xref: /freebsd/sys/kern/uipc_shm.c (revision 389e4940069316fe667ffa263fa7d6390d0a960f)
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  * TODO:
40  *
41  * (1) Need to export data to a userland tool via a sysctl.  Should ipcs(1)
42  *     and ipcrm(1) be expanded or should new tools to manage both POSIX
43  *     kernel semaphores and POSIX shared memory be written?
44  *
45  * (2) Add support for this file type to fstat(1).
46  *
47  * (3) Resource limits?  Does this need its own resource limits or are the
48  *     existing limits in mmap(2) sufficient?
49  */
50 
51 #include <sys/cdefs.h>
52 __FBSDID("$FreeBSD$");
53 
54 #include "opt_capsicum.h"
55 #include "opt_ktrace.h"
56 
57 #include <sys/param.h>
58 #include <sys/capsicum.h>
59 #include <sys/conf.h>
60 #include <sys/fcntl.h>
61 #include <sys/file.h>
62 #include <sys/filedesc.h>
63 #include <sys/fnv_hash.h>
64 #include <sys/kernel.h>
65 #include <sys/uio.h>
66 #include <sys/signal.h>
67 #include <sys/jail.h>
68 #include <sys/ktrace.h>
69 #include <sys/lock.h>
70 #include <sys/malloc.h>
71 #include <sys/mman.h>
72 #include <sys/mutex.h>
73 #include <sys/priv.h>
74 #include <sys/proc.h>
75 #include <sys/refcount.h>
76 #include <sys/resourcevar.h>
77 #include <sys/rwlock.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/vnode.h>
86 #include <sys/unistd.h>
87 #include <sys/user.h>
88 
89 #include <security/audit/audit.h>
90 #include <security/mac/mac_framework.h>
91 
92 #include <vm/vm.h>
93 #include <vm/vm_param.h>
94 #include <vm/pmap.h>
95 #include <vm/vm_extern.h>
96 #include <vm/vm_map.h>
97 #include <vm/vm_kern.h>
98 #include <vm/vm_object.h>
99 #include <vm/vm_page.h>
100 #include <vm/vm_pageout.h>
101 #include <vm/vm_pager.h>
102 #include <vm/swap_pager.h>
103 
104 struct shm_mapping {
105 	char		*sm_path;
106 	Fnv32_t		sm_fnv;
107 	struct shmfd	*sm_shmfd;
108 	LIST_ENTRY(shm_mapping) sm_link;
109 };
110 
111 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor");
112 static LIST_HEAD(, shm_mapping) *shm_dictionary;
113 static struct sx shm_dict_lock;
114 static struct mtx shm_timestamp_lock;
115 static u_long shm_hash;
116 static struct unrhdr *shm_ino_unr;
117 static dev_t shm_dev_ino;
118 
119 #define	SHM_HASH(fnv)	(&shm_dictionary[(fnv) & shm_hash])
120 
121 static void	shm_init(void *arg);
122 static void	shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd);
123 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv);
124 static int	shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred);
125 
126 static fo_rdwr_t	shm_read;
127 static fo_rdwr_t	shm_write;
128 static fo_truncate_t	shm_truncate;
129 static fo_stat_t	shm_stat;
130 static fo_close_t	shm_close;
131 static fo_chmod_t	shm_chmod;
132 static fo_chown_t	shm_chown;
133 static fo_seek_t	shm_seek;
134 static fo_fill_kinfo_t	shm_fill_kinfo;
135 static fo_mmap_t	shm_mmap;
136 
137 /* File descriptor operations. */
138 struct fileops shm_ops = {
139 	.fo_read = shm_read,
140 	.fo_write = shm_write,
141 	.fo_truncate = shm_truncate,
142 	.fo_ioctl = invfo_ioctl,
143 	.fo_poll = invfo_poll,
144 	.fo_kqfilter = invfo_kqfilter,
145 	.fo_stat = shm_stat,
146 	.fo_close = shm_close,
147 	.fo_chmod = shm_chmod,
148 	.fo_chown = shm_chown,
149 	.fo_sendfile = vn_sendfile,
150 	.fo_seek = shm_seek,
151 	.fo_fill_kinfo = shm_fill_kinfo,
152 	.fo_mmap = shm_mmap,
153 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
154 };
155 
156 FEATURE(posix_shm, "POSIX shared memory");
157 
158 static int
159 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio)
160 {
161 	vm_page_t m;
162 	vm_pindex_t idx;
163 	size_t tlen;
164 	int error, offset, rv;
165 
166 	idx = OFF_TO_IDX(uio->uio_offset);
167 	offset = uio->uio_offset & PAGE_MASK;
168 	tlen = MIN(PAGE_SIZE - offset, len);
169 
170 	VM_OBJECT_WLOCK(obj);
171 
172 	/*
173 	 * Read I/O without either a corresponding resident page or swap
174 	 * page: use zero_region.  This is intended to avoid instantiating
175 	 * pages on read from a sparse region.
176 	 */
177 	if (uio->uio_rw == UIO_READ && vm_page_lookup(obj, idx) == NULL &&
178 	    !vm_pager_has_page(obj, idx, NULL, NULL)) {
179 		VM_OBJECT_WUNLOCK(obj);
180 		return (uiomove(__DECONST(void *, zero_region), tlen, uio));
181 	}
182 
183 	/*
184 	 * Parallel reads of the page content from disk are prevented
185 	 * by exclusive busy.
186 	 *
187 	 * Although the tmpfs vnode lock is held here, it is
188 	 * nonetheless safe to sleep waiting for a free page.  The
189 	 * pageout daemon does not need to acquire the tmpfs vnode
190 	 * lock to page out tobj's pages because tobj is a OBJT_SWAP
191 	 * type object.
192 	 */
193 	m = vm_page_grab(obj, idx, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
194 	if (m->valid != VM_PAGE_BITS_ALL) {
195 		vm_page_xbusy(m);
196 		if (vm_pager_has_page(obj, idx, NULL, NULL)) {
197 			rv = vm_pager_get_pages(obj, &m, 1, NULL, NULL);
198 			if (rv != VM_PAGER_OK) {
199 				printf(
200 	    "uiomove_object: vm_obj %p idx %jd valid %x pager error %d\n",
201 				    obj, idx, m->valid, rv);
202 				vm_page_lock(m);
203 				vm_page_free(m);
204 				vm_page_unlock(m);
205 				VM_OBJECT_WUNLOCK(obj);
206 				return (EIO);
207 			}
208 		} else
209 			vm_page_zero_invalid(m, TRUE);
210 		vm_page_xunbusy(m);
211 	}
212 	vm_page_lock(m);
213 	vm_page_hold(m);
214 	if (vm_page_active(m))
215 		vm_page_reference(m);
216 	else
217 		vm_page_activate(m);
218 	vm_page_unlock(m);
219 	VM_OBJECT_WUNLOCK(obj);
220 	error = uiomove_fromphys(&m, offset, tlen, uio);
221 	if (uio->uio_rw == UIO_WRITE && error == 0) {
222 		VM_OBJECT_WLOCK(obj);
223 		vm_page_dirty(m);
224 		vm_pager_page_unswapped(m);
225 		VM_OBJECT_WUNLOCK(obj);
226 	}
227 	vm_page_lock(m);
228 	vm_page_unhold(m);
229 	vm_page_unlock(m);
230 
231 	return (error);
232 }
233 
234 int
235 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
236 {
237 	ssize_t resid;
238 	size_t len;
239 	int error;
240 
241 	error = 0;
242 	while ((resid = uio->uio_resid) > 0) {
243 		if (obj_size <= uio->uio_offset)
244 			break;
245 		len = MIN(obj_size - uio->uio_offset, resid);
246 		if (len == 0)
247 			break;
248 		error = uiomove_object_page(obj, len, uio);
249 		if (error != 0 || resid == uio->uio_resid)
250 			break;
251 	}
252 	return (error);
253 }
254 
255 static int
256 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
257 {
258 	struct shmfd *shmfd;
259 	off_t foffset;
260 	int error;
261 
262 	shmfd = fp->f_data;
263 	foffset = foffset_lock(fp, 0);
264 	error = 0;
265 	switch (whence) {
266 	case L_INCR:
267 		if (foffset < 0 ||
268 		    (offset > 0 && foffset > OFF_MAX - offset)) {
269 			error = EOVERFLOW;
270 			break;
271 		}
272 		offset += foffset;
273 		break;
274 	case L_XTND:
275 		if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
276 			error = EOVERFLOW;
277 			break;
278 		}
279 		offset += shmfd->shm_size;
280 		break;
281 	case L_SET:
282 		break;
283 	default:
284 		error = EINVAL;
285 	}
286 	if (error == 0) {
287 		if (offset < 0 || offset > shmfd->shm_size)
288 			error = EINVAL;
289 		else
290 			td->td_uretoff.tdu_off = offset;
291 	}
292 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
293 	return (error);
294 }
295 
296 static int
297 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
298     int flags, struct thread *td)
299 {
300 	struct shmfd *shmfd;
301 	void *rl_cookie;
302 	int error;
303 
304 	shmfd = fp->f_data;
305 #ifdef MAC
306 	error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
307 	if (error)
308 		return (error);
309 #endif
310 	foffset_lock_uio(fp, uio, flags);
311 	rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset,
312 	    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
313 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
314 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
315 	foffset_unlock_uio(fp, uio, flags);
316 	return (error);
317 }
318 
319 static int
320 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
321     int flags, struct thread *td)
322 {
323 	struct shmfd *shmfd;
324 	void *rl_cookie;
325 	int error;
326 
327 	shmfd = fp->f_data;
328 #ifdef MAC
329 	error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
330 	if (error)
331 		return (error);
332 #endif
333 	foffset_lock_uio(fp, uio, flags);
334 	if ((flags & FOF_OFFSET) == 0) {
335 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
336 		    &shmfd->shm_mtx);
337 	} else {
338 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset,
339 		    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
340 	}
341 
342 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
343 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
344 	foffset_unlock_uio(fp, uio, flags);
345 	return (error);
346 }
347 
348 static int
349 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
350     struct thread *td)
351 {
352 	struct shmfd *shmfd;
353 #ifdef MAC
354 	int error;
355 #endif
356 
357 	shmfd = fp->f_data;
358 #ifdef MAC
359 	error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
360 	if (error)
361 		return (error);
362 #endif
363 	return (shm_dotruncate(shmfd, length));
364 }
365 
366 static int
367 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
368     struct thread *td)
369 {
370 	struct shmfd *shmfd;
371 #ifdef MAC
372 	int error;
373 #endif
374 
375 	shmfd = fp->f_data;
376 
377 #ifdef MAC
378 	error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
379 	if (error)
380 		return (error);
381 #endif
382 
383 	/*
384 	 * Attempt to return sanish values for fstat() on a memory file
385 	 * descriptor.
386 	 */
387 	bzero(sb, sizeof(*sb));
388 	sb->st_blksize = PAGE_SIZE;
389 	sb->st_size = shmfd->shm_size;
390 	sb->st_blocks = howmany(sb->st_size, sb->st_blksize);
391 	mtx_lock(&shm_timestamp_lock);
392 	sb->st_atim = shmfd->shm_atime;
393 	sb->st_ctim = shmfd->shm_ctime;
394 	sb->st_mtim = shmfd->shm_mtime;
395 	sb->st_birthtim = shmfd->shm_birthtime;
396 	sb->st_mode = S_IFREG | shmfd->shm_mode;		/* XXX */
397 	sb->st_uid = shmfd->shm_uid;
398 	sb->st_gid = shmfd->shm_gid;
399 	mtx_unlock(&shm_timestamp_lock);
400 	sb->st_dev = shm_dev_ino;
401 	sb->st_ino = shmfd->shm_ino;
402 
403 	return (0);
404 }
405 
406 static int
407 shm_close(struct file *fp, struct thread *td)
408 {
409 	struct shmfd *shmfd;
410 
411 	shmfd = fp->f_data;
412 	fp->f_data = NULL;
413 	shm_drop(shmfd);
414 
415 	return (0);
416 }
417 
418 int
419 shm_dotruncate(struct shmfd *shmfd, off_t length)
420 {
421 	vm_object_t object;
422 	vm_page_t m;
423 	vm_pindex_t idx, nobjsize;
424 	vm_ooffset_t delta;
425 	int base, rv;
426 
427 	KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
428 	object = shmfd->shm_object;
429 	VM_OBJECT_WLOCK(object);
430 	if (length == shmfd->shm_size) {
431 		VM_OBJECT_WUNLOCK(object);
432 		return (0);
433 	}
434 	nobjsize = OFF_TO_IDX(length + PAGE_MASK);
435 
436 	/* Are we shrinking?  If so, trim the end. */
437 	if (length < shmfd->shm_size) {
438 		/*
439 		 * Disallow any requests to shrink the size if this
440 		 * object is mapped into the kernel.
441 		 */
442 		if (shmfd->shm_kmappings > 0) {
443 			VM_OBJECT_WUNLOCK(object);
444 			return (EBUSY);
445 		}
446 
447 		/*
448 		 * Zero the truncated part of the last page.
449 		 */
450 		base = length & PAGE_MASK;
451 		if (base != 0) {
452 			idx = OFF_TO_IDX(length);
453 retry:
454 			m = vm_page_lookup(object, idx);
455 			if (m != NULL) {
456 				if (vm_page_sleep_if_busy(m, "shmtrc"))
457 					goto retry;
458 			} else if (vm_pager_has_page(object, idx, NULL, NULL)) {
459 				m = vm_page_alloc(object, idx,
460 				    VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL);
461 				if (m == NULL)
462 					goto retry;
463 				rv = vm_pager_get_pages(object, &m, 1, NULL,
464 				    NULL);
465 				vm_page_lock(m);
466 				if (rv == VM_PAGER_OK) {
467 					/*
468 					 * Since the page was not resident,
469 					 * and therefore not recently
470 					 * accessed, immediately enqueue it
471 					 * for asynchronous laundering.  The
472 					 * current operation is not regarded
473 					 * as an access.
474 					 */
475 					vm_page_launder(m);
476 					vm_page_unlock(m);
477 					vm_page_xunbusy(m);
478 				} else {
479 					vm_page_free(m);
480 					vm_page_unlock(m);
481 					VM_OBJECT_WUNLOCK(object);
482 					return (EIO);
483 				}
484 			}
485 			if (m != NULL) {
486 				pmap_zero_page_area(m, base, PAGE_SIZE - base);
487 				KASSERT(m->valid == VM_PAGE_BITS_ALL,
488 				    ("shm_dotruncate: page %p is invalid", m));
489 				vm_page_dirty(m);
490 				vm_pager_page_unswapped(m);
491 			}
492 		}
493 		delta = IDX_TO_OFF(object->size - nobjsize);
494 
495 		/* Toss in memory pages. */
496 		if (nobjsize < object->size)
497 			vm_object_page_remove(object, nobjsize, object->size,
498 			    0);
499 
500 		/* Toss pages from swap. */
501 		if (object->type == OBJT_SWAP)
502 			swap_pager_freespace(object, nobjsize, delta);
503 
504 		/* Free the swap accounted for shm */
505 		swap_release_by_cred(delta, object->cred);
506 		object->charge -= delta;
507 	} else {
508 		/* Try to reserve additional swap space. */
509 		delta = IDX_TO_OFF(nobjsize - object->size);
510 		if (!swap_reserve_by_cred(delta, object->cred)) {
511 			VM_OBJECT_WUNLOCK(object);
512 			return (ENOMEM);
513 		}
514 		object->charge += delta;
515 	}
516 	shmfd->shm_size = length;
517 	mtx_lock(&shm_timestamp_lock);
518 	vfs_timestamp(&shmfd->shm_ctime);
519 	shmfd->shm_mtime = shmfd->shm_ctime;
520 	mtx_unlock(&shm_timestamp_lock);
521 	object->size = nobjsize;
522 	VM_OBJECT_WUNLOCK(object);
523 	return (0);
524 }
525 
526 /*
527  * shmfd object management including creation and reference counting
528  * routines.
529  */
530 struct shmfd *
531 shm_alloc(struct ucred *ucred, mode_t mode)
532 {
533 	struct shmfd *shmfd;
534 	int ino;
535 
536 	shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
537 	shmfd->shm_size = 0;
538 	shmfd->shm_uid = ucred->cr_uid;
539 	shmfd->shm_gid = ucred->cr_gid;
540 	shmfd->shm_mode = mode;
541 	shmfd->shm_object = vm_pager_allocate(OBJT_DEFAULT, NULL,
542 	    shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
543 	KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
544 	shmfd->shm_object->pg_color = 0;
545 	VM_OBJECT_WLOCK(shmfd->shm_object);
546 	vm_object_clear_flag(shmfd->shm_object, OBJ_ONEMAPPING);
547 	vm_object_set_flag(shmfd->shm_object, OBJ_COLORED | OBJ_NOSPLIT);
548 	VM_OBJECT_WUNLOCK(shmfd->shm_object);
549 	vfs_timestamp(&shmfd->shm_birthtime);
550 	shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
551 	    shmfd->shm_birthtime;
552 	ino = alloc_unr(shm_ino_unr);
553 	if (ino == -1)
554 		shmfd->shm_ino = 0;
555 	else
556 		shmfd->shm_ino = ino;
557 	refcount_init(&shmfd->shm_refs, 1);
558 	mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
559 	rangelock_init(&shmfd->shm_rl);
560 #ifdef MAC
561 	mac_posixshm_init(shmfd);
562 	mac_posixshm_create(ucred, shmfd);
563 #endif
564 
565 	return (shmfd);
566 }
567 
568 struct shmfd *
569 shm_hold(struct shmfd *shmfd)
570 {
571 
572 	refcount_acquire(&shmfd->shm_refs);
573 	return (shmfd);
574 }
575 
576 void
577 shm_drop(struct shmfd *shmfd)
578 {
579 
580 	if (refcount_release(&shmfd->shm_refs)) {
581 #ifdef MAC
582 		mac_posixshm_destroy(shmfd);
583 #endif
584 		rangelock_destroy(&shmfd->shm_rl);
585 		mtx_destroy(&shmfd->shm_mtx);
586 		vm_object_deallocate(shmfd->shm_object);
587 		if (shmfd->shm_ino != 0)
588 			free_unr(shm_ino_unr, shmfd->shm_ino);
589 		free(shmfd, M_SHMFD);
590 	}
591 }
592 
593 /*
594  * Determine if the credentials have sufficient permissions for a
595  * specified combination of FREAD and FWRITE.
596  */
597 int
598 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
599 {
600 	accmode_t accmode;
601 	int error;
602 
603 	accmode = 0;
604 	if (flags & FREAD)
605 		accmode |= VREAD;
606 	if (flags & FWRITE)
607 		accmode |= VWRITE;
608 	mtx_lock(&shm_timestamp_lock);
609 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
610 	    accmode, ucred, NULL);
611 	mtx_unlock(&shm_timestamp_lock);
612 	return (error);
613 }
614 
615 /*
616  * Dictionary management.  We maintain an in-kernel dictionary to map
617  * paths to shmfd objects.  We use the FNV hash on the path to store
618  * the mappings in a hash table.
619  */
620 static void
621 shm_init(void *arg)
622 {
623 
624 	mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
625 	sx_init(&shm_dict_lock, "shm dictionary");
626 	shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
627 	shm_ino_unr = new_unrhdr(1, INT32_MAX, NULL);
628 	KASSERT(shm_ino_unr != NULL, ("shm fake inodes not initialized"));
629 	shm_dev_ino = devfs_alloc_cdp_inode();
630 	KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
631 }
632 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
633 
634 static struct shmfd *
635 shm_lookup(char *path, Fnv32_t fnv)
636 {
637 	struct shm_mapping *map;
638 
639 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
640 		if (map->sm_fnv != fnv)
641 			continue;
642 		if (strcmp(map->sm_path, path) == 0)
643 			return (map->sm_shmfd);
644 	}
645 
646 	return (NULL);
647 }
648 
649 static void
650 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
651 {
652 	struct shm_mapping *map;
653 
654 	map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
655 	map->sm_path = path;
656 	map->sm_fnv = fnv;
657 	map->sm_shmfd = shm_hold(shmfd);
658 	shmfd->shm_path = path;
659 	LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
660 }
661 
662 static int
663 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
664 {
665 	struct shm_mapping *map;
666 	int error;
667 
668 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
669 		if (map->sm_fnv != fnv)
670 			continue;
671 		if (strcmp(map->sm_path, path) == 0) {
672 #ifdef MAC
673 			error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
674 			if (error)
675 				return (error);
676 #endif
677 			error = shm_access(map->sm_shmfd, ucred,
678 			    FREAD | FWRITE);
679 			if (error)
680 				return (error);
681 			map->sm_shmfd->shm_path = NULL;
682 			LIST_REMOVE(map, sm_link);
683 			shm_drop(map->sm_shmfd);
684 			free(map->sm_path, M_SHMFD);
685 			free(map, M_SHMFD);
686 			return (0);
687 		}
688 	}
689 
690 	return (ENOENT);
691 }
692 
693 int
694 kern_shm_open(struct thread *td, const char *userpath, int flags, mode_t mode,
695     struct filecaps *fcaps)
696 {
697 	struct filedesc *fdp;
698 	struct shmfd *shmfd;
699 	struct file *fp;
700 	char *path;
701 	const char *pr_path;
702 	size_t pr_pathlen;
703 	Fnv32_t fnv;
704 	mode_t cmode;
705 	int fd, error;
706 
707 #ifdef CAPABILITY_MODE
708 	/*
709 	 * shm_open(2) is only allowed for anonymous objects.
710 	 */
711 	if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON))
712 		return (ECAPMODE);
713 #endif
714 
715 	AUDIT_ARG_FFLAGS(flags);
716 	AUDIT_ARG_MODE(mode);
717 
718 	if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
719 		return (EINVAL);
720 
721 	if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
722 		return (EINVAL);
723 
724 	fdp = td->td_proc->p_fd;
725 	cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS;
726 
727 	error = falloc_caps(td, &fp, &fd, O_CLOEXEC, fcaps);
728 	if (error)
729 		return (error);
730 
731 	/* A SHM_ANON path pointer creates an anonymous object. */
732 	if (userpath == SHM_ANON) {
733 		/* A read-only anonymous object is pointless. */
734 		if ((flags & O_ACCMODE) == O_RDONLY) {
735 			fdclose(td, fp, fd);
736 			fdrop(fp, td);
737 			return (EINVAL);
738 		}
739 		shmfd = shm_alloc(td->td_ucred, cmode);
740 	} else {
741 		path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
742 		pr_path = td->td_ucred->cr_prison->pr_path;
743 
744 		/* Construct a full pathname for jailed callers. */
745 		pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
746 		    : strlcpy(path, pr_path, MAXPATHLEN);
747 		error = copyinstr(userpath, path + pr_pathlen,
748 		    MAXPATHLEN - pr_pathlen, NULL);
749 #ifdef KTRACE
750 		if (error == 0 && KTRPOINT(curthread, KTR_NAMEI))
751 			ktrnamei(path);
752 #endif
753 		/* Require paths to start with a '/' character. */
754 		if (error == 0 && path[pr_pathlen] != '/')
755 			error = EINVAL;
756 		if (error) {
757 			fdclose(td, fp, fd);
758 			fdrop(fp, td);
759 			free(path, M_SHMFD);
760 			return (error);
761 		}
762 
763 		AUDIT_ARG_UPATH1_CANON(path);
764 		fnv = fnv_32_str(path, FNV1_32_INIT);
765 		sx_xlock(&shm_dict_lock);
766 		shmfd = shm_lookup(path, fnv);
767 		if (shmfd == NULL) {
768 			/* Object does not yet exist, create it if requested. */
769 			if (flags & O_CREAT) {
770 #ifdef MAC
771 				error = mac_posixshm_check_create(td->td_ucred,
772 				    path);
773 				if (error == 0) {
774 #endif
775 					shmfd = shm_alloc(td->td_ucred, cmode);
776 					shm_insert(path, fnv, shmfd);
777 #ifdef MAC
778 				}
779 #endif
780 			} else {
781 				free(path, M_SHMFD);
782 				error = ENOENT;
783 			}
784 		} else {
785 			/*
786 			 * Object already exists, obtain a new
787 			 * reference if requested and permitted.
788 			 */
789 			free(path, M_SHMFD);
790 			if ((flags & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
791 				error = EEXIST;
792 			else {
793 #ifdef MAC
794 				error = mac_posixshm_check_open(td->td_ucred,
795 				    shmfd, FFLAGS(flags & O_ACCMODE));
796 				if (error == 0)
797 #endif
798 				error = shm_access(shmfd, td->td_ucred,
799 				    FFLAGS(flags & O_ACCMODE));
800 			}
801 
802 			/*
803 			 * Truncate the file back to zero length if
804 			 * O_TRUNC was specified and the object was
805 			 * opened with read/write.
806 			 */
807 			if (error == 0 &&
808 			    (flags & (O_ACCMODE | O_TRUNC)) ==
809 			    (O_RDWR | O_TRUNC)) {
810 #ifdef MAC
811 				error = mac_posixshm_check_truncate(
812 					td->td_ucred, fp->f_cred, shmfd);
813 				if (error == 0)
814 #endif
815 					shm_dotruncate(shmfd, 0);
816 			}
817 			if (error == 0)
818 				shm_hold(shmfd);
819 		}
820 		sx_xunlock(&shm_dict_lock);
821 
822 		if (error) {
823 			fdclose(td, fp, fd);
824 			fdrop(fp, td);
825 			return (error);
826 		}
827 	}
828 
829 	finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
830 
831 	td->td_retval[0] = fd;
832 	fdrop(fp, td);
833 
834 	return (0);
835 }
836 
837 /* System calls. */
838 int
839 sys_shm_open(struct thread *td, struct shm_open_args *uap)
840 {
841 
842 	return (kern_shm_open(td, uap->path, uap->flags, uap->mode, NULL));
843 }
844 
845 int
846 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
847 {
848 	char *path;
849 	const char *pr_path;
850 	size_t pr_pathlen;
851 	Fnv32_t fnv;
852 	int error;
853 
854 	path = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
855 	pr_path = td->td_ucred->cr_prison->pr_path;
856 	pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
857 	    : strlcpy(path, pr_path, MAXPATHLEN);
858 	error = copyinstr(uap->path, path + pr_pathlen, MAXPATHLEN - pr_pathlen,
859 	    NULL);
860 	if (error) {
861 		free(path, M_TEMP);
862 		return (error);
863 	}
864 #ifdef KTRACE
865 	if (KTRPOINT(curthread, KTR_NAMEI))
866 		ktrnamei(path);
867 #endif
868 	AUDIT_ARG_UPATH1_CANON(path);
869 	fnv = fnv_32_str(path, FNV1_32_INIT);
870 	sx_xlock(&shm_dict_lock);
871 	error = shm_remove(path, fnv, td->td_ucred);
872 	sx_xunlock(&shm_dict_lock);
873 	free(path, M_TEMP);
874 
875 	return (error);
876 }
877 
878 int
879 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
880     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
881     vm_ooffset_t foff, struct thread *td)
882 {
883 	struct shmfd *shmfd;
884 	vm_prot_t maxprot;
885 	int error;
886 
887 	shmfd = fp->f_data;
888 	maxprot = VM_PROT_NONE;
889 
890 	/* FREAD should always be set. */
891 	if ((fp->f_flag & FREAD) != 0)
892 		maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
893 	if ((fp->f_flag & FWRITE) != 0)
894 		maxprot |= VM_PROT_WRITE;
895 
896 	/* Don't permit shared writable mappings on read-only descriptors. */
897 	if ((flags & MAP_SHARED) != 0 &&
898 	    (maxprot & VM_PROT_WRITE) == 0 &&
899 	    (prot & VM_PROT_WRITE) != 0)
900 		return (EACCES);
901 	maxprot &= cap_maxprot;
902 
903 	/* See comment in vn_mmap(). */
904 	if (
905 #ifdef _LP64
906 	    objsize > OFF_MAX ||
907 #endif
908 	    foff < 0 || foff > OFF_MAX - objsize)
909 		return (EINVAL);
910 
911 #ifdef MAC
912 	error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
913 	if (error != 0)
914 		return (error);
915 #endif
916 
917 	mtx_lock(&shm_timestamp_lock);
918 	vfs_timestamp(&shmfd->shm_atime);
919 	mtx_unlock(&shm_timestamp_lock);
920 	vm_object_reference(shmfd->shm_object);
921 
922 	error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
923 	    shmfd->shm_object, foff, FALSE, td);
924 	if (error != 0)
925 		vm_object_deallocate(shmfd->shm_object);
926 	return (error);
927 }
928 
929 static int
930 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
931     struct thread *td)
932 {
933 	struct shmfd *shmfd;
934 	int error;
935 
936 	error = 0;
937 	shmfd = fp->f_data;
938 	mtx_lock(&shm_timestamp_lock);
939 	/*
940 	 * SUSv4 says that x bits of permission need not be affected.
941 	 * Be consistent with our shm_open there.
942 	 */
943 #ifdef MAC
944 	error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
945 	if (error != 0)
946 		goto out;
947 #endif
948 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid,
949 	    shmfd->shm_gid, VADMIN, active_cred, NULL);
950 	if (error != 0)
951 		goto out;
952 	shmfd->shm_mode = mode & ACCESSPERMS;
953 out:
954 	mtx_unlock(&shm_timestamp_lock);
955 	return (error);
956 }
957 
958 static int
959 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
960     struct thread *td)
961 {
962 	struct shmfd *shmfd;
963 	int error;
964 
965 	error = 0;
966 	shmfd = fp->f_data;
967 	mtx_lock(&shm_timestamp_lock);
968 #ifdef MAC
969 	error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
970 	if (error != 0)
971 		goto out;
972 #endif
973 	if (uid == (uid_t)-1)
974 		uid = shmfd->shm_uid;
975 	if (gid == (gid_t)-1)
976                  gid = shmfd->shm_gid;
977 	if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
978 	    (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
979 	    (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN, 0)))
980 		goto out;
981 	shmfd->shm_uid = uid;
982 	shmfd->shm_gid = gid;
983 out:
984 	mtx_unlock(&shm_timestamp_lock);
985 	return (error);
986 }
987 
988 /*
989  * Helper routines to allow the backing object of a shared memory file
990  * descriptor to be mapped in the kernel.
991  */
992 int
993 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
994 {
995 	struct shmfd *shmfd;
996 	vm_offset_t kva, ofs;
997 	vm_object_t obj;
998 	int rv;
999 
1000 	if (fp->f_type != DTYPE_SHM)
1001 		return (EINVAL);
1002 	shmfd = fp->f_data;
1003 	obj = shmfd->shm_object;
1004 	VM_OBJECT_WLOCK(obj);
1005 	/*
1006 	 * XXXRW: This validation is probably insufficient, and subject to
1007 	 * sign errors.  It should be fixed.
1008 	 */
1009 	if (offset >= shmfd->shm_size ||
1010 	    offset + size > round_page(shmfd->shm_size)) {
1011 		VM_OBJECT_WUNLOCK(obj);
1012 		return (EINVAL);
1013 	}
1014 
1015 	shmfd->shm_kmappings++;
1016 	vm_object_reference_locked(obj);
1017 	VM_OBJECT_WUNLOCK(obj);
1018 
1019 	/* Map the object into the kernel_map and wire it. */
1020 	kva = vm_map_min(kernel_map);
1021 	ofs = offset & PAGE_MASK;
1022 	offset = trunc_page(offset);
1023 	size = round_page(size + ofs);
1024 	rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1025 	    VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1026 	    VM_PROT_READ | VM_PROT_WRITE, 0);
1027 	if (rv == KERN_SUCCESS) {
1028 		rv = vm_map_wire(kernel_map, kva, kva + size,
1029 		    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1030 		if (rv == KERN_SUCCESS) {
1031 			*memp = (void *)(kva + ofs);
1032 			return (0);
1033 		}
1034 		vm_map_remove(kernel_map, kva, kva + size);
1035 	} else
1036 		vm_object_deallocate(obj);
1037 
1038 	/* On failure, drop our mapping reference. */
1039 	VM_OBJECT_WLOCK(obj);
1040 	shmfd->shm_kmappings--;
1041 	VM_OBJECT_WUNLOCK(obj);
1042 
1043 	return (vm_mmap_to_errno(rv));
1044 }
1045 
1046 /*
1047  * We require the caller to unmap the entire entry.  This allows us to
1048  * safely decrement shm_kmappings when a mapping is removed.
1049  */
1050 int
1051 shm_unmap(struct file *fp, void *mem, size_t size)
1052 {
1053 	struct shmfd *shmfd;
1054 	vm_map_entry_t entry;
1055 	vm_offset_t kva, ofs;
1056 	vm_object_t obj;
1057 	vm_pindex_t pindex;
1058 	vm_prot_t prot;
1059 	boolean_t wired;
1060 	vm_map_t map;
1061 	int rv;
1062 
1063 	if (fp->f_type != DTYPE_SHM)
1064 		return (EINVAL);
1065 	shmfd = fp->f_data;
1066 	kva = (vm_offset_t)mem;
1067 	ofs = kva & PAGE_MASK;
1068 	kva = trunc_page(kva);
1069 	size = round_page(size + ofs);
1070 	map = kernel_map;
1071 	rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1072 	    &obj, &pindex, &prot, &wired);
1073 	if (rv != KERN_SUCCESS)
1074 		return (EINVAL);
1075 	if (entry->start != kva || entry->end != kva + size) {
1076 		vm_map_lookup_done(map, entry);
1077 		return (EINVAL);
1078 	}
1079 	vm_map_lookup_done(map, entry);
1080 	if (obj != shmfd->shm_object)
1081 		return (EINVAL);
1082 	vm_map_remove(map, kva, kva + size);
1083 	VM_OBJECT_WLOCK(obj);
1084 	KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1085 	shmfd->shm_kmappings--;
1086 	VM_OBJECT_WUNLOCK(obj);
1087 	return (0);
1088 }
1089 
1090 static int
1091 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1092 {
1093 	const char *path, *pr_path;
1094 	struct shmfd *shmfd;
1095 	size_t pr_pathlen;
1096 
1097 	kif->kf_type = KF_TYPE_SHM;
1098 	shmfd = fp->f_data;
1099 
1100 	mtx_lock(&shm_timestamp_lock);
1101 	kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;	/* XXX */
1102 	mtx_unlock(&shm_timestamp_lock);
1103 	kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1104 	if (shmfd->shm_path != NULL) {
1105 		sx_slock(&shm_dict_lock);
1106 		if (shmfd->shm_path != NULL) {
1107 			path = shmfd->shm_path;
1108 			pr_path = curthread->td_ucred->cr_prison->pr_path;
1109 			if (strcmp(pr_path, "/") != 0) {
1110 				/* Return the jail-rooted pathname. */
1111 				pr_pathlen = strlen(pr_path);
1112 				if (strncmp(path, pr_path, pr_pathlen) == 0 &&
1113 				    path[pr_pathlen] == '/')
1114 					path += pr_pathlen;
1115 			}
1116 			strlcpy(kif->kf_path, path, sizeof(kif->kf_path));
1117 		}
1118 		sx_sunlock(&shm_dict_lock);
1119 	}
1120 	return (0);
1121 }
1122