAGER_OK);
300 }
301 
302 static boolean_t
303 shm_largepage_phys_haspage(vm_object_t object, vm_pindex_t pindex,
304     int *before, int *after)
305 {
306 	int psind;
307 
308 	psind = object->un_pager.phys.data_val;
309 	if (psind == 0 || pindex >= object->size)
310 		return (FALSE);
311 	if (before != NULL) {
312 		*before = pindex - rounddown2(pindex, pagesizes[psind] /
313 		    PAGE_SIZE);
314 	}
315 	if (after != NULL) {
316 		*after = roundup2(pindex, pagesizes[psind] / PAGE_SIZE) -
317 		    pindex;
318 	}
319 	return (TRUE);
320 }
321 
322 static void
323 shm_largepage_phys_ctor(vm_object_t object, vm_prot_t prot,
324     vm_ooffset_t foff, struct ucred *cred)
325 {
326 }
327 
328 static void
329 shm_largepage_phys_dtor(vm_object_t object)
330 {
331 	int psind;
332 
333 	psind = object->un_pager.phys.data_val;
334 	if (psind != 0) {
335 		atomic_subtract_long(&count_largepages[psind],
336 		    object->size / (pagesizes[psind] / PAGE_SIZE));
337 		vm_wire_sub(object->size);
338 	} else {
339 		KASSERT(object->size == 0,
340 		    ("largepage phys obj %p not initialized bit size %#jx > 0",
341 		    object, (uintmax_t)object->size));
342 	}
343 }
344 
345 static const struct phys_pager_ops shm_largepage_phys_ops = {
346 	.phys_pg_populate =	shm_largepage_phys_populate,
347 	.phys_pg_haspage =	shm_largepage_phys_haspage,
348 	.phys_pg_ctor =		shm_largepage_phys_ctor,
349 	.phys_pg_dtor =		shm_largepage_phys_dtor,
350 };
351 
352 bool
353 shm_largepage(struct shmfd *shmfd)
354 {
355 	return (shmfd->shm_object->type == OBJT_PHYS);
356 }
357 
358 static void
359 shm_pager_freespace(vm_object_t obj, vm_pindex_t start, vm_size_t size)
360 {
361 	struct shmfd *shm;
362 	vm_size_t c;
363 
364 	swap_pager_freespace(obj, start, size, &c);
365 	if (c == 0)
366 		return;
367 
368 	shm = obj->un_pager.swp.swp_priv;
369 	if (shm == NULL)
370 		return;
371 	KASSERT(shm->shm_pages >= c,
372 	    ("shm %p pages %jd free %jd", shm,
373 	    (uintmax_t)shm->shm_pages, (uintmax_t)c));
374 	shm->shm_pages -= c;
375 }
376 
377 static void
378 shm_page_inserted(vm_object_t obj, vm_page_t m)
379 {
380 	struct shmfd *shm;
381 
382 	shm = obj->un_pager.swp.swp_priv;
383 	if (shm == NULL)
384 		return;
385 	if (!vm_pager_has_page(obj, m->pindex, NULL, NULL))
386 		shm->shm_pages += 1;
387 }
388 
389 static void
390 shm_page_removed(vm_object_t obj, vm_page_t m)
391 {
392 	struct shmfd *shm;
393 
394 	shm = obj->un_pager.swp.swp_priv;
395 	if (shm == NULL)
396 		return;
397 	if (!vm_pager_has_page(obj, m->pindex, NULL, NULL)) {
398 		KASSERT(shm->shm_pages >= 1,
399 		    ("shm %p pages %jd free 1", shm,
400 		    (uintmax_t)shm->shm_pages));
401 		shm->shm_pages -= 1;
402 	}
403 }
404 
405 static struct pagerops shm_swap_pager_ops = {
406 	.pgo_kvme_type = KVME_TYPE_SWAP,
407 	.pgo_freespace = shm_pager_freespace,
408 	.pgo_page_inserted = shm_page_inserted,
409 	.pgo_page_removed = shm_page_removed,
410 };
411 static int shmfd_pager_type = -1;
412 
413 static int
414 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
415 {
416 	struct shmfd *shmfd;
417 	off_t foffset;
418 	int error;
419 
420 	shmfd = fp->f_data;
421 	foffset = foffset_lock(fp, 0);
422 	error = 0;
423 	switch (whence) {
424 	case L_INCR:
425 		if (foffset < 0 ||
426 		    (offset > 0 && foffset > OFF_MAX - offset)) {
427 			error = EOVERFLOW;
428 			break;
429 		}
430 		offset += foffset;
431 		break;
432 	case L_XTND:
433 		if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
434 			error = EOVERFLOW;
435 			break;
436 		}
437 		offset += shmfd->shm_size;
438 		break;
439 	case L_SET:
440 		break;
441 	default:
442 		error = EINVAL;
443 	}
444 	if (error == 0) {
445 		if (offset < 0 || offset > shmfd->shm_size)
446 			error = EINVAL;
447 		else
448 			td->td_uretoff.tdu_off = offset;
449 	}
450 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
451 	return (error);
452 }
453 
454 static int
455 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
456     int flags, struct thread *td)
457 {
458 	struct shmfd *shmfd;
459 	void *rl_cookie;
460 	int error;
461 
462 	shmfd = fp->f_data;
463 #ifdef MAC
464 	error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
465 	if (error)
466 		return (error);
467 #endif
468 	foffset_lock_uio(fp, uio, flags);
469 	rl_cookie = shm_rangelock_rlock(shmfd, uio->uio_offset,
470 	    uio->uio_offset + uio->uio_resid);
471 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
472 	shm_rangelock_unlock(shmfd, rl_cookie);
473 	foffset_unlock_uio(fp, uio, flags);
474 	return (error);
475 }
476 
477 static int
478 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
479     int flags, struct thread *td)
480 {
481 	struct shmfd *shmfd;
482 	void *rl_cookie;
483 	int error;
484 	off_t size;
485 
486 	shmfd = fp->f_data;
487 #ifdef MAC
488 	error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
489 	if (error)
490 		return (error);
491 #endif
492 	if (shm_largepage(shmfd) && shmfd->shm_lp_psind == 0)
493 		return (EINVAL);
494 	foffset_lock_uio(fp, uio, flags);
495 	if (uio->uio_resid > OFF_MAX - uio->uio_offset) {
496 		/*
497 		 * Overflow is only an error if we're supposed to expand on
498 		 * write.  Otherwise, we'll just truncate the write to the
499 		 * size of the file, which can only grow up to OFF_MAX.
500 		 */
501 		if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0) {
502 			foffset_unlock_uio(fp, uio, flags);
503 			return (EFBIG);
504 		}
505 
506 		size = shmfd->shm_size;
507 	} else {
508 		size = uio->uio_offset + uio->uio_resid;
509 	}
510 	if ((flags & FOF_OFFSET) == 0)
511 		rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
512 	else
513 		rl_cookie = shm_rangelock_wlock(shmfd, uio->uio_offset, size);
514 	if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) {
515 		error = EPERM;
516 	} else {
517 		error = 0;
518 		if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0 &&
519 		    size > shmfd->shm_size) {
520 			error = shm_dotruncate_cookie(shmfd, size, rl_cookie);
521 		}
522 		if (error == 0)
523 			error = uiomove_object(shmfd->shm_object,
524 			    shmfd->shm_size, uio);
525 	}
526 	shm_rangelock_unlock(shmfd, rl_cookie);
527 	foffset_unlock_uio(fp, uio, flags);
528 	return (error);
529 }
530 
531 static int
532 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
533     struct thread *td)
534 {
535 	struct shmfd *shmfd;
536 #ifdef MAC
537 	int error;
538 #endif
539 
540 	shmfd = fp->f_data;
541 #ifdef MAC
542 	error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
543 	if (error)
544 		return (error);
545 #endif
546 	return (shm_dotruncate(shmfd, length));
547 }
548 
549 int
550 shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
551     struct thread *td)
552 {
553 	struct shmfd *shmfd;
554 	struct shm_largepage_conf *conf;
555 	void *rl_cookie;
556 
557 	shmfd = fp->f_data;
558 	switch (com) {
559 	case FIONBIO:
560 	case FIOASYNC:
561 		/*
562 		 * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work,
563 		 * just like it would on an unlinked regular file
564 		 */
565 		return (0);
566 	case FIOSSHMLPGCNF:
567 		if (!shm_largepage(shmfd))
568 			return (ENOTTY);
569 		conf = data;
570 		if (shmfd->shm_lp_psind != 0 &&
571 		    conf->psind != shmfd->shm_lp_psind)
572 			return (EINVAL);
573 		if (conf->psind <= 0 || conf->psind >= MAXPAGESIZES ||
574 		    pagesizes[conf->psind] == 0)
575 			return (EINVAL);
576 		if (conf->alloc_policy != SHM_LARGEPAGE_ALLOC_DEFAULT &&
577 		    conf->alloc_policy != SHM_LARGEPAGE_ALLOC_NOWAIT &&
578 		    conf->alloc_policy != SHM_LARGEPAGE_ALLOC_HARD)
579 			return (EINVAL);
580 
581 		rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
582 		shmfd->shm_lp_psind = conf->psind;
583 		shmfd->shm_lp_alloc_policy = conf->alloc_policy;
584 		shmfd->shm_object->un_pager.phys.data_val = conf->psind;
585 		shm_rangelock_unlock(shmfd, rl_cookie);
586 		return (0);
587 	case FIOGSHMLPGCNF:
588 		if (!shm_largepage(shmfd))
589 			return (ENOTTY);
590 		conf = data;
591 		rl_cookie = shm_rangelock_rlock(shmfd, 0, OFF_MAX);
592 		conf->psind = shmfd->shm_lp_psind;
593 		conf->alloc_policy = shmfd->shm_lp_alloc_policy;
594 		shm_rangelock_unlock(shmfd, rl_cookie);
595 		return (0);
596 	default:
597 		return (ENOTTY);
598 	}
599 }
600 
601 static int
602 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred)
603 {
604 	struct shmfd *shmfd;
605 #ifdef MAC
606 	int error;
607 #endif
608 
609 	shmfd = fp->f_data;
610 
611 #ifdef MAC
612 	error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
613 	if (error)
614 		return (error);
615 #endif
616 
617 	/*
618 	 * Attempt to return sanish values for fstat() on a memory file
619 	 * descriptor.
620 	 */
621 	bzero(sb, sizeof(*sb));
622 	sb->st_blksize = PAGE_SIZE;
623 	sb->st_size = shmfd->shm_size;
624 	mtx_lock(&shm_timestamp_lock);
625 	sb->st_atim = shmfd->shm_atime;
626 	sb->st_ctim = shmfd->shm_ctime;
627 	sb->st_mtim = shmfd->shm_mtime;
628 	sb->st_birthtim = shmfd->shm_birthtime;
629 	sb->st_mode = S_IFREG | shmfd->shm_mode;		/* XXX */
630 	sb->st_uid = shmfd->shm_uid;
631 	sb->st_gid = shmfd->shm_gid;
632 	mtx_unlock(&shm_timestamp_lock);
633 	sb->st_dev = shm_dev_ino;
634 	sb->st_ino = shmfd->shm_ino;
635 	sb->st_nlink = shmfd->shm_object->ref_count;
636 	if (shm_largepage(shmfd)) {
637 		sb->st_blocks = shmfd->shm_object->size /
638 		    (pagesizes[shmfd->shm_lp_psind] >> PAGE_SHIFT);
639 	} else {
640 		sb->st_blocks = shmfd->shm_pages;
641 	}
642 
643 	return (0);
644 }
645 
646 static int
647 shm_close(struct file *fp, struct thread *td)
648 {
649 	struct shmfd *shmfd;
650 
651 	shmfd = fp->f_data;
652 	fp->f_data = NULL;
653 	shm_drop(shmfd);
654 
655 	return (0);
656 }
657 
658 static int
659 shm_copyin_path(struct thread *td, const char *userpath_in, char **path_out) {
660 	int error;
661 	char *path;
662 	const char *pr_path;
663 	size_t pr_pathlen;
664 
665 	path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
666 	pr_path = td->td_ucred->cr_prison->pr_path;
667 
668 	/* Construct a full pathname for jailed callers. */
669 	pr_pathlen = strcmp(pr_path, "/") ==
670 	    0 ? 0 : strlcpy(path, pr_path, MAXPATHLEN);
671 	error = copyinstr(userpath_in, path + pr_pathlen,
672 	    MAXPATHLEN - pr_pathlen, NULL);
673 	if (error != 0)
674 		goto out;
675 
676 #ifdef KTRACE
677 	if (KTRPOINT(curthread, KTR_NAMEI))
678 		ktrnamei(path);
679 #endif
680 
681 	/* Require paths to start with a '/' character. */
682 	if (path[pr_pathlen] != '/') {
683 		error = EINVAL;
684 		goto out;
685 	}
686 
687 	*path_out = path;
688 
689 out:
690 	if (error != 0)
691 		free(path, M_SHMFD);
692 
693 	return (error);
694 }
695 
696 static int
697 shm_partial_page_invalidate(vm_object_t object, vm_pindex_t idx, int base,
698     int end)
699 {
700 	vm_page_t m;
701 	int rv;
702 
703 	VM_OBJECT_ASSERT_WLOCKED(object);
704 	KASSERT(base >= 0, ("%s: base %d", __func__, base));
705 	KASSERT(end - base <= PAGE_SIZE, ("%s: base %d end %d", __func__, base,
706 	    end));
707 
708 retry:
709 	m = vm_page_grab(object, idx, VM_ALLOC_NOCREAT);
710 	if (m != NULL) {
711 		MPASS(vm_page_all_valid(m));
712 	} else if (vm_pager_has_page(object, idx, NULL, NULL)) {
713 		m = vm_page_alloc(object, idx,
714 		    VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL);
715 		if (m == NULL)
716 			goto retry;
717 		vm_object_pip_add(object, 1);
718 		VM_OBJECT_WUNLOCK(object);
719 		rv = vm_pager_get_pages(object, &m, 1, NULL, NULL);
720 		VM_OBJECT_WLOCK(object);
721 		vm_object_pip_wakeup(object);
722 		if (rv == VM_PAGER_OK) {
723 			/*
724 			 * Since the page was not resident, and therefore not
725 			 * recently accessed, immediately enqueue it for
726 			 * asynchronous laundering.  The current operation is
727 			 * not regarded as an access.
728 			 */
729 			vm_page_launder(m);
730 		} else {
731 			vm_page_free(m);
732 			VM_OBJECT_WUNLOCK(object);
733 			return (EIO);
734 		}
735 	}
736 	if (m != NULL) {
737 		pmap_zero_page_area(m, base, end - base);
738 		KASSERT(vm_page_all_valid(m), ("%s: page %p is invalid",
739 		    __func__, m));
740 		vm_page_set_dirty(m);
741 		vm_page_xunbusy(m);
742 	}
743 
744 	return (0);
745 }
746 
747 static int
748 shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie)
749 {
750 	vm_object_t object;
751 	vm_pindex_t nobjsize;
752 	vm_ooffset_t delta;
753 	int base, error;
754 
755 	KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
756 	object = shmfd->shm_object;
757 	VM_OBJECT_ASSERT_WLOCKED(object);
758 	rangelock_cookie_assert(rl_cookie, RA_WLOCKED);
759 	if (length == shmfd->shm_size)
760 		return (0);
761 	nobjsize = OFF_TO_IDX(length + PAGE_MASK);
762 
763 	/* Are we shrinking?  If so, trim the end. */
764 	if (length < shmfd->shm_size) {
765 		if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0)
766 			return (EPERM);
767 
768 		/*
769 		 * Disallow any requests to shrink the size if this
770 		 * object is mapped into the kernel.
771 		 */
772 		if (shmfd->shm_kmappings > 0)
773 			return (EBUSY);
774 
775 		/*
776 		 * Zero the truncated part of the last page.
777 		 */
778 		base = length & PAGE_MASK;
779 		if (base != 0) {
780 			error = shm_partial_page_invalidate(object,
781 			    OFF_TO_IDX(length), base, PAGE_SIZE);
782 			if (error)
783 				return (error);
784 		}
785 		delta = IDX_TO_OFF(object->size - nobjsize);
786 
787 		if (nobjsize < object->size)
788 			vm_object_page_remove(object, nobjsize, object->size,
789 			    0);
790 
791 		/* Free the swap accounted for shm */
792 		swap_release_by_cred(delta, object->cred);
793 		object->charge -= delta;
794 	} else {
795 		if ((shmfd->shm_seals & F_SEAL_GROW) != 0)
796 			return (EPERM);
797 
798 		/* Try to reserve additional swap space. */
799 		delta = IDX_TO_OFF(nobjsize - object->size);
800 		if (!swap_reserve_by_cred(delta, object->cred))
801 			return (ENOMEM);
802 		object->charge += delta;
803 	}
804 	shmfd->shm_size = length;
805 	mtx_lock(&shm_timestamp_lock);
806 	vfs_timestamp(&shmfd->shm_ctime);
807 	shmfd->shm_mtime = shmfd->shm_ctime;
808 	mtx_unlock(&shm_timestamp_lock);
809 	object->size = nobjsize;
810 	return (0);
811 }
812 
813 static int
814 shm_dotruncate_largepage(struct shmfd *shmfd, off_t length, void *rl_cookie)
815 {
816 	vm_object_t object;
817 	vm_page_t m;
818 	vm_pindex_t newobjsz;
819 	vm_pindex_t oldobjsz __unused;
820 	int aflags, error, i, psind, try;
821 
822 	KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
823 	object = shmfd->shm_object;
824 	VM_OBJECT_ASSERT_WLOCKED(object);
825 	rangelock_cookie_assert(rl_cookie, RA_WLOCKED);
826 
827 	oldobjsz = object->size;
828 	newobjsz = OFF_TO_IDX(length);
829 	if (length == shmfd->shm_size)
830 		return (0);
831 	psind = shmfd->shm_lp_psind;
832 	if (psind == 0 && length != 0)
833 		return (EINVAL);
834 	if ((length & (pagesizes[psind] - 1)) != 0)
835 		return (EINVAL);
836 
837 	if (length < shmfd->shm_size) {
838 		if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0)
839 			return (EPERM);
840 		if (shmfd->shm_kmappings > 0)
841 			return (EBUSY);
842 		return (ENOTSUP);	/* Pages are unmanaged. */
843 #if 0
844 		vm_object_page_remove(object, newobjsz, oldobjsz, 0);
845 		object->size = newobjsz;
846 		shmfd->shm_size = length;
847 		return (0);
848 #endif
849 	}
850 
851 	if ((shmfd->shm_seals & F_SEAL_GROW) != 0)
852 		return (EPERM);
853 
854 	aflags = VM_ALLOC_NORMAL | VM_ALLOC_ZERO;
855 	if (shmfd->shm_lp_alloc_policy == SHM_LARGEPAGE_ALLOC_NOWAIT)
856 		aflags |= VM_ALLOC_WAITFAIL;
857 	try = 0;
858 
859 	/*
860 	 * Extend shmfd and object, keeping all already fully
861 	 * allocated large pages intact even on error, because dropped
862 	 * object lock might allowed mapping of them.
863 	 */
864 	while (object->size < newobjsz) {
865 		m = vm_page_alloc_contig(object, object->size, aflags,
866 		    pagesizes[psind] / PAGE_SIZE, 0, ~0,
867 		    pagesizes[psind], 0,
868 		    VM_MEMATTR_DEFAULT);
869 		if (m == NULL) {
870 			VM_OBJECT_WUNLOCK(object);
871 			if (shmfd->shm_lp_alloc_policy ==
872 			    SHM_LARGEPAGE_ALLOC_NOWAIT ||
873 			    (shmfd->shm_lp_alloc_policy ==
874 			    SHM_LARGEPAGE_ALLOC_DEFAULT &&
875 			    try >= largepage_reclaim_tries)) {
876 				VM_OBJECT_WLOCK(object);
877 				return (ENOMEM);
878 			}
879 			error = vm_page_reclaim_contig(aflags,
880 			    pagesizes[psind] / PAGE_SIZE, 0, ~0,
881 			    pagesizes[psind], 0);
882 			if (error == ENOMEM)
883 				error = vm_wait_intr(object);
884 			if (error != 0) {
885 				VM_OBJECT_WLOCK(object);
886 				return (error);
887 			}
888 			try++;
889 			VM_OBJECT_WLOCK(object);
890 			continue;
891 		}
892 		try = 0;
893 		for (i = 0; i < pagesizes[psind] / PAGE_SIZE; i++) {
894 			if ((m[i].flags & PG_ZERO) == 0)
895 				pmap_zero_page(&m[i]);
896 			vm_page_valid(&m[i]);
897 			vm_page_xunbusy(&m[i]);
898 		}
899 		object->size += OFF_TO_IDX(pagesizes[psind]);
900 		shmfd->shm_size += pagesizes[psind];
901 		atomic_add_long(&count_largepages[psind], 1);
902 		vm_wire_add(atop(pagesizes[psind]));
903 	}
904 	return (0);
905 }
906 
907 static int
908 shm_dotruncate_cookie(struct shmfd *shmfd, off_t length, void *rl_cookie)
909 {
910 	int error;
911 
912 	VM_OBJECT_WLOCK(shmfd->shm_object);
913 	error = shm_largepage(shmfd) ? shm_dotruncate_largepage(shmfd,
914 	    length, rl_cookie) : shm_dotruncate_locked(shmfd, length,
915 	    rl_cookie);
916 	VM_OBJECT_WUNLOCK(shmfd->shm_object);
917 	return (error);
918 }
919 
920 int
921 shm_dotruncate(struct shmfd *shmfd, off_t length)
922 {
923 	void *rl_cookie;
924 	int error;
925 
926 	rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
927 	error = shm_dotruncate_cookie(shmfd, length, rl_cookie);
928 	shm_rangelock_unlock(shmfd, rl_cookie);
929 	return (error);
930 }
931 
932 /*
933  * shmfd object management including creation and reference counting
934  * routines.
935  */
936 struct shmfd *
937 shm_alloc(struct ucred *ucred, mode_t mode, bool largepage)
938 {
939 	struct shmfd *shmfd;
940 	vm_object_t obj;
941 
942 	shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
943 	shmfd->shm_size = 0;
944 	shmfd->shm_uid = ucred->cr_uid;
945 	shmfd->shm_gid = ucred->cr_gid;
946 	shmfd->shm_mode = mode;
947 	if (largepage) {
948 		shmfd->shm_object = phys_pager_allocate(NULL,
949 		    &shm_largepage_phys_ops, NULL, shmfd->shm_size,
950 		    VM_PROT_DEFAULT, 0, ucred);
951 		shmfd->shm_lp_alloc_policy = SHM_LARGEPAGE_ALLOC_DEFAULT;
952 	} else {
953 		obj = vm_pager_allocate(shmfd_pager_type, NULL,
954 		    shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
955 		VM_OBJECT_WLOCK(obj);
956 		obj->un_pager.swp.swp_priv = shmfd;
957 		VM_OBJECT_WUNLOCK(obj);
958 		shmfd->shm_object = obj;
959 	}
960 	KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
961 	vfs_timestamp(&shmfd->shm_birthtime);
962 	shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
963 	    shmfd->shm_birthtime;
964 	shmfd->shm_ino = alloc_unr64(&shm_ino_unr);
965 	refcount_init(&shmfd->shm_refs, 1);
966 	mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
967 	rangelock_init(&shmfd->shm_rl);
968 #ifdef MAC
969 	mac_posixshm_init(shmfd);
970 	mac_posixshm_create(ucred, shmfd);
971 #endif
972 
973 	return (shmfd);
974 }
975 
976 struct shmfd *
977 shm_hold(struct shmfd *shmfd)
978 {
979 
980 	refcount_acquire(&shmfd->shm_refs);
981 	return (shmfd);
982 }
983 
984 void
985 shm_drop(struct shmfd *shmfd)
986 {
987 	vm_object_t obj;
988 
989 	if (refcount_release(&shmfd->shm_refs)) {
990 #ifdef MAC
991 		mac_posixshm_destroy(shmfd);
992 #endif
993 		rangelock_destroy(&shmfd->shm_rl);
994 		mtx_destroy(&shmfd->shm_mtx);
995 		obj = shmfd->shm_object;
996 		if (!shm_largepage(shmfd)) {
997 			VM_OBJECT_WLOCK(obj);
998 			obj->un_pager.swp.swp_priv = NULL;
999 			VM_OBJECT_WUNLOCK(obj);
1000 		}
1001 		vm_object_deallocate(obj);
1002 		free(shmfd, M_SHMFD);
1003 	}
1004 }
1005 
1006 /*
1007  * Determine if the credentials have sufficient permissions for a
1008  * specified combination of FREAD and FWRITE.
1009  */
1010 int
1011 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
1012 {
1013 	accmode_t accmode;
1014 	int error;
1015 
1016 	accmode = 0;
1017 	if (flags & FREAD)
1018 		accmode |= VREAD;
1019 	if (flags & FWRITE)
1020 		accmode |= VWRITE;
1021 	mtx_lock(&shm_timestamp_lock);
1022 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
1023 	    accmode, ucred);
1024 	mtx_unlock(&shm_timestamp_lock);
1025 	return (error);
1026 }
1027 
1028 static void
1029 shm_init(void *arg)
1030 {
1031 	char name[32];
1032 	int i;
1033 
1034 	mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
1035 	sx_init(&shm_dict_lock, "shm dictionary");
1036 	shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
1037 	new_unrhdr64(&shm_ino_unr, 1);
1038 	shm_dev_ino = devfs_alloc_cdp_inode();
1039 	KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
1040 	shmfd_pager_type = vm_pager_alloc_dyn_type(&shm_swap_pager_ops,
1041 	    OBJT_SWAP);
1042 	MPASS(shmfd_pager_type != -1);
1043 
1044 	for (i = 1; i < MAXPAGESIZES; i++) {
1045 		if (pagesizes[i] == 0)
1046 			break;
1047 #define	M	(1024 * 1024)
1048 #define	G	(1024 * M)
1049 		if (pagesizes[i] >= G)
1050 			snprintf(name, sizeof(name), "%luG", pagesizes[i] / G);
1051 		else if (pagesizes[i] >= M)
1052 			snprintf(name, sizeof(name), "%luM", pagesizes[i] / M);
1053 		else
1054 			snprintf(name, sizeof(name), "%lu", pagesizes[i]);
1055 #undef G
1056 #undef M
1057 		SYSCTL_ADD_ULONG(NULL, SYSCTL_STATIC_CHILDREN(_vm_largepages),
1058 		    OID_AUTO, name, CTLFLAG_RD, &count_largepages[i],
1059 		    "number of non-transient largepages allocated");
1060 	}
1061 }
1062 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
1063 
1064 /*
1065  * Remove all shared memory objects that belong to a prison.
1066  */
1067 void
1068 shm_remove_prison(struct prison *pr)
1069 {
1070 	struct shm_mapping *shmm, *tshmm;
1071 	u_long i;
1072 
1073 	sx_xlock(&shm_dict_lock);
1074 	for (i = 0; i < shm_hash + 1; i++) {
1075 		LIST_FOREACH_SAFE(shmm, &shm_dictionary[i], sm_link, tshmm) {
1076 			if (shmm->sm_shmfd->shm_object->cred &&
1077 			    shmm->sm_shmfd->shm_object->cred->cr_prison == pr)
1078 				shm_doremove(shmm);
1079 		}
1080 	}
1081 	sx_xunlock(&shm_dict_lock);
1082 }
1083 
1084 /*
1085  * Dictionary management.  We maintain an in-kernel dictionary to map
1086  * paths to shmfd objects.  We use the FNV hash on the path to store
1087  * the mappings in a hash table.
1088  */
1089 static struct shmfd *
1090 shm_lookup(char *path, Fnv32_t fnv)
1091 {
1092 	struct shm_mapping *map;
1093 
1094 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
1095 		if (map->sm_fnv != fnv)
1096 			continue;
1097 		if (strcmp(map->sm_path, path) == 0)
1098 			return (map->sm_shmfd);
1099 	}
1100 
1101 	return (NULL);
1102 }
1103 
1104 static void
1105 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
1106 {
1107 	struct shm_mapping *map;
1108 
1109 	map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
1110 	map->sm_path = path;
1111 	map->sm_fnv = fnv;
1112 	map->sm_shmfd = shm_hold(shmfd);
1113 	shmfd->shm_path = path;
1114 	LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
1115 }
1116 
1117 static int
1118 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
1119 {
1120 	struct shm_mapping *map;
1121 	int error;
1122 
1123 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
1124 		if (map->sm_fnv != fnv)
1125 			continue;
1126 		if (strcmp(map->sm_path, path) == 0) {
1127 #ifdef MAC
1128 			error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
1129 			if (error)
1130 				return (error);
1131 #endif
1132 			error = shm_access(map->sm_shmfd, ucred,
1133 			    FREAD | FWRITE);
1134 			if (error)
1135 				return (error);
1136 			shm_doremove(map);
1137 			return (0);
1138 		}
1139 	}
1140 
1141 	return (ENOENT);
1142 }
1143 
1144 static void
1145 shm_doremove(struct shm_mapping *map)
1146 {
1147 	map->sm_shmfd->shm_path = NULL;
1148 	LIST_REMOVE(map, sm_link);
1149 	shm_drop(map->sm_shmfd);
1150 	free(map->sm_path, M_SHMFD);
1151 	free(map, M_SHMFD);
1152 }
1153 
1154 int
1155 kern_shm_open2(struct thread *td, const char *userpath, int flags, mode_t mode,
1156     int shmflags, struct filecaps *fcaps, const char *name __unused)
1157 {
1158 	struct pwddesc *pdp;
1159 	struct shmfd *shmfd;
1160 	struct file *fp;
1161 	char *path;
1162 	void *rl_cookie;
1163 	Fnv32_t fnv;
1164 	mode_t cmode;
1165 	int error, fd, initial_seals;
1166 	bool largepage;
1167 
1168 	if ((shmflags & ~(SHM_ALLOW_SEALING | SHM_GROW_ON_WRITE |
1169 	    SHM_LARGEPAGE)) != 0)
1170 		return (EINVAL);
1171 
1172 	initial_seals = F_SEAL_SEAL;
1173 	if ((shmflags & SHM_ALLOW_SEALING) != 0)
1174 		initial_seals &= ~F_SEAL_SEAL;
1175 
1176 	AUDIT_ARG_FFLAGS(flags);
1177 	AUDIT_ARG_MODE(mode);
1178 
1179 	if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
1180 		return (EINVAL);
1181 
1182 	if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
1183 		return (EINVAL);
1184 
1185 	largepage = (shmflags & SHM_LARGEPAGE) != 0;
1186 	if (largepage && !PMAP_HAS_LARGEPAGES)
1187 		return (ENOTTY);
1188 
1189 	/*
1190 	 * Currently only F_SEAL_SEAL may be set when creating or opening shmfd.
1191 	 * If the decision is made later to allow additional seals, care must be
1192 	 * taken below to ensure that the seals are properly set if the shmfd
1193 	 * already existed -- this currently assumes that only F_SEAL_SEAL can
1194 	 * be set and doesn't take further precautions to ensure the validity of
1195 	 * the seals being added with respect to current mappings.
1196 	 */
1197 	if ((initial_seals & ~F_SEAL_SEAL) != 0)
1198 		return (EINVAL);
1199 
1200 	if (userpath != SHM_ANON) {
1201 		error = shm_copyin_path(td, userpath, &path);
1202 		if (error != 0)
1203 			return (error);
1204 
1205 #ifdef CAPABILITY_MODE
1206 		/*
1207 		 * shm_open(2) is only allowed for anonymous objects.
1208 		 */
1209 		if (CAP_TRACING(td))
1210 			ktrcapfail(CAPFAIL_NAMEI, path);
1211 		if (IN_CAPABILITY_MODE(td)) {
1212 			free(path, M_SHMFD);
1213 			return (ECAPMODE);
1214 		}
1215 #endif
1216 
1217 		AUDIT_ARG_UPATH1_CANON(path);
1218 	} else {
1219 		path = NULL;
1220 	}
1221 
1222 	pdp = td->td_proc->p_pd;
1223 	cmode = (mode & ~pdp->pd_cmask) & ACCESSPERMS;
1224 
1225 	/*
1226 	 * shm_open(2) created shm should always have O_CLOEXEC set, as mandated
1227 	 * by POSIX.  We allow it to be unset here so that an in-kernel
1228 	 * interface may be written as a thin layer around shm, optionally not
1229 	 * setting CLOEXEC.  For shm_open(2), O_CLOEXEC is set unconditionally
1230 	 * in sys_shm_open() to keep this implementation compliant.
1231 	 */
1232 	error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps);
1233 	if (error) {
1234 		free(path, M_SHMFD);
1235 		return (error);
1236 	}
1237 
1238 	/* A SHM_ANON path pointer creates an anonymous object. */
1239 	if (userpath == SHM_ANON) {
1240 		/* A read-only anonymous object is pointless. */
1241 		if ((flags & O_ACCMODE) == O_RDONLY) {
1242 			fdclose(td, fp, fd);
1243 			fdrop(fp, td);
1244 			return (EINVAL);
1245 		}
1246 		shmfd = shm_alloc(td->td_ucred, cmode, largepage);
1247 		shmfd->shm_seals = initial_seals;
1248 		shmfd->shm_flags = shmflags;
1249 	} else {
1250 		fnv = fnv_32_str(path, FNV1_32_INIT);
1251 		sx_xlock(&shm_dict_lock);
1252 		shmfd = shm_lookup(path, fnv);
1253 		if (shmfd == NULL) {
1254 			/* Object does not yet exist, create it if requested. */
1255 			if (flags & O_CREAT) {
1256 #ifdef MAC
1257 				error = mac_posixshm_check_create(td->td_ucred,
1258 				    path);
1259 				if (error == 0) {
1260 #endif
1261 					shmfd = shm_alloc(td->td_ucred, cmode,
1262 					    largepage);
1263 					shmfd->shm_seals = initial_seals;
1264 					shmfd->shm_flags = shmflags;
1265 					shm_insert(path, fnv, shmfd);
1266 #ifdef MAC
1267 				}
1268 #endif
1269 			} else {
1270 				free(path, M_SHMFD);
1271 				error = ENOENT;
1272 			}
1273 		} else {
1274 			rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
1275 
1276 			/*
1277 			 * kern_shm_open() likely shouldn't ever error out on
1278 			 * trying to set a seal that already exists, unlike
1279 			 * F_ADD_SEALS.  This would break terribly as
1280 			 * shm_open(2) actually sets F_SEAL_SEAL to maintain
1281 			 * historical behavior where the underlying file could
1282 			 * not be sealed.
1283 			 */
1284 			initial_seals &= ~shmfd->shm_seals;
1285 
1286 			/*
1287 			 * Object already exists, obtain a new
1288 			 * reference if requested and permitted.
1289 			 */
1290 			free(path, M_SHMFD);
1291 
1292 			/*
1293 			 * initial_seals can't set additional seals if we've
1294 			 * already been set F_SEAL_SEAL.  If F_SEAL_SEAL is set,
1295 			 * then we've already removed that one from
1296 			 * initial_seals.  This is currently redundant as we
1297 			 * only allow setting F_SEAL_SEAL at creation time, but
1298 			 * it's cheap to check and decreases the effort required
1299 			 * to allow additional seals.
1300 			 */
1301 			if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 &&
1302 			    initial_seals != 0)
1303 				error = EPERM;
1304 			else if ((flags & (O_CREAT | O_EXCL)) ==
1305 			    (O_CREAT | O_EXCL))
1306 				error = EEXIST;
1307 			else if (shmflags != 0 && shmflags != shmfd->shm_flags)
1308 				error = EINVAL;
1309 			else {
1310 #ifdef MAC
1311 				error = mac_posixshm_check_open(td->td_ucred,
1312 				    shmfd, FFLAGS(flags & O_ACCMODE));
1313 				if (error == 0)
1314 #endif
1315 				error = shm_access(shmfd, td->td_ucred,
1316 				    FFLAGS(flags & O_ACCMODE));
1317 			}
1318 
1319 			/*
1320 			 * Truncate the file back to zero length if
1321 			 * O_TRUNC was specified and the object was
1322 			 * opened with read/write.
1323 			 */
1324 			if (error == 0 &&
1325 			    (flags & (O_ACCMODE | O_TRUNC)) ==
1326 			    (O_RDWR | O_TRUNC)) {
1327 				VM_OBJECT_WLOCK(shmfd->shm_object);
1328 #ifdef MAC
1329 				error = mac_posixshm_check_truncate(
1330 					td->td_ucred, fp->f_cred, shmfd);
1331 				if (error == 0)
1332 #endif
1333 					error = shm_dotruncate_locked(shmfd, 0,
1334 					    rl_cookie);
1335 				VM_OBJECT_WUNLOCK(shmfd->shm_object);
1336 			}
1337 			if (error == 0) {
1338 				/*
1339 				 * Currently we only allow F_SEAL_SEAL to be
1340 				 * set initially.  As noted above, this would
1341 				 * need to be reworked should that change.
1342 				 */
1343 				shmfd->shm_seals |= initial_seals;
1344 				shm_hold(shmfd);
1345 			}
1346 			shm_rangelock_unlock(shmfd, rl_cookie);
1347 		}
1348 		sx_xunlock(&shm_dict_lock);
1349 
1350 		if (error) {
1351 			fdclose(td, fp, fd);
1352 			fdrop(fp, td);
1353 			return (error);
1354 		}
1355 	}
1356 
1357 	finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
1358 
1359 	td->td_retval[0] = fd;
1360 	fdrop(fp, td);
1361 
1362 	return (0);
1363 }
1364 
1365 /* System calls. */
1366 #ifdef COMPAT_FREEBSD12
1367 int
1368 freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap)
1369 {
1370 
1371 	return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC,
1372 	    uap->mode, NULL));
1373 }
1374 #endif
1375 
1376 int
1377 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
1378 {
1379 	char *path;
1380 	Fnv32_t fnv;
1381 	int error;
1382 
1383 	error = shm_copyin_path(td, uap->path, &path);
1384 	if (error != 0)
1385 		return (error);
1386 
1387 	AUDIT_ARG_UPATH1_CANON(path);
1388 	fnv = fnv_32_str(path, FNV1_32_INIT);
1389 	sx_xlock(&shm_dict_lock);
1390 	error = shm_remove(path, fnv, td->td_ucred);
1391 	sx_xunlock(&shm_dict_lock);
1392 	free(path, M_SHMFD);
1393 
1394 	return (error);
1395 }
1396 
1397 int
1398 sys_shm_rename(struct thread *td, struct shm_rename_args *uap)
1399 {
1400 	char *path_from = NULL, *path_to = NULL;
1401 	Fnv32_t fnv_from, fnv_to;
1402 	struct shmfd *fd_from;
1403 	struct shmfd *fd_to;
1404 	int error;
1405 	int flags;
1406 
1407 	flags = uap->flags;
1408 	AUDIT_ARG_FFLAGS(flags);
1409 
1410 	/*
1411 	 * Make sure the user passed only valid flags.
1412 	 * If you add a new flag, please add a new term here.
1413 	 */
1414 	if ((flags & ~(
1415 	    SHM_RENAME_NOREPLACE |
1416 	    SHM_RENAME_EXCHANGE
1417 	    )) != 0) {
1418 		error = EINVAL;
1419 		goto out;
1420 	}
1421 
1422 	/*
1423 	 * EXCHANGE and NOREPLACE don't quite make sense together. Let's
1424 	 * force the user to choose one or the other.
1425 	 */
1426 	if ((flags & SHM_RENAME_NOREPLACE) != 0 &&
1427 	    (flags & SHM_RENAME_EXCHANGE) != 0) {
1428 		error = EINVAL;
1429 		goto out;
1430 	}
1431 
1432 	/* Renaming to or from anonymous makes no sense */
1433 	if (uap->path_from == SHM_ANON || uap->path_to == SHM_ANON) {
1434 		error = EINVAL;
1435 		goto out;
1436 	}
1437 
1438 	error = shm_copyin_path(td, uap->path_from, &path_from);
1439 	if (error != 0)
1440 		goto out;
1441 
1442 	error = shm_copyin_path(td, uap->path_to, &path_to);
1443 	if (error != 0)
1444 		goto out;
1445 
1446 	AUDIT_ARG_UPATH1_CANON(path_from);
1447 	AUDIT_ARG_UPATH2_CANON(path_to);
1448 
1449 	/* Rename with from/to equal is a no-op */
1450 	if (strcmp(path_from, path_to) == 0)
1451 		goto out;
1452 
1453 	fnv_from = fnv_32_str(path_from, FNV1_32_INIT);
1454 	fnv_to = fnv_32_str(path_to, FNV1_32_INIT);
1455 
1456 	sx_xlock(&shm_dict_lock);
1457 
1458 	fd_from = shm_lookup(path_from, fnv_from);
1459 	if (fd_from == NULL) {
1460 		error = ENOENT;
1461 		goto out_locked;
1462 	}
1463 
1464 	fd_to = shm_lookup(path_to, fnv_to);
1465 	if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) {
1466 		error = EEXIST;
1467 		goto out_locked;
1468 	}
1469 
1470 	/*
1471 	 * Unconditionally prevents shm_remove from invalidating the 'from'
1472 	 * shm's state.
1473 	 */
1474 	shm_hold(fd_from);
1475 	error = shm_remove(path_from, fnv_from, td->td_ucred);
1476 
1477 	/*
1478 	 * One of my assumptions failed if ENOENT (e.g. locking didn't
1479 	 * protect us)
1480 	 */
1481 	KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s",
1482 	    path_from));
1483 	if (error != 0) {
1484 		shm_drop(fd_from);
1485 		goto out_locked;
1486 	}
1487 
1488 	/*
1489 	 * If we are exchanging, we need to ensure the shm_remove below
1490 	 * doesn't invalidate the dest shm's state.
1491 	 */
1492 	if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL)
1493 		shm_hold(fd_to);
1494 
1495 	/*
1496 	 * NOTE: if path_to is not already in the hash, c'est la vie;
1497 	 * it simply means we have nothing already at path_to to unlink.
1498 	 * That is the ENOENT case.
1499 	 *
1500 	 * If we somehow don't have access to unlink this guy, but
1501 	 * did for the shm at path_from, then relink the shm to path_from
1502 	 * and abort with EACCES.
1503 	 *
1504 	 * All other errors: that is weird; let's relink and abort the
1505 	 * operation.
1506 	 */
1507 	error = shm_remove(path_to, fnv_to, td->td_ucred);
1508 	if (error != 0 && error != ENOENT) {
1509 		shm_insert(path_from, fnv_from, fd_from);
1510 		shm_drop(fd_from);
1511 		/* Don't free path_from now, since the hash references it */
1512 		path_from = NULL;
1513 		goto out_locked;
1514 	}
1515 
1516 	error = 0;
1517 
1518 	shm_insert(path_to, fnv_to, fd_from);
1519 
1520 	/* Don't free path_to now, since the hash references it */
1521 	path_to = NULL;
1522 
1523 	/* We kept a ref when we removed, and incremented again in insert */
1524 	shm_drop(fd_from);
1525 	KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1526 	    fd_from->shm_refs));
1527 
1528 	if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) {
1529 		shm_insert(path_from, fnv_from, fd_to);
1530 		path_from = NULL;
1531 		shm_drop(fd_to);
1532 		KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1533 		    fd_to->shm_refs));
1534 	}
1535 
1536 out_locked:
1537 	sx_xunlock(&shm_dict_lock);
1538 
1539 out:
1540 	free(path_from, M_SHMFD);
1541 	free(path_to, M_SHMFD);
1542 	return (error);
1543 }
1544 
1545 static int
1546 shm_mmap_large(struct shmfd *shmfd, vm_map_t map, vm_offset_t *addr,
1547     vm_size_t size, vm_prot_t prot, vm_prot_t max_prot, int flags,
1548     vm_ooffset_t foff, struct thread *td)
1549 {
1550 	struct vmspace *vms;
1551 	vm_map_entry_t next_entry, prev_entry;
1552 	vm_offset_t align, mask, maxaddr;
1553 	int docow, error, rv, try;
1554 	bool curmap;
1555 
1556 	if (shmfd->shm_lp_psind == 0)
1557 		return (EINVAL);
1558 
1559 	/* MAP_PRIVATE is disabled */
1560 	if ((flags & ~(MAP_SHARED | MAP_FIXED | MAP_EXCL |
1561 	    MAP_NOCORE | MAP_32BIT | MAP_ALIGNMENT_MASK)) != 0)
1562 		return (EINVAL);
1563 
1564 	vms = td->td_proc->p_vmspace;
1565 	curmap = map == &vms->vm_map;
1566 	if (curmap) {
1567 		error = kern_mmap_racct_check(td, map, size);
1568 		if (error != 0)
1569 			return (error);
1570 	}
1571 
1572 	docow = shmfd->shm_lp_psind << MAP_SPLIT_BOUNDARY_SHIFT;
1573 	docow |= MAP_INHERIT_SHARE;
1574 	if ((flags & MAP_NOCORE) != 0)
1575 		docow |= MAP_DISABLE_COREDUMP;
1576 
1577 	mask = pagesizes[shmfd->shm_lp_psind] - 1;
1578 	if ((foff & mask) != 0)
1579 		return (EINVAL);
1580 	maxaddr = vm_map_max(map);
1581 	if ((flags & MAP_32BIT) != 0 && maxaddr > MAP_32BIT_MAX_ADDR)
1582 		maxaddr = MAP_32BIT_MAX_ADDR;
1583 	if (size == 0 || (size & mask) != 0 ||
1584 	    (*addr != 0 && ((*addr & mask) != 0 ||
1585 	    *addr + size < *addr || *addr + size > maxaddr)))
1586 		return (EINVAL);
1587 
1588 	align = flags & MAP_ALIGNMENT_MASK;
1589 	if (align == 0) {
1590 		align = pagesizes[shmfd->shm_lp_psind];
1591 	} else if (align == MAP_ALIGNED_SUPER) {
1592 		/*
1593 		 * MAP_ALIGNED_SUPER is only supported on superpage sizes,
1594 		 * i.e., [1, VM_NRESERVLEVEL].  shmfd->shm_lp_psind < 1 is
1595 		 * handled above.
1596 		 */
1597 		if (
1598 #if VM_NRESERVLEVEL > 0
1599 		    shmfd->shm_lp_psind > VM_NRESERVLEVEL
1600 #else
1601 		    shmfd->shm_lp_psind > 1
1602 #endif
1603 		    )
1604 			return (EINVAL);
1605 		align = pagesizes[shmfd->shm_lp_psind];
1606 	} else {
1607 		align >>= MAP_ALIGNMENT_SHIFT;
1608 		align = 1ULL << align;
1609 		/* Also handles overflow. */
1610 		if (align < pagesizes[shmfd->shm_lp_psind])
1611 			return (EINVAL);
1612 	}
1613 
1614 	vm_map_lock(map);
1615 	if ((flags & MAP_FIXED) == 0) {
1616 		try = 1;
1617 		if (curmap && (*addr == 0 ||
1618 		    (*addr >= round_page((vm_offset_t)vms->vm_taddr) &&
1619 		    *addr < round_page((vm_offset_t)vms->vm_daddr +
1620 		    lim_max(td, RLIMIT_DATA))))) {
1621 			*addr = roundup2((vm_offset_t)vms->vm_daddr +
1622 			    lim_max(td, RLIMIT_DATA),
1623 			    pagesizes[shmfd->shm_lp_psind]);
1624 		}
1625 again:
1626 		rv = vm_map_find_aligned(map, addr, size, maxaddr, align);
1627 		if (rv != KERN_SUCCESS) {
1628 			if (try == 1) {
1629 				try = 2;
1630 				*addr = vm_map_min(map);
1631 				if ((*addr & mask) != 0)
1632 					*addr = (*addr + mask) & mask;
1633 				goto again;
1634 			}
1635 			goto fail1;
1636 		}
1637 	} else if ((flags & MAP_EXCL) == 0) {
1638 		rv = vm_map_delete(map, *addr, *addr + size);
1639 		if (rv != KERN_SUCCESS)
1640 			goto fail1;
1641 	} else {
1642 		error = ENOSPC;
1643 		if (vm_map_lookup_entry(map, *addr, &prev_entry))
1644 			goto fail;
1645 		next_entry = vm_map_entry_succ(prev_entry);
1646 		if (next_entry->start < *addr + size)
1647 			goto fail;
1648 	}
1649 
1650 	rv = vm_map_insert(map, shmfd->shm_object, foff, *addr, *addr + size,
1651 	    prot, max_prot, docow);
1652 fail1:
1653 	error = vm_mmap_to_errno(rv);
1654 fail:
1655 	vm_map_unlock(map);
1656 	return (error);
1657 }
1658 
1659 static int
1660 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
1661     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
1662     vm_ooffset_t foff, struct thread *td)
1663 {
1664 	struct shmfd *shmfd;
1665 	vm_prot_t maxprot;
1666 	int error;
1667 	bool writecnt;
1668 	void *rl_cookie;
1669 
1670 	shmfd = fp->f_data;
1671 	maxprot = VM_PROT_NONE;
1672 
1673 	rl_cookie = shm_rangelock_rlock(shmfd, 0, objsize);
1674 	/* FREAD should always be set. */
1675 	if ((fp->f_flag & FREAD) != 0)
1676 		maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
1677 
1678 	/*
1679 	 * If FWRITE's set, we can allow VM_PROT_WRITE unless it's a shared
1680 	 * mapping with a write seal applied.  Private mappings are always
1681 	 * writeable.
1682 	 */
1683 	if ((flags & MAP_SHARED) == 0) {
1684 		cap_maxprot |= VM_PROT_WRITE;
1685 		maxprot |= VM_PROT_WRITE;
1686 		writecnt = false;
1687 	} else {
1688 		if ((fp->f_flag & FWRITE) != 0 &&
1689 		    (shmfd->shm_seals & F_SEAL_WRITE) == 0)
1690 			maxprot |= VM_PROT_WRITE;
1691 
1692 		/*
1693 		 * Any mappings from a writable descriptor may be upgraded to
1694 		 * VM_PROT_WRITE with mprotect(2), unless a write-seal was
1695 		 * applied between the open and subsequent mmap(2).  We want to
1696 		 * reject application of a write seal as long as any such
1697 		 * mapping exists so that the seal cannot be trivially bypassed.
1698 		 */
1699 		writecnt = (maxprot & VM_PROT_WRITE) != 0;
1700 		if (!writecnt && (prot & VM_PROT_WRITE) != 0) {
1701 			error = EACCES;
1702 			goto out;
1703 		}
1704 	}
1705 	maxprot &= cap_maxprot;
1706 
1707 	/* See comment in vn_mmap(). */
1708 	if (
1709 #ifdef _LP64
1710 	    objsize > OFF_MAX ||
1711 #endif
1712 	    foff > OFF_MAX - objsize) {
1713 		error = EINVAL;
1714 		goto out;
1715 	}
1716 
1717 #ifdef MAC
1718 	error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
1719 	if (error != 0)
1720 		goto out;
1721 #endif
1722 
1723 	mtx_lock(&shm_timestamp_lock);
1724 	vfs_timestamp(&shmfd->shm_atime);
1725 	mtx_unlock(&shm_timestamp_lock);
1726 	vm_object_reference(shmfd->shm_object);
1727 
1728 	if (shm_largepage(shmfd)) {
1729 		writecnt = false;
1730 		error = shm_mmap_large(shmfd, map, addr, objsize, prot,
1731 		    maxprot, flags, foff, td);
1732 	} else {
1733 		if (writecnt) {
1734 			vm_pager_update_writecount(shmfd->shm_object, 0,
1735 			    objsize);
1736 		}
1737 		error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
1738 		    shmfd->shm_object, foff, writecnt, td);
1739 	}
1740 	if (error != 0) {
1741 		if (writecnt)
1742 			vm_pager_release_writecount(shmfd->shm_object, 0,
1743 			    objsize);
1744 		vm_object_deallocate(shmfd->shm_object);
1745 	}
1746 out:
1747 	shm_rangelock_unlock(shmfd, rl_cookie);
1748 	return (error);
1749 }
1750 
1751 static int
1752 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
1753     struct thread *td)
1754 {
1755 	struct shmfd *shmfd;
1756 	int error;
1757 
1758 	error = 0;
1759 	shmfd = fp->f_data;
1760 	mtx_lock(&shm_timestamp_lock);
1761 	/*
1762 	 * SUSv4 says that x bits of permission need not be affected.
1763 	 * Be consistent with our shm_open there.
1764 	 */
1765 #ifdef MAC
1766 	error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
1767 	if (error != 0)
1768 		goto out;
1769 #endif
1770 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
1771 	    VADMIN, active_cred);
1772 	if (error != 0)
1773 		goto out;
1774 	shmfd->shm_mode = mode & ACCESSPERMS;
1775 out:
1776 	mtx_unlock(&shm_timestamp_lock);
1777 	return (error);
1778 }
1779 
1780 static int
1781 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1782     struct thread *td)
1783 {
1784 	struct shmfd *shmfd;
1785 	int error;
1786 
1787 	error = 0;
1788 	shmfd = fp->f_data;
1789 	mtx_lock(&shm_timestamp_lock);
1790 #ifdef MAC
1791 	error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
1792 	if (error != 0)
1793 		goto out;
1794 #endif
1795 	if (uid == (uid_t)-1)
1796 		uid = shmfd->shm_uid;
1797 	if (gid == (gid_t)-1)
1798                  gid = shmfd->shm_gid;
1799 	if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
1800 	    (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
1801 	    (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN)))
1802 		goto out;
1803 	shmfd->shm_uid = uid;
1804 	shmfd->shm_gid = gid;
1805 out:
1806 	mtx_unlock(&shm_timestamp_lock);
1807 	return (error);
1808 }
1809 
1810 /*
1811  * Helper routines to allow the backing object of a shared memory file
1812  * descriptor to be mapped in the kernel.
1813  */
1814 int
1815 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
1816 {
1817 	struct shmfd *shmfd;
1818 	vm_offset_t kva, ofs;
1819 	vm_object_t obj;
1820 	int rv;
1821 
1822 	if (fp->f_type != DTYPE_SHM)
1823 		return (EINVAL);
1824 	shmfd = fp->f_data;
1825 	obj = shmfd->shm_object;
1826 	VM_OBJECT_WLOCK(obj);
1827 	/*
1828 	 * XXXRW: This validation is probably insufficient, and subject to
1829 	 * sign errors.  It should be fixed.
1830 	 */
1831 	if (offset >= shmfd->shm_size ||
1832 	    offset + size > round_page(shmfd->shm_size)) {
1833 		VM_OBJECT_WUNLOCK(obj);
1834 		return (EINVAL);
1835 	}
1836 
1837 	shmfd->shm_kmappings++;
1838 	vm_object_reference_locked(obj);
1839 	VM_OBJECT_WUNLOCK(obj);
1840 
1841 	/* Map the object into the kernel_map and wire it. */
1842 	kva = vm_map_min(kernel_map);
1843 	ofs = offset & PAGE_MASK;
1844 	offset = trunc_page(offset);
1845 	size = round_page(size + ofs);
1846 	rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1847 	    VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1848 	    VM_PROT_READ | VM_PROT_WRITE, 0);
1849 	if (rv == KERN_SUCCESS) {
1850 		rv = vm_map_wire(kernel_map, kva, kva + size,
1851 		    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1852 		if (rv == KERN_SUCCESS) {
1853 			*memp = (void *)(kva + ofs);
1854 			return (0);
1855 		}
1856 		vm_map_remove(kernel_map, kva, kva + size);
1857 	} else
1858 		vm_object_deallocate(obj);
1859 
1860 	/* On failure, drop our mapping reference. */
1861 	VM_OBJECT_WLOCK(obj);
1862 	shmfd->shm_kmappings--;
1863 	VM_OBJECT_WUNLOCK(obj);
1864 
1865 	return (vm_mmap_to_errno(rv));
1866 }
1867 
1868 /*
1869  * We require the caller to unmap the entire entry.  This allows us to
1870  * safely decrement shm_kmappings when a mapping is removed.
1871  */
1872 int
1873 shm_unmap(struct file *fp, void *mem, size_t size)
1874 {
1875 	struct shmfd *shmfd;
1876 	vm_map_entry_t entry;
1877 	vm_offset_t kva, ofs;
1878 	vm_object_t obj;
1879 	vm_pindex_t pindex;
1880 	vm_prot_t prot;
1881 	boolean_t wired;
1882 	vm_map_t map;
1883 	int rv;
1884 
1885 	if (fp->f_type != DTYPE_SHM)
1886 		return (EINVAL);
1887 	shmfd = fp->f_data;
1888 	kva = (vm_offset_t)mem;
1889 	ofs = kva & PAGE_MASK;
1890 	kva = trunc_page(kva);
1891 	size = round_page(size + ofs);
1892 	map = kernel_map;
1893 	rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1894 	    &obj, &pindex, &prot, &wired);
1895 	if (rv != KERN_SUCCESS)
1896 		return (EINVAL);
1897 	if (entry->start != kva || entry->end != kva + size) {
1898 		vm_map_lookup_done(map, entry);
1899 		return (EINVAL);
1900 	}
1901 	vm_map_lookup_done(map, entry);
1902 	if (obj != shmfd->shm_object)
1903 		return (EINVAL);
1904 	vm_map_remove(map, kva, kva + size);
1905 	VM_OBJECT_WLOCK(obj);
1906 	KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1907 	shmfd->shm_kmappings--;
1908 	VM_OBJECT_WUNLOCK(obj);
1909 	return (0);
1910 }
1911 
1912 static int
1913 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list)
1914 {
1915 	const char *path, *pr_path;
1916 	size_t pr_pathlen;
1917 	bool visible;
1918 
1919 	sx_assert(&shm_dict_lock, SA_LOCKED);
1920 	kif->kf_type = KF_TYPE_SHM;
1921 	kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;
1922 	kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1923 	if (shmfd->shm_path != NULL) {
1924 		path = shmfd->shm_path;
1925 		pr_path = curthread->td_ucred->cr_prison->pr_path;
1926 		if (strcmp(pr_path, "/") != 0) {
1927 			/* Return the jail-rooted pathname. */
1928 			pr_pathlen = strlen(pr_path);
1929 			visible = strncmp(path, pr_path, pr_pathlen) == 0 &&
1930 			    path[pr_pathlen] == '/';
1931 			if (list && !visible)
1932 				return (EPERM);
1933 			if (visible)
1934 				path += pr_pathlen;
1935 		}
1936 		strlcpy(kif->kf_path, path, sizeof(kif->kf_path));
1937 	}
1938 	return (0);
1939 }
1940 
1941 static int
1942 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif,
1943     struct filedesc *fdp __unused)
1944 {
1945 	int res;
1946 
1947 	sx_slock(&shm_dict_lock);
1948 	res = shm_fill_kinfo_locked(fp->f_data, kif, false);
1949 	sx_sunlock(&shm_dict_lock);
1950 	return (res);
1951 }
1952 
1953 static int
1954 shm_add_seals(struct file *fp, int seals)
1955 {
1956 	struct shmfd *shmfd;
1957 	void *rl_cookie;
1958 	vm_ooffset_t writemappings;
1959 	int error, nseals;
1960 
1961 	error = 0;
1962 	shmfd = fp->f_data;
1963 	rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
1964 
1965 	/* Even already-set seals should result in EPERM. */
1966 	if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) {
1967 		error = EPERM;
1968 		goto out;
1969 	}
1970 	nseals = seals & ~shmfd->shm_seals;
1971 	if ((nseals & F_SEAL_WRITE) != 0) {
1972 		if (shm_largepage(shmfd)) {
1973 			error = ENOTSUP;
1974 			goto out;
1975 		}
1976 
1977 		/*
1978 		 * The rangelock above prevents writable mappings from being
1979 		 * added after we've started applying seals.  The RLOCK here
1980 		 * is to avoid torn reads on ILP32 arches as unmapping/reducing
1981 		 * writemappings will be done without a rangelock.
1982 		 */
1983 		VM_OBJECT_RLOCK(shmfd->shm_object);
1984 		writemappings = shmfd->shm_object->un_pager.swp.writemappings;
1985 		VM_OBJECT_RUNLOCK(shmfd->shm_object);
1986 		/* kmappings are also writable */
1987 		if (writemappings > 0) {
1988 			error = EBUSY;
1989 			goto out;
1990 		}
1991 	}
1992 	shmfd->shm_seals |= nseals;
1993 out:
1994 	shm_rangelock_unlock(shmfd, rl_cookie);
1995 	return (error);
1996 }
1997 
1998 static int
1999 shm_get_seals(struct file *fp, int *seals)
2000 {
2001 	struct shmfd *shmfd;
2002 
2003 	shmfd = fp->f_data;
2004 	*seals = shmfd->shm_seals;
2005 	return (0);
2006 }
2007 
2008 static int
2009 shm_deallocate(struct shmfd *shmfd, off_t *offset, off_t *length, int flags)
2010 {
2011 	vm_object_t object;
2012 	vm_pindex_t pistart, pi, piend;
2013 	vm_ooffset_t off, len;
2014 	int startofs, endofs, end;
2015 	int error;
2016 
2017 	off = *offset;
2018 	len = *length;
2019 	KASSERT(off + len <= (vm_ooffset_t)OFF_MAX, ("off + len overflows"));
2020 	if (off + len > shmfd->shm_size)
2021 		len = shmfd->shm_size - off;
2022 	object = shmfd->shm_object;
2023 	startofs = off & PAGE_MASK;
2024 	endofs = (off + len) & PAGE_MASK;
2025 	pistart = OFF_TO_IDX(off);
2026 	piend = OFF_TO_IDX(off + len);
2027 	pi = OFF_TO_IDX(off + PAGE_MASK);
2028 	error = 0;
2029 
2030 	/* Handle the case when offset is on or beyond shm size. */
2031 	if ((off_t)len <= 0) {
2032 		*length = 0;
2033 		return (0);
2034 	}
2035 
2036 	VM_OBJECT_WLOCK(object);
2037 
2038 	if (startofs != 0) {
2039 		end = pistart != piend ? PAGE_SIZE : endofs;
2040 		error = shm_partial_page_invalidate(object, pistart, startofs,
2041 		    end);
2042 		if (error)
2043 			goto out;
2044 		off += end - startofs;
2045 		len -= end - startofs;
2046 	}
2047 
2048 	if (pi < piend) {
2049 		vm_object_page_remove(object, pi, piend, 0);
2050 		off += IDX_TO_OFF(piend - pi);
2051 		len -= IDX_TO_OFF(piend - pi);
2052 	}
2053 
2054 	if (endofs != 0 && pistart != piend) {
2055 		error = shm_partial_page_invalidate(object, piend, 0, endofs);
2056 		if (error)
2057 			goto out;
2058 		off += endofs;
2059 		len -= endofs;
2060 	}
2061 
2062 out:
2063 	VM_OBJECT_WUNLOCK(shmfd->shm_object);
2064 	*offset = off;
2065 	*length = len;
2066 	return (error);
2067 }
2068 
2069 static int
2070 shm_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags,
2071     struct ucred *active_cred, struct thread *td)
2072 {
2073 	void *rl_cookie;
2074 	struct shmfd *shmfd;
2075 	off_t off, len;
2076 	int error;
2077 
2078 	KASSERT(cmd == SPACECTL_DEALLOC, ("shm_fspacectl: Invalid cmd"));
2079 	KASSERT((flags & ~SPACECTL_F_SUPPORTED) == 0,
2080 	    ("shm_fspacectl: non-zero flags"));
2081 	KASSERT(*offset >= 0 && *length > 0 && *length <= OFF_MAX - *offset,
2082 	    ("shm_fspacectl: offset/length overflow or underflow"));
2083 	error = EINVAL;
2084 	shmfd = fp->f_data;
2085 	off = *offset;
2086 	len = *length;
2087 
2088 	rl_cookie = shm_rangelock_wlock(shmfd, off, off + len);
2089 	switch (cmd) {
2090 	case SPACECTL_DEALLOC:
2091 		if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) {
2092 			error = EPERM;
2093 			break;
2094 		}
2095 		error = shm_deallocate(shmfd, &off, &len, flags);
2096 		*offset = off;
2097 		*length = len;
2098 		break;
2099 	default:
2100 		__assert_unreachable();
2101 	}
2102 	shm_rangelock_unlock(shmfd, rl_cookie);
2103 	return (error);
2104 }
2105 
2106 
2107 static int
2108 shm_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
2109 {
2110 	void *rl_cookie;
2111 	struct shmfd *shmfd;
2112 	size_t size;
2113 	int error;
2114 
2115 	/* This assumes that the caller already checked for overflow. */
2116 	error = 0;
2117 	shmfd = fp->f_data;
2118 	size = offset + len;
2119 
2120 	/*
2121 	 * Just grab the rangelock for the range that we may be attempting to
2122 	 * grow, rather than blocking read/write for regions we won't be
2123 	 * touching while this (potential) resize is in progress.  Other
2124 	 * attempts to resize the shmfd will have to take a write lock from 0 to
2125 	 * OFF_MAX, so this being potentially beyond the current usable range of
2126 	 * the shmfd is not necessarily a concern.  If other mechanisms are
2127 	 * added to grow a shmfd, this may need to be re-evaluated.
2128 	 */
2129 	rl_cookie = shm_rangelock_wlock(shmfd, offset, size);
2130 	if (size > shmfd->shm_size)
2131 		error = shm_dotruncate_cookie(shmfd, size, rl_cookie);
2132 	shm_rangelock_unlock(shmfd, rl_cookie);
2133 	/* Translate to posix_fallocate(2) return value as needed. */
2134 	if (error == ENOMEM)
2135 		error = ENOSPC;
2136 	return (error);
2137 }
2138 
2139 static int
2140 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS)
2141 {
2142 	struct shm_mapping *shmm;
2143 	struct sbuf sb;
2144 	struct kinfo_file kif;
2145 	u_long i;
2146 	int error, error2;
2147 
2148 	sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req);
2149 	sbuf_clear_flags(&sb, SBUF_INCLUDENUL);
2150 	error = 0;
2151 	sx_slock(&shm_dict_lock);
2152 	for (i = 0; i < shm_hash + 1; i++) {
2153 		LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) {
2154 			error = shm_fill_kinfo_locked(shmm->sm_shmfd,
2155 			    &kif, true);
2156 			if (error == EPERM) {
2157 				error = 0;
2158 				continue;
2159 			}
2160 			if (error != 0)
2161 				break;
2162 			pack_kinfo(&kif);
2163 			error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ?
2164 			    0 : ENOMEM;
2165 			if (error != 0)
2166 				break;
2167 		}
2168 	}
2169 	sx_sunlock(&shm_dict_lock);
2170 	error2 = sbuf_finish(&sb);
2171 	sbuf_delete(&sb);
2172 	return (error != 0 ? error : error2);
2173 }
2174 
2175 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list,
2176     CTLFLAG_RD | CTLFLAG_PRISON | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE,
2177     NULL, 0, sysctl_posix_shm_list, "",
2178     "POSIX SHM list");
2179 
2180 int
2181 kern_shm_open(struct thread *td, const char *path, int flags, mode_t mode,
2182     struct filecaps *caps)
2183 {
2184 
2185 	return (kern_shm_open2(td, path, flags, mode, 0, caps, NULL));
2186 }
2187 
2188 /*
2189  * This version of the shm_open() interface leaves CLOEXEC behavior up to the
2190  * caller, and libc will enforce it for the traditional shm_open() call.  This
2191  * allows other consumers, like memfd_create(), to opt-in for CLOEXEC.  This
2192  * interface also includes a 'name' argument that is currently unused, but could
2193  * potentially be exported later via some interface for debugging purposes.
2194  * From the kernel's perspective, it is optional.  Individual consumers like
2195  * memfd_create() may require it in order to be compatible with other systems
2196  * implementing the same function.
2197  */
2198 int
2199 sys_shm_open2(struct thread *td, struct shm_open2_args *uap)
2200 {
2201 
2202 	return (kern_shm_open2(td, uap->path, uap->flags, uap->mode,
2203 	    uap->shmflags, NULL, uap->name));
2204 }
2205