xref: /freebsd/sys/kern/uipc_shm.c (revision 2c3f47a727372086b41ef9ce06eb1f1eb83a67d3)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2006, 2011, 2016-2017 Robert N. M. Watson
5  * All rights reserved.
6  *
7  * Portions of this software were developed by BAE Systems, the University of
8  * Cambridge Computer Laboratory, and Memorial University under DARPA/AFRL
9  * contract FA8650-15-C-7558 ("CADETS"), as part of the DARPA Transparent
10  * Computing (TC) research program.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 /*
35  * Support for shared swap-backed anonymous memory objects via
36  * shm_open(2) and shm_unlink(2).  While most of the implementation is
37  * here, vm_mmap.c contains mapping logic changes.
38  *
39  * posixshmcontrol(1) allows users to inspect the state of the memory
40  * objects.  Per-uid swap resource limit controls total amount of
41  * memory that user can consume for anonymous objects, including
42  * shared.
43  */
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include "opt_capsicum.h"
49 #include "opt_ktrace.h"
50 
51 #include <sys/param.h>
52 #include <sys/capsicum.h>
53 #include <sys/conf.h>
54 #include <sys/fcntl.h>
55 #include <sys/file.h>
56 #include <sys/filedesc.h>
57 #include <sys/filio.h>
58 #include <sys/fnv_hash.h>
59 #include <sys/kernel.h>
60 #include <sys/limits.h>
61 #include <sys/uio.h>
62 #include <sys/signal.h>
63 #include <sys/jail.h>
64 #include <sys/ktrace.h>
65 #include <sys/lock.h>
66 #include <sys/malloc.h>
67 #include <sys/mman.h>
68 #include <sys/mutex.h>
69 #include <sys/priv.h>
70 #include <sys/proc.h>
71 #include <sys/refcount.h>
72 #include <sys/resourcevar.h>
73 #include <sys/rwlock.h>
74 #include <sys/sbuf.h>
75 #include <sys/stat.h>
76 #include <sys/syscallsubr.h>
77 #include <sys/sysctl.h>
78 #include <sys/sysproto.h>
79 #include <sys/systm.h>
80 #include <sys/sx.h>
81 #include <sys/time.h>
82 #include <sys/vnode.h>
83 #include <sys/unistd.h>
84 #include <sys/user.h>
85 
86 #include <security/audit/audit.h>
87 #include <security/mac/mac_framework.h>
88 
89 #include <vm/vm.h>
90 #include <vm/vm_param.h>
91 #include <vm/pmap.h>
92 #include <vm/vm_extern.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_kern.h>
95 #include <vm/vm_object.h>
96 #include <vm/vm_page.h>
97 #include <vm/vm_pageout.h>
98 #include <vm/vm_pager.h>
99 #include <vm/swap_pager.h>
100 
101 struct shm_mapping {
102 	char		*sm_path;
103 	Fnv32_t		sm_fnv;
104 	struct shmfd	*sm_shmfd;
105 	LIST_ENTRY(shm_mapping) sm_link;
106 };
107 
108 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor");
109 static LIST_HEAD(, shm_mapping) *shm_dictionary;
110 static struct sx shm_dict_lock;
111 static struct mtx shm_timestamp_lock;
112 static u_long shm_hash;
113 static struct unrhdr64 shm_ino_unr;
114 static dev_t shm_dev_ino;
115 
116 #define	SHM_HASH(fnv)	(&shm_dictionary[(fnv) & shm_hash])
117 
118 static void	shm_init(void *arg);
119 static void	shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd);
120 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv);
121 static int	shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred);
122 
123 static fo_rdwr_t	shm_read;
124 static fo_rdwr_t	shm_write;
125 static fo_truncate_t	shm_truncate;
126 static fo_ioctl_t	shm_ioctl;
127 static fo_stat_t	shm_stat;
128 static fo_close_t	shm_close;
129 static fo_chmod_t	shm_chmod;
130 static fo_chown_t	shm_chown;
131 static fo_seek_t	shm_seek;
132 static fo_fill_kinfo_t	shm_fill_kinfo;
133 static fo_mmap_t	shm_mmap;
134 
135 /* File descriptor operations. */
136 struct fileops shm_ops = {
137 	.fo_read = shm_read,
138 	.fo_write = shm_write,
139 	.fo_truncate = shm_truncate,
140 	.fo_ioctl = shm_ioctl,
141 	.fo_poll = invfo_poll,
142 	.fo_kqfilter = invfo_kqfilter,
143 	.fo_stat = shm_stat,
144 	.fo_close = shm_close,
145 	.fo_chmod = shm_chmod,
146 	.fo_chown = shm_chown,
147 	.fo_sendfile = vn_sendfile,
148 	.fo_seek = shm_seek,
149 	.fo_fill_kinfo = shm_fill_kinfo,
150 	.fo_mmap = shm_mmap,
151 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
152 };
153 
154 FEATURE(posix_shm, "POSIX shared memory");
155 
156 static int
157 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio)
158 {
159 	vm_page_t m;
160 	vm_pindex_t idx;
161 	size_t tlen;
162 	int error, offset, rv;
163 
164 	idx = OFF_TO_IDX(uio->uio_offset);
165 	offset = uio->uio_offset & PAGE_MASK;
166 	tlen = MIN(PAGE_SIZE - offset, len);
167 
168 	VM_OBJECT_WLOCK(obj);
169 
170 	/*
171 	 * Read I/O without either a corresponding resident page or swap
172 	 * page: use zero_region.  This is intended to avoid instantiating
173 	 * pages on read from a sparse region.
174 	 */
175 	if (uio->uio_rw == UIO_READ && vm_page_lookup(obj, idx) == NULL &&
176 	    !vm_pager_has_page(obj, idx, NULL, NULL)) {
177 		VM_OBJECT_WUNLOCK(obj);
178 		return (uiomove(__DECONST(void *, zero_region), tlen, uio));
179 	}
180 
181 	/*
182 	 * Parallel reads of the page content from disk are prevented
183 	 * by exclusive busy.
184 	 *
185 	 * Although the tmpfs vnode lock is held here, it is
186 	 * nonetheless safe to sleep waiting for a free page.  The
187 	 * pageout daemon does not need to acquire the tmpfs vnode
188 	 * lock to page out tobj's pages because tobj is a OBJT_SWAP
189 	 * type object.
190 	 */
191 	rv = vm_page_grab_valid(&m, obj, idx,
192 	    VM_ALLOC_NORMAL | VM_ALLOC_WIRED | VM_ALLOC_NOBUSY);
193 	if (rv != VM_PAGER_OK) {
194 		VM_OBJECT_WUNLOCK(obj);
195 		printf("uiomove_object: vm_obj %p idx %jd pager error %d\n",
196 		    obj, idx, rv);
197 		return (EIO);
198 	}
199 	VM_OBJECT_WUNLOCK(obj);
200 	error = uiomove_fromphys(&m, offset, tlen, uio);
201 	if (uio->uio_rw == UIO_WRITE && error == 0) {
202 		VM_OBJECT_WLOCK(obj);
203 		vm_page_dirty(m);
204 		vm_pager_page_unswapped(m);
205 		VM_OBJECT_WUNLOCK(obj);
206 	}
207 	vm_page_unwire(m, PQ_ACTIVE);
208 
209 	return (error);
210 }
211 
212 int
213 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
214 {
215 	ssize_t resid;
216 	size_t len;
217 	int error;
218 
219 	error = 0;
220 	while ((resid = uio->uio_resid) > 0) {
221 		if (obj_size <= uio->uio_offset)
222 			break;
223 		len = MIN(obj_size - uio->uio_offset, resid);
224 		if (len == 0)
225 			break;
226 		error = uiomove_object_page(obj, len, uio);
227 		if (error != 0 || resid == uio->uio_resid)
228 			break;
229 	}
230 	return (error);
231 }
232 
233 static int
234 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
235 {
236 	struct shmfd *shmfd;
237 	off_t foffset;
238 	int error;
239 
240 	shmfd = fp->f_data;
241 	foffset = foffset_lock(fp, 0);
242 	error = 0;
243 	switch (whence) {
244 	case L_INCR:
245 		if (foffset < 0 ||
246 		    (offset > 0 && foffset > OFF_MAX - offset)) {
247 			error = EOVERFLOW;
248 			break;
249 		}
250 		offset += foffset;
251 		break;
252 	case L_XTND:
253 		if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
254 			error = EOVERFLOW;
255 			break;
256 		}
257 		offset += shmfd->shm_size;
258 		break;
259 	case L_SET:
260 		break;
261 	default:
262 		error = EINVAL;
263 	}
264 	if (error == 0) {
265 		if (offset < 0 || offset > shmfd->shm_size)
266 			error = EINVAL;
267 		else
268 			td->td_uretoff.tdu_off = offset;
269 	}
270 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
271 	return (error);
272 }
273 
274 static int
275 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
276     int flags, struct thread *td)
277 {
278 	struct shmfd *shmfd;
279 	void *rl_cookie;
280 	int error;
281 
282 	shmfd = fp->f_data;
283 #ifdef MAC
284 	error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
285 	if (error)
286 		return (error);
287 #endif
288 	foffset_lock_uio(fp, uio, flags);
289 	rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset,
290 	    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
291 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
292 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
293 	foffset_unlock_uio(fp, uio, flags);
294 	return (error);
295 }
296 
297 static int
298 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
299     int flags, struct thread *td)
300 {
301 	struct shmfd *shmfd;
302 	void *rl_cookie;
303 	int error;
304 
305 	shmfd = fp->f_data;
306 #ifdef MAC
307 	error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
308 	if (error)
309 		return (error);
310 #endif
311 	foffset_lock_uio(fp, uio, flags);
312 	if ((flags & FOF_OFFSET) == 0) {
313 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
314 		    &shmfd->shm_mtx);
315 	} else {
316 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset,
317 		    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
318 	}
319 
320 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
321 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
322 	foffset_unlock_uio(fp, uio, flags);
323 	return (error);
324 }
325 
326 static int
327 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
328     struct thread *td)
329 {
330 	struct shmfd *shmfd;
331 #ifdef MAC
332 	int error;
333 #endif
334 
335 	shmfd = fp->f_data;
336 #ifdef MAC
337 	error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
338 	if (error)
339 		return (error);
340 #endif
341 	return (shm_dotruncate(shmfd, length));
342 }
343 
344 int
345 shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
346     struct thread *td)
347 {
348 
349 	switch (com) {
350 	case FIONBIO:
351 	case FIOASYNC:
352 		/*
353 		 * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work,
354 		 * just like it would on an unlinked regular file
355 		 */
356 		return (0);
357 	default:
358 		return (ENOTTY);
359 	}
360 }
361 
362 static int
363 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
364     struct thread *td)
365 {
366 	struct shmfd *shmfd;
367 #ifdef MAC
368 	int error;
369 #endif
370 
371 	shmfd = fp->f_data;
372 
373 #ifdef MAC
374 	error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
375 	if (error)
376 		return (error);
377 #endif
378 
379 	/*
380 	 * Attempt to return sanish values for fstat() on a memory file
381 	 * descriptor.
382 	 */
383 	bzero(sb, sizeof(*sb));
384 	sb->st_blksize = PAGE_SIZE;
385 	sb->st_size = shmfd->shm_size;
386 	sb->st_blocks = howmany(sb->st_size, sb->st_blksize);
387 	mtx_lock(&shm_timestamp_lock);
388 	sb->st_atim = shmfd->shm_atime;
389 	sb->st_ctim = shmfd->shm_ctime;
390 	sb->st_mtim = shmfd->shm_mtime;
391 	sb->st_birthtim = shmfd->shm_birthtime;
392 	sb->st_mode = S_IFREG | shmfd->shm_mode;		/* XXX */
393 	sb->st_uid = shmfd->shm_uid;
394 	sb->st_gid = shmfd->shm_gid;
395 	mtx_unlock(&shm_timestamp_lock);
396 	sb->st_dev = shm_dev_ino;
397 	sb->st_ino = shmfd->shm_ino;
398 	sb->st_nlink = shmfd->shm_object->ref_count;
399 
400 	return (0);
401 }
402 
403 static int
404 shm_close(struct file *fp, struct thread *td)
405 {
406 	struct shmfd *shmfd;
407 
408 	shmfd = fp->f_data;
409 	fp->f_data = NULL;
410 	shm_drop(shmfd);
411 
412 	return (0);
413 }
414 
415 int
416 shm_dotruncate(struct shmfd *shmfd, off_t length)
417 {
418 	vm_object_t object;
419 	vm_page_t m;
420 	vm_pindex_t idx, nobjsize;
421 	vm_ooffset_t delta;
422 	int base, rv;
423 
424 	KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
425 	object = shmfd->shm_object;
426 	VM_OBJECT_WLOCK(object);
427 	if (length == shmfd->shm_size) {
428 		VM_OBJECT_WUNLOCK(object);
429 		return (0);
430 	}
431 	nobjsize = OFF_TO_IDX(length + PAGE_MASK);
432 
433 	/* Are we shrinking?  If so, trim the end. */
434 	if (length < shmfd->shm_size) {
435 		/*
436 		 * Disallow any requests to shrink the size if this
437 		 * object is mapped into the kernel.
438 		 */
439 		if (shmfd->shm_kmappings > 0) {
440 			VM_OBJECT_WUNLOCK(object);
441 			return (EBUSY);
442 		}
443 
444 		/*
445 		 * Zero the truncated part of the last page.
446 		 */
447 		base = length & PAGE_MASK;
448 		if (base != 0) {
449 			idx = OFF_TO_IDX(length);
450 retry:
451 			m = vm_page_lookup(object, idx);
452 			if (m != NULL) {
453 				if (vm_page_sleep_if_busy(m, "shmtrc"))
454 					goto retry;
455 			} else if (vm_pager_has_page(object, idx, NULL, NULL)) {
456 				m = vm_page_alloc(object, idx,
457 				    VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL);
458 				if (m == NULL)
459 					goto retry;
460 				rv = vm_pager_get_pages(object, &m, 1, NULL,
461 				    NULL);
462 				if (rv == VM_PAGER_OK) {
463 					/*
464 					 * Since the page was not resident,
465 					 * and therefore not recently
466 					 * accessed, immediately enqueue it
467 					 * for asynchronous laundering.  The
468 					 * current operation is not regarded
469 					 * as an access.
470 					 */
471 					vm_page_launder(m);
472 					vm_page_xunbusy(m);
473 				} else {
474 					vm_page_free(m);
475 					VM_OBJECT_WUNLOCK(object);
476 					return (EIO);
477 				}
478 			}
479 			if (m != NULL) {
480 				pmap_zero_page_area(m, base, PAGE_SIZE - base);
481 				KASSERT(m->valid == VM_PAGE_BITS_ALL,
482 				    ("shm_dotruncate: page %p is invalid", m));
483 				vm_page_dirty(m);
484 				vm_pager_page_unswapped(m);
485 			}
486 		}
487 		delta = IDX_TO_OFF(object->size - nobjsize);
488 
489 		/* Toss in memory pages. */
490 		if (nobjsize < object->size)
491 			vm_object_page_remove(object, nobjsize, object->size,
492 			    0);
493 
494 		/* Toss pages from swap. */
495 		if (object->type == OBJT_SWAP)
496 			swap_pager_freespace(object, nobjsize, delta);
497 
498 		/* Free the swap accounted for shm */
499 		swap_release_by_cred(delta, object->cred);
500 		object->charge -= delta;
501 	} else {
502 		/* Try to reserve additional swap space. */
503 		delta = IDX_TO_OFF(nobjsize - object->size);
504 		if (!swap_reserve_by_cred(delta, object->cred)) {
505 			VM_OBJECT_WUNLOCK(object);
506 			return (ENOMEM);
507 		}
508 		object->charge += delta;
509 	}
510 	shmfd->shm_size = length;
511 	mtx_lock(&shm_timestamp_lock);
512 	vfs_timestamp(&shmfd->shm_ctime);
513 	shmfd->shm_mtime = shmfd->shm_ctime;
514 	mtx_unlock(&shm_timestamp_lock);
515 	object->size = nobjsize;
516 	VM_OBJECT_WUNLOCK(object);
517 	return (0);
518 }
519 
520 /*
521  * shmfd object management including creation and reference counting
522  * routines.
523  */
524 struct shmfd *
525 shm_alloc(struct ucred *ucred, mode_t mode)
526 {
527 	struct shmfd *shmfd;
528 
529 	shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
530 	shmfd->shm_size = 0;
531 	shmfd->shm_uid = ucred->cr_uid;
532 	shmfd->shm_gid = ucred->cr_gid;
533 	shmfd->shm_mode = mode;
534 	shmfd->shm_object = vm_pager_allocate(OBJT_SWAP, NULL,
535 	    shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
536 	KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
537 	shmfd->shm_object->pg_color = 0;
538 	VM_OBJECT_WLOCK(shmfd->shm_object);
539 	vm_object_clear_flag(shmfd->shm_object, OBJ_ONEMAPPING);
540 	vm_object_set_flag(shmfd->shm_object, OBJ_COLORED | OBJ_NOSPLIT);
541 	VM_OBJECT_WUNLOCK(shmfd->shm_object);
542 	vfs_timestamp(&shmfd->shm_birthtime);
543 	shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
544 	    shmfd->shm_birthtime;
545 	shmfd->shm_ino = alloc_unr64(&shm_ino_unr);
546 	refcount_init(&shmfd->shm_refs, 1);
547 	mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
548 	rangelock_init(&shmfd->shm_rl);
549 #ifdef MAC
550 	mac_posixshm_init(shmfd);
551 	mac_posixshm_create(ucred, shmfd);
552 #endif
553 
554 	return (shmfd);
555 }
556 
557 struct shmfd *
558 shm_hold(struct shmfd *shmfd)
559 {
560 
561 	refcount_acquire(&shmfd->shm_refs);
562 	return (shmfd);
563 }
564 
565 void
566 shm_drop(struct shmfd *shmfd)
567 {
568 
569 	if (refcount_release(&shmfd->shm_refs)) {
570 #ifdef MAC
571 		mac_posixshm_destroy(shmfd);
572 #endif
573 		rangelock_destroy(&shmfd->shm_rl);
574 		mtx_destroy(&shmfd->shm_mtx);
575 		vm_object_deallocate(shmfd->shm_object);
576 		free(shmfd, M_SHMFD);
577 	}
578 }
579 
580 /*
581  * Determine if the credentials have sufficient permissions for a
582  * specified combination of FREAD and FWRITE.
583  */
584 int
585 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
586 {
587 	accmode_t accmode;
588 	int error;
589 
590 	accmode = 0;
591 	if (flags & FREAD)
592 		accmode |= VREAD;
593 	if (flags & FWRITE)
594 		accmode |= VWRITE;
595 	mtx_lock(&shm_timestamp_lock);
596 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
597 	    accmode, ucred, NULL);
598 	mtx_unlock(&shm_timestamp_lock);
599 	return (error);
600 }
601 
602 /*
603  * Dictionary management.  We maintain an in-kernel dictionary to map
604  * paths to shmfd objects.  We use the FNV hash on the path to store
605  * the mappings in a hash table.
606  */
607 static void
608 shm_init(void *arg)
609 {
610 
611 	mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
612 	sx_init(&shm_dict_lock, "shm dictionary");
613 	shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
614 	new_unrhdr64(&shm_ino_unr, 1);
615 	shm_dev_ino = devfs_alloc_cdp_inode();
616 	KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
617 }
618 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
619 
620 static struct shmfd *
621 shm_lookup(char *path, Fnv32_t fnv)
622 {
623 	struct shm_mapping *map;
624 
625 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
626 		if (map->sm_fnv != fnv)
627 			continue;
628 		if (strcmp(map->sm_path, path) == 0)
629 			return (map->sm_shmfd);
630 	}
631 
632 	return (NULL);
633 }
634 
635 static void
636 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
637 {
638 	struct shm_mapping *map;
639 
640 	map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
641 	map->sm_path = path;
642 	map->sm_fnv = fnv;
643 	map->sm_shmfd = shm_hold(shmfd);
644 	shmfd->shm_path = path;
645 	LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
646 }
647 
648 static int
649 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
650 {
651 	struct shm_mapping *map;
652 	int error;
653 
654 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
655 		if (map->sm_fnv != fnv)
656 			continue;
657 		if (strcmp(map->sm_path, path) == 0) {
658 #ifdef MAC
659 			error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
660 			if (error)
661 				return (error);
662 #endif
663 			error = shm_access(map->sm_shmfd, ucred,
664 			    FREAD | FWRITE);
665 			if (error)
666 				return (error);
667 			map->sm_shmfd->shm_path = NULL;
668 			LIST_REMOVE(map, sm_link);
669 			shm_drop(map->sm_shmfd);
670 			free(map->sm_path, M_SHMFD);
671 			free(map, M_SHMFD);
672 			return (0);
673 		}
674 	}
675 
676 	return (ENOENT);
677 }
678 
679 int
680 kern_shm_open(struct thread *td, const char *userpath, int flags, mode_t mode,
681     struct filecaps *fcaps)
682 {
683 	struct filedesc *fdp;
684 	struct shmfd *shmfd;
685 	struct file *fp;
686 	char *path;
687 	const char *pr_path;
688 	size_t pr_pathlen;
689 	Fnv32_t fnv;
690 	mode_t cmode;
691 	int fd, error;
692 
693 #ifdef CAPABILITY_MODE
694 	/*
695 	 * shm_open(2) is only allowed for anonymous objects.
696 	 */
697 	if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON))
698 		return (ECAPMODE);
699 #endif
700 
701 	AUDIT_ARG_FFLAGS(flags);
702 	AUDIT_ARG_MODE(mode);
703 
704 	if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
705 		return (EINVAL);
706 
707 	if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
708 		return (EINVAL);
709 
710 	fdp = td->td_proc->p_fd;
711 	cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS;
712 
713 	/*
714 	 * shm_open(2) created shm should always have O_CLOEXEC set, as mandated
715 	 * by POSIX.  We allow it to be unset here so that an in-kernel
716 	 * interface may be written as a thin layer around shm, optionally not
717 	 * setting CLOEXEC.  For shm_open(2), O_CLOEXEC is set unconditionally
718 	 * in sys_shm_open() to keep this implementation compliant.
719 	 */
720 	error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps);
721 	if (error)
722 		return (error);
723 
724 	/* A SHM_ANON path pointer creates an anonymous object. */
725 	if (userpath == SHM_ANON) {
726 		/* A read-only anonymous object is pointless. */
727 		if ((flags & O_ACCMODE) == O_RDONLY) {
728 			fdclose(td, fp, fd);
729 			fdrop(fp, td);
730 			return (EINVAL);
731 		}
732 		shmfd = shm_alloc(td->td_ucred, cmode);
733 	} else {
734 		path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
735 		pr_path = td->td_ucred->cr_prison->pr_path;
736 
737 		/* Construct a full pathname for jailed callers. */
738 		pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
739 		    : strlcpy(path, pr_path, MAXPATHLEN);
740 		error = copyinstr(userpath, path + pr_pathlen,
741 		    MAXPATHLEN - pr_pathlen, NULL);
742 #ifdef KTRACE
743 		if (error == 0 && KTRPOINT(curthread, KTR_NAMEI))
744 			ktrnamei(path);
745 #endif
746 		/* Require paths to start with a '/' character. */
747 		if (error == 0 && path[pr_pathlen] != '/')
748 			error = EINVAL;
749 		if (error) {
750 			fdclose(td, fp, fd);
751 			fdrop(fp, td);
752 			free(path, M_SHMFD);
753 			return (error);
754 		}
755 
756 		AUDIT_ARG_UPATH1_CANON(path);
757 		fnv = fnv_32_str(path, FNV1_32_INIT);
758 		sx_xlock(&shm_dict_lock);
759 		shmfd = shm_lookup(path, fnv);
760 		if (shmfd == NULL) {
761 			/* Object does not yet exist, create it if requested. */
762 			if (flags & O_CREAT) {
763 #ifdef MAC
764 				error = mac_posixshm_check_create(td->td_ucred,
765 				    path);
766 				if (error == 0) {
767 #endif
768 					shmfd = shm_alloc(td->td_ucred, cmode);
769 					shm_insert(path, fnv, shmfd);
770 #ifdef MAC
771 				}
772 #endif
773 			} else {
774 				free(path, M_SHMFD);
775 				error = ENOENT;
776 			}
777 		} else {
778 			/*
779 			 * Object already exists, obtain a new
780 			 * reference if requested and permitted.
781 			 */
782 			free(path, M_SHMFD);
783 			if ((flags & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
784 				error = EEXIST;
785 			else {
786 #ifdef MAC
787 				error = mac_posixshm_check_open(td->td_ucred,
788 				    shmfd, FFLAGS(flags & O_ACCMODE));
789 				if (error == 0)
790 #endif
791 				error = shm_access(shmfd, td->td_ucred,
792 				    FFLAGS(flags & O_ACCMODE));
793 			}
794 
795 			/*
796 			 * Truncate the file back to zero length if
797 			 * O_TRUNC was specified and the object was
798 			 * opened with read/write.
799 			 */
800 			if (error == 0 &&
801 			    (flags & (O_ACCMODE | O_TRUNC)) ==
802 			    (O_RDWR | O_TRUNC)) {
803 #ifdef MAC
804 				error = mac_posixshm_check_truncate(
805 					td->td_ucred, fp->f_cred, shmfd);
806 				if (error == 0)
807 #endif
808 					shm_dotruncate(shmfd, 0);
809 			}
810 			if (error == 0)
811 				shm_hold(shmfd);
812 		}
813 		sx_xunlock(&shm_dict_lock);
814 
815 		if (error) {
816 			fdclose(td, fp, fd);
817 			fdrop(fp, td);
818 			return (error);
819 		}
820 	}
821 
822 	finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
823 
824 	td->td_retval[0] = fd;
825 	fdrop(fp, td);
826 
827 	return (0);
828 }
829 
830 /* System calls. */
831 int
832 sys_shm_open(struct thread *td, struct shm_open_args *uap)
833 {
834 
835 	return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC, uap->mode,
836 	    NULL));
837 }
838 
839 int
840 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
841 {
842 	char *path;
843 	const char *pr_path;
844 	size_t pr_pathlen;
845 	Fnv32_t fnv;
846 	int error;
847 
848 	path = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
849 	pr_path = td->td_ucred->cr_prison->pr_path;
850 	pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
851 	    : strlcpy(path, pr_path, MAXPATHLEN);
852 	error = copyinstr(uap->path, path + pr_pathlen, MAXPATHLEN - pr_pathlen,
853 	    NULL);
854 	if (error) {
855 		free(path, M_TEMP);
856 		return (error);
857 	}
858 #ifdef KTRACE
859 	if (KTRPOINT(curthread, KTR_NAMEI))
860 		ktrnamei(path);
861 #endif
862 	AUDIT_ARG_UPATH1_CANON(path);
863 	fnv = fnv_32_str(path, FNV1_32_INIT);
864 	sx_xlock(&shm_dict_lock);
865 	error = shm_remove(path, fnv, td->td_ucred);
866 	sx_xunlock(&shm_dict_lock);
867 	free(path, M_TEMP);
868 
869 	return (error);
870 }
871 
872 int
873 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
874     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
875     vm_ooffset_t foff, struct thread *td)
876 {
877 	struct shmfd *shmfd;
878 	vm_prot_t maxprot;
879 	int error;
880 	bool writecnt;
881 
882 	shmfd = fp->f_data;
883 	maxprot = VM_PROT_NONE;
884 
885 	/* FREAD should always be set. */
886 	if ((fp->f_flag & FREAD) != 0)
887 		maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
888 	if ((fp->f_flag & FWRITE) != 0)
889 		maxprot |= VM_PROT_WRITE;
890 
891 	writecnt = (flags & MAP_SHARED) != 0 && (prot & VM_PROT_WRITE) != 0;
892 
893 	/* Don't permit shared writable mappings on read-only descriptors. */
894 	if (writecnt && (maxprot & VM_PROT_WRITE) == 0)
895 		return (EACCES);
896 	maxprot &= cap_maxprot;
897 
898 	/* See comment in vn_mmap(). */
899 	if (
900 #ifdef _LP64
901 	    objsize > OFF_MAX ||
902 #endif
903 	    foff < 0 || foff > OFF_MAX - objsize)
904 		return (EINVAL);
905 
906 #ifdef MAC
907 	error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
908 	if (error != 0)
909 		return (error);
910 #endif
911 
912 	mtx_lock(&shm_timestamp_lock);
913 	vfs_timestamp(&shmfd->shm_atime);
914 	mtx_unlock(&shm_timestamp_lock);
915 	vm_object_reference(shmfd->shm_object);
916 
917 	if (writecnt)
918 		vm_pager_update_writecount(shmfd->shm_object, 0, objsize);
919 	error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
920 	    shmfd->shm_object, foff, writecnt, td);
921 	if (error != 0) {
922 		if (writecnt)
923 			vm_pager_release_writecount(shmfd->shm_object, 0,
924 			    objsize);
925 		vm_object_deallocate(shmfd->shm_object);
926 	}
927 	return (error);
928 }
929 
930 static int
931 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
932     struct thread *td)
933 {
934 	struct shmfd *shmfd;
935 	int error;
936 
937 	error = 0;
938 	shmfd = fp->f_data;
939 	mtx_lock(&shm_timestamp_lock);
940 	/*
941 	 * SUSv4 says that x bits of permission need not be affected.
942 	 * Be consistent with our shm_open there.
943 	 */
944 #ifdef MAC
945 	error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
946 	if (error != 0)
947 		goto out;
948 #endif
949 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid,
950 	    shmfd->shm_gid, VADMIN, active_cred, NULL);
951 	if (error != 0)
952 		goto out;
953 	shmfd->shm_mode = mode & ACCESSPERMS;
954 out:
955 	mtx_unlock(&shm_timestamp_lock);
956 	return (error);
957 }
958 
959 static int
960 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
961     struct thread *td)
962 {
963 	struct shmfd *shmfd;
964 	int error;
965 
966 	error = 0;
967 	shmfd = fp->f_data;
968 	mtx_lock(&shm_timestamp_lock);
969 #ifdef MAC
970 	error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
971 	if (error != 0)
972 		goto out;
973 #endif
974 	if (uid == (uid_t)-1)
975 		uid = shmfd->shm_uid;
976 	if (gid == (gid_t)-1)
977                  gid = shmfd->shm_gid;
978 	if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
979 	    (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
980 	    (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN)))
981 		goto out;
982 	shmfd->shm_uid = uid;
983 	shmfd->shm_gid = gid;
984 out:
985 	mtx_unlock(&shm_timestamp_lock);
986 	return (error);
987 }
988 
989 /*
990  * Helper routines to allow the backing object of a shared memory file
991  * descriptor to be mapped in the kernel.
992  */
993 int
994 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
995 {
996 	struct shmfd *shmfd;
997 	vm_offset_t kva, ofs;
998 	vm_object_t obj;
999 	int rv;
1000 
1001 	if (fp->f_type != DTYPE_SHM)
1002 		return (EINVAL);
1003 	shmfd = fp->f_data;
1004 	obj = shmfd->shm_object;
1005 	VM_OBJECT_WLOCK(obj);
1006 	/*
1007 	 * XXXRW: This validation is probably insufficient, and subject to
1008 	 * sign errors.  It should be fixed.
1009 	 */
1010 	if (offset >= shmfd->shm_size ||
1011 	    offset + size > round_page(shmfd->shm_size)) {
1012 		VM_OBJECT_WUNLOCK(obj);
1013 		return (EINVAL);
1014 	}
1015 
1016 	shmfd->shm_kmappings++;
1017 	vm_object_reference_locked(obj);
1018 	VM_OBJECT_WUNLOCK(obj);
1019 
1020 	/* Map the object into the kernel_map and wire it. */
1021 	kva = vm_map_min(kernel_map);
1022 	ofs = offset & PAGE_MASK;
1023 	offset = trunc_page(offset);
1024 	size = round_page(size + ofs);
1025 	rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1026 	    VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1027 	    VM_PROT_READ | VM_PROT_WRITE, 0);
1028 	if (rv == KERN_SUCCESS) {
1029 		rv = vm_map_wire(kernel_map, kva, kva + size,
1030 		    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1031 		if (rv == KERN_SUCCESS) {
1032 			*memp = (void *)(kva + ofs);
1033 			return (0);
1034 		}
1035 		vm_map_remove(kernel_map, kva, kva + size);
1036 	} else
1037 		vm_object_deallocate(obj);
1038 
1039 	/* On failure, drop our mapping reference. */
1040 	VM_OBJECT_WLOCK(obj);
1041 	shmfd->shm_kmappings--;
1042 	VM_OBJECT_WUNLOCK(obj);
1043 
1044 	return (vm_mmap_to_errno(rv));
1045 }
1046 
1047 /*
1048  * We require the caller to unmap the entire entry.  This allows us to
1049  * safely decrement shm_kmappings when a mapping is removed.
1050  */
1051 int
1052 shm_unmap(struct file *fp, void *mem, size_t size)
1053 {
1054 	struct shmfd *shmfd;
1055 	vm_map_entry_t entry;
1056 	vm_offset_t kva, ofs;
1057 	vm_object_t obj;
1058 	vm_pindex_t pindex;
1059 	vm_prot_t prot;
1060 	boolean_t wired;
1061 	vm_map_t map;
1062 	int rv;
1063 
1064 	if (fp->f_type != DTYPE_SHM)
1065 		return (EINVAL);
1066 	shmfd = fp->f_data;
1067 	kva = (vm_offset_t)mem;
1068 	ofs = kva & PAGE_MASK;
1069 	kva = trunc_page(kva);
1070 	size = round_page(size + ofs);
1071 	map = kernel_map;
1072 	rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1073 	    &obj, &pindex, &prot, &wired);
1074 	if (rv != KERN_SUCCESS)
1075 		return (EINVAL);
1076 	if (entry->start != kva || entry->end != kva + size) {
1077 		vm_map_lookup_done(map, entry);
1078 		return (EINVAL);
1079 	}
1080 	vm_map_lookup_done(map, entry);
1081 	if (obj != shmfd->shm_object)
1082 		return (EINVAL);
1083 	vm_map_remove(map, kva, kva + size);
1084 	VM_OBJECT_WLOCK(obj);
1085 	KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1086 	shmfd->shm_kmappings--;
1087 	VM_OBJECT_WUNLOCK(obj);
1088 	return (0);
1089 }
1090 
1091 static int
1092 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list)
1093 {
1094 	const char *path, *pr_path;
1095 	size_t pr_pathlen;
1096 	bool visible;
1097 
1098 	sx_assert(&shm_dict_lock, SA_LOCKED);
1099 	kif->kf_type = KF_TYPE_SHM;
1100 	kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;
1101 	kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1102 	if (shmfd->shm_path != NULL) {
1103 		if (shmfd->shm_path != NULL) {
1104 			path = shmfd->shm_path;
1105 			pr_path = curthread->td_ucred->cr_prison->pr_path;
1106 			if (strcmp(pr_path, "/") != 0) {
1107 				/* Return the jail-rooted pathname. */
1108 				pr_pathlen = strlen(pr_path);
1109 				visible = strncmp(path, pr_path, pr_pathlen)
1110 				    == 0 && path[pr_pathlen] == '/';
1111 				if (list && !visible)
1112 					return (EPERM);
1113 				if (visible)
1114 					path += pr_pathlen;
1115 			}
1116 			strlcpy(kif->kf_path, path, sizeof(kif->kf_path));
1117 		}
1118 	}
1119 	return (0);
1120 }
1121 
1122 static int
1123 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif,
1124     struct filedesc *fdp __unused)
1125 {
1126 	int res;
1127 
1128 	sx_slock(&shm_dict_lock);
1129 	res = shm_fill_kinfo_locked(fp->f_data, kif, false);
1130 	sx_sunlock(&shm_dict_lock);
1131 	return (res);
1132 }
1133 
1134 static int
1135 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS)
1136 {
1137 	struct shm_mapping *shmm;
1138 	struct sbuf sb;
1139 	struct kinfo_file kif;
1140 	u_long i;
1141 	ssize_t curlen;
1142 	int error, error2;
1143 
1144 	sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req);
1145 	sbuf_clear_flags(&sb, SBUF_INCLUDENUL);
1146 	curlen = 0;
1147 	error = 0;
1148 	sx_slock(&shm_dict_lock);
1149 	for (i = 0; i < shm_hash + 1; i++) {
1150 		LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) {
1151 			error = shm_fill_kinfo_locked(shmm->sm_shmfd,
1152 			    &kif, true);
1153 			if (error == EPERM)
1154 				continue;
1155 			if (error != 0)
1156 				break;
1157 			pack_kinfo(&kif);
1158 			if (req->oldptr != NULL &&
1159 			    kif.kf_structsize + curlen > req->oldlen)
1160 				break;
1161 			error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ?
1162 			    0 : ENOMEM;
1163 			if (error != 0)
1164 				break;
1165 			curlen += kif.kf_structsize;
1166 		}
1167 	}
1168 	sx_sunlock(&shm_dict_lock);
1169 	error2 = sbuf_finish(&sb);
1170 	sbuf_delete(&sb);
1171 	return (error != 0 ? error : error2);
1172 }
1173 
1174 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list,
1175     CTLFLAG_RD | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE,
1176     NULL, 0, sysctl_posix_shm_list, "",
1177     "POSIX SHM list");
1178