xref: /freebsd/sys/kern/uipc_shm.c (revision d6eb98610fa65663bf0df4574b7cb2c5c4ffda71)
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/uio.h>
61 #include <sys/signal.h>
62 #include <sys/jail.h>
63 #include <sys/ktrace.h>
64 #include <sys/lock.h>
65 #include <sys/malloc.h>
66 #include <sys/mman.h>
67 #include <sys/mutex.h>
68 #include <sys/priv.h>
69 #include <sys/proc.h>
70 #include <sys/refcount.h>
71 #include <sys/resourcevar.h>
72 #include <sys/rwlock.h>
73 #include <sys/sbuf.h>
74 #include <sys/stat.h>
75 #include <sys/syscallsubr.h>
76 #include <sys/sysctl.h>
77 #include <sys/sysproto.h>
78 #include <sys/systm.h>
79 #include <sys/sx.h>
80 #include <sys/time.h>
81 #include <sys/vnode.h>
82 #include <sys/unistd.h>
83 #include <sys/user.h>
84 
85 #include <security/audit/audit.h>
86 #include <security/mac/mac_framework.h>
87 
88 #include <vm/vm.h>
89 #include <vm/vm_param.h>
90 #include <vm/pmap.h>
91 #include <vm/vm_extern.h>
92 #include <vm/vm_map.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_object.h>
95 #include <vm/vm_page.h>
96 #include <vm/vm_pageout.h>
97 #include <vm/vm_pager.h>
98 #include <vm/swap_pager.h>
99 
100 struct shm_mapping {
101 	char		*sm_path;
102 	Fnv32_t		sm_fnv;
103 	struct shmfd	*sm_shmfd;
104 	LIST_ENTRY(shm_mapping) sm_link;
105 };
106 
107 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor");
108 static LIST_HEAD(, shm_mapping) *shm_dictionary;
109 static struct sx shm_dict_lock;
110 static struct mtx shm_timestamp_lock;
111 static u_long shm_hash;
112 static struct unrhdr64 shm_ino_unr;
113 static dev_t shm_dev_ino;
114 
115 #define	SHM_HASH(fnv)	(&shm_dictionary[(fnv) & shm_hash])
116 
117 static void	shm_init(void *arg);
118 static void	shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd);
119 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv);
120 static int	shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred);
121 
122 static fo_rdwr_t	shm_read;
123 static fo_rdwr_t	shm_write;
124 static fo_truncate_t	shm_truncate;
125 static fo_ioctl_t	shm_ioctl;
126 static fo_stat_t	shm_stat;
127 static fo_close_t	shm_close;
128 static fo_chmod_t	shm_chmod;
129 static fo_chown_t	shm_chown;
130 static fo_seek_t	shm_seek;
131 static fo_fill_kinfo_t	shm_fill_kinfo;
132 static fo_mmap_t	shm_mmap;
133 
134 /* File descriptor operations. */
135 struct fileops shm_ops = {
136 	.fo_read = shm_read,
137 	.fo_write = shm_write,
138 	.fo_truncate = shm_truncate,
139 	.fo_ioctl = shm_ioctl,
140 	.fo_poll = invfo_poll,
141 	.fo_kqfilter = invfo_kqfilter,
142 	.fo_stat = shm_stat,
143 	.fo_close = shm_close,
144 	.fo_chmod = shm_chmod,
145 	.fo_chown = shm_chown,
146 	.fo_sendfile = vn_sendfile,
147 	.fo_seek = shm_seek,
148 	.fo_fill_kinfo = shm_fill_kinfo,
149 	.fo_mmap = shm_mmap,
150 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
151 };
152 
153 FEATURE(posix_shm, "POSIX shared memory");
154 
155 static int
156 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio)
157 {
158 	vm_page_t m;
159 	vm_pindex_t idx;
160 	size_t tlen;
161 	int error, offset, rv;
162 
163 	idx = OFF_TO_IDX(uio->uio_offset);
164 	offset = uio->uio_offset & PAGE_MASK;
165 	tlen = MIN(PAGE_SIZE - offset, len);
166 
167 	VM_OBJECT_WLOCK(obj);
168 
169 	/*
170 	 * Read I/O without either a corresponding resident page or swap
171 	 * page: use zero_region.  This is intended to avoid instantiating
172 	 * pages on read from a sparse region.
173 	 */
174 	if (uio->uio_rw == UIO_READ && vm_page_lookup(obj, idx) == NULL &&
175 	    !vm_pager_has_page(obj, idx, NULL, NULL)) {
176 		VM_OBJECT_WUNLOCK(obj);
177 		return (uiomove(__DECONST(void *, zero_region), tlen, uio));
178 	}
179 
180 	/*
181 	 * Parallel reads of the page content from disk are prevented
182 	 * by exclusive busy.
183 	 *
184 	 * Although the tmpfs vnode lock is held here, it is
185 	 * nonetheless safe to sleep waiting for a free page.  The
186 	 * pageout daemon does not need to acquire the tmpfs vnode
187 	 * lock to page out tobj's pages because tobj is a OBJT_SWAP
188 	 * type object.
189 	 */
190 	m = vm_page_grab(obj, idx, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
191 	if (m->valid != VM_PAGE_BITS_ALL) {
192 		vm_page_xbusy(m);
193 		if (vm_pager_has_page(obj, idx, NULL, NULL)) {
194 			rv = vm_pager_get_pages(obj, &m, 1, NULL, NULL);
195 			if (rv != VM_PAGER_OK) {
196 				printf(
197 	    "uiomove_object: vm_obj %p idx %jd valid %x pager error %d\n",
198 				    obj, idx, m->valid, rv);
199 				vm_page_lock(m);
200 				vm_page_free(m);
201 				vm_page_unlock(m);
202 				VM_OBJECT_WUNLOCK(obj);
203 				return (EIO);
204 			}
205 		} else
206 			vm_page_zero_invalid(m, TRUE);
207 		vm_page_xunbusy(m);
208 	}
209 	vm_page_lock(m);
210 	vm_page_wire(m);
211 	vm_page_unlock(m);
212 	VM_OBJECT_WUNLOCK(obj);
213 	error = uiomove_fromphys(&m, offset, tlen, uio);
214 	if (uio->uio_rw == UIO_WRITE && error == 0) {
215 		VM_OBJECT_WLOCK(obj);
216 		vm_page_dirty(m);
217 		vm_pager_page_unswapped(m);
218 		VM_OBJECT_WUNLOCK(obj);
219 	}
220 	vm_page_lock(m);
221 	vm_page_unwire(m, PQ_ACTIVE);
222 	vm_page_unlock(m);
223 
224 	return (error);
225 }
226 
227 int
228 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
229 {
230 	ssize_t resid;
231 	size_t len;
232 	int error;
233 
234 	error = 0;
235 	while ((resid = uio->uio_resid) > 0) {
236 		if (obj_size <= uio->uio_offset)
237 			break;
238 		len = MIN(obj_size - uio->uio_offset, resid);
239 		if (len == 0)
240 			break;
241 		error = uiomove_object_page(obj, len, uio);
242 		if (error != 0 || resid == uio->uio_resid)
243 			break;
244 	}
245 	return (error);
246 }
247 
248 static int
249 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
250 {
251 	struct shmfd *shmfd;
252 	off_t foffset;
253 	int error;
254 
255 	shmfd = fp->f_data;
256 	foffset = foffset_lock(fp, 0);
257 	error = 0;
258 	switch (whence) {
259 	case L_INCR:
260 		if (foffset < 0 ||
261 		    (offset > 0 && foffset > OFF_MAX - offset)) {
262 			error = EOVERFLOW;
263 			break;
264 		}
265 		offset += foffset;
266 		break;
267 	case L_XTND:
268 		if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
269 			error = EOVERFLOW;
270 			break;
271 		}
272 		offset += shmfd->shm_size;
273 		break;
274 	case L_SET:
275 		break;
276 	default:
277 		error = EINVAL;
278 	}
279 	if (error == 0) {
280 		if (offset < 0 || offset > shmfd->shm_size)
281 			error = EINVAL;
282 		else
283 			td->td_uretoff.tdu_off = offset;
284 	}
285 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
286 	return (error);
287 }
288 
289 static int
290 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
291     int flags, struct thread *td)
292 {
293 	struct shmfd *shmfd;
294 	void *rl_cookie;
295 	int error;
296 
297 	shmfd = fp->f_data;
298 #ifdef MAC
299 	error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
300 	if (error)
301 		return (error);
302 #endif
303 	foffset_lock_uio(fp, uio, flags);
304 	rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset,
305 	    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
306 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
307 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
308 	foffset_unlock_uio(fp, uio, flags);
309 	return (error);
310 }
311 
312 static int
313 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
314     int flags, struct thread *td)
315 {
316 	struct shmfd *shmfd;
317 	void *rl_cookie;
318 	int error;
319 
320 	shmfd = fp->f_data;
321 #ifdef MAC
322 	error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
323 	if (error)
324 		return (error);
325 #endif
326 	foffset_lock_uio(fp, uio, flags);
327 	if ((flags & FOF_OFFSET) == 0) {
328 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
329 		    &shmfd->shm_mtx);
330 	} else {
331 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset,
332 		    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
333 	}
334 
335 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
336 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
337 	foffset_unlock_uio(fp, uio, flags);
338 	return (error);
339 }
340 
341 static int
342 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
343     struct thread *td)
344 {
345 	struct shmfd *shmfd;
346 #ifdef MAC
347 	int error;
348 #endif
349 
350 	shmfd = fp->f_data;
351 #ifdef MAC
352 	error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
353 	if (error)
354 		return (error);
355 #endif
356 	return (shm_dotruncate(shmfd, length));
357 }
358 
359 int
360 shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
361     struct thread *td)
362 {
363 
364 	switch (com) {
365 	case FIONBIO:
366 	case FIOASYNC:
367 		/*
368 		 * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work,
369 		 * just like it would on an unlinked regular file
370 		 */
371 		return (0);
372 	default:
373 		return (ENOTTY);
374 	}
375 }
376 
377 static int
378 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
379     struct thread *td)
380 {
381 	struct shmfd *shmfd;
382 #ifdef MAC
383 	int error;
384 #endif
385 
386 	shmfd = fp->f_data;
387 
388 #ifdef MAC
389 	error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
390 	if (error)
391 		return (error);
392 #endif
393 
394 	/*
395 	 * Attempt to return sanish values for fstat() on a memory file
396 	 * descriptor.
397 	 */
398 	bzero(sb, sizeof(*sb));
399 	sb->st_blksize = PAGE_SIZE;
400 	sb->st_size = shmfd->shm_size;
401 	sb->st_blocks = howmany(sb->st_size, sb->st_blksize);
402 	mtx_lock(&shm_timestamp_lock);
403 	sb->st_atim = shmfd->shm_atime;
404 	sb->st_ctim = shmfd->shm_ctime;
405 	sb->st_mtim = shmfd->shm_mtime;
406 	sb->st_birthtim = shmfd->shm_birthtime;
407 	sb->st_mode = S_IFREG | shmfd->shm_mode;		/* XXX */
408 	sb->st_uid = shmfd->shm_uid;
409 	sb->st_gid = shmfd->shm_gid;
410 	mtx_unlock(&shm_timestamp_lock);
411 	sb->st_dev = shm_dev_ino;
412 	sb->st_ino = shmfd->shm_ino;
413 	sb->st_nlink = shmfd->shm_object->ref_count;
414 
415 	return (0);
416 }
417 
418 static int
419 shm_close(struct file *fp, struct thread *td)
420 {
421 	struct shmfd *shmfd;
422 
423 	shmfd = fp->f_data;
424 	fp->f_data = NULL;
425 	shm_drop(shmfd);
426 
427 	return (0);
428 }
429 
430 int
431 shm_dotruncate(struct shmfd *shmfd, off_t length)
432 {
433 	vm_object_t object;
434 	vm_page_t m;
435 	vm_pindex_t idx, nobjsize;
436 	vm_ooffset_t delta;
437 	int base, rv;
438 
439 	KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
440 	object = shmfd->shm_object;
441 	VM_OBJECT_WLOCK(object);
442 	if (length == shmfd->shm_size) {
443 		VM_OBJECT_WUNLOCK(object);
444 		return (0);
445 	}
446 	nobjsize = OFF_TO_IDX(length + PAGE_MASK);
447 
448 	/* Are we shrinking?  If so, trim the end. */
449 	if (length < shmfd->shm_size) {
450 		/*
451 		 * Disallow any requests to shrink the size if this
452 		 * object is mapped into the kernel.
453 		 */
454 		if (shmfd->shm_kmappings > 0) {
455 			VM_OBJECT_WUNLOCK(object);
456 			return (EBUSY);
457 		}
458 
459 		/*
460 		 * Zero the truncated part of the last page.
461 		 */
462 		base = length & PAGE_MASK;
463 		if (base != 0) {
464 			idx = OFF_TO_IDX(length);
465 retry:
466 			m = vm_page_lookup(object, idx);
467 			if (m != NULL) {
468 				if (vm_page_sleep_if_busy(m, "shmtrc"))
469 					goto retry;
470 			} else if (vm_pager_has_page(object, idx, NULL, NULL)) {
471 				m = vm_page_alloc(object, idx,
472 				    VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL);
473 				if (m == NULL)
474 					goto retry;
475 				rv = vm_pager_get_pages(object, &m, 1, NULL,
476 				    NULL);
477 				vm_page_lock(m);
478 				if (rv == VM_PAGER_OK) {
479 					/*
480 					 * Since the page was not resident,
481 					 * and therefore not recently
482 					 * accessed, immediately enqueue it
483 					 * for asynchronous laundering.  The
484 					 * current operation is not regarded
485 					 * as an access.
486 					 */
487 					vm_page_launder(m);
488 					vm_page_unlock(m);
489 					vm_page_xunbusy(m);
490 				} else {
491 					vm_page_free(m);
492 					vm_page_unlock(m);
493 					VM_OBJECT_WUNLOCK(object);
494 					return (EIO);
495 				}
496 			}
497 			if (m != NULL) {
498 				pmap_zero_page_area(m, base, PAGE_SIZE - base);
499 				KASSERT(m->valid == VM_PAGE_BITS_ALL,
500 				    ("shm_dotruncate: page %p is invalid", m));
501 				vm_page_dirty(m);
502 				vm_pager_page_unswapped(m);
503 			}
504 		}
505 		delta = IDX_TO_OFF(object->size - nobjsize);
506 
507 		/* Toss in memory pages. */
508 		if (nobjsize < object->size)
509 			vm_object_page_remove(object, nobjsize, object->size,
510 			    0);
511 
512 		/* Toss pages from swap. */
513 		if (object->type == OBJT_SWAP)
514 			swap_pager_freespace(object, nobjsize, delta);
515 
516 		/* Free the swap accounted for shm */
517 		swap_release_by_cred(delta, object->cred);
518 		object->charge -= delta;
519 	} else {
520 		/* Try to reserve additional swap space. */
521 		delta = IDX_TO_OFF(nobjsize - object->size);
522 		if (!swap_reserve_by_cred(delta, object->cred)) {
523 			VM_OBJECT_WUNLOCK(object);
524 			return (ENOMEM);
525 		}
526 		object->charge += delta;
527 	}
528 	shmfd->shm_size = length;
529 	mtx_lock(&shm_timestamp_lock);
530 	vfs_timestamp(&shmfd->shm_ctime);
531 	shmfd->shm_mtime = shmfd->shm_ctime;
532 	mtx_unlock(&shm_timestamp_lock);
533 	object->size = nobjsize;
534 	VM_OBJECT_WUNLOCK(object);
535 	return (0);
536 }
537 
538 /*
539  * shmfd object management including creation and reference counting
540  * routines.
541  */
542 struct shmfd *
543 shm_alloc(struct ucred *ucred, mode_t mode)
544 {
545 	struct shmfd *shmfd;
546 
547 	shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
548 	shmfd->shm_size = 0;
549 	shmfd->shm_uid = ucred->cr_uid;
550 	shmfd->shm_gid = ucred->cr_gid;
551 	shmfd->shm_mode = mode;
552 	shmfd->shm_object = vm_pager_allocate(OBJT_DEFAULT, NULL,
553 	    shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
554 	KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
555 	shmfd->shm_object->pg_color = 0;
556 	VM_OBJECT_WLOCK(shmfd->shm_object);
557 	vm_object_clear_flag(shmfd->shm_object, OBJ_ONEMAPPING);
558 	vm_object_set_flag(shmfd->shm_object, OBJ_COLORED | OBJ_NOSPLIT);
559 	VM_OBJECT_WUNLOCK(shmfd->shm_object);
560 	vfs_timestamp(&shmfd->shm_birthtime);
561 	shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
562 	    shmfd->shm_birthtime;
563 	shmfd->shm_ino = alloc_unr64(&shm_ino_unr);
564 	refcount_init(&shmfd->shm_refs, 1);
565 	mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
566 	rangelock_init(&shmfd->shm_rl);
567 #ifdef MAC
568 	mac_posixshm_init(shmfd);
569 	mac_posixshm_create(ucred, shmfd);
570 #endif
571 
572 	return (shmfd);
573 }
574 
575 struct shmfd *
576 shm_hold(struct shmfd *shmfd)
577 {
578 
579 	refcount_acquire(&shmfd->shm_refs);
580 	return (shmfd);
581 }
582 
583 void
584 shm_drop(struct shmfd *shmfd)
585 {
586 
587 	if (refcount_release(&shmfd->shm_refs)) {
588 #ifdef MAC
589 		mac_posixshm_destroy(shmfd);
590 #endif
591 		rangelock_destroy(&shmfd->shm_rl);
592 		mtx_destroy(&shmfd->shm_mtx);
593 		vm_object_deallocate(shmfd->shm_object);
594 		free(shmfd, M_SHMFD);
595 	}
596 }
597 
598 /*
599  * Determine if the credentials have sufficient permissions for a
600  * specified combination of FREAD and FWRITE.
601  */
602 int
603 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
604 {
605 	accmode_t accmode;
606 	int error;
607 
608 	accmode = 0;
609 	if (flags & FREAD)
610 		accmode |= VREAD;
611 	if (flags & FWRITE)
612 		accmode |= VWRITE;
613 	mtx_lock(&shm_timestamp_lock);
614 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
615 	    accmode, ucred, NULL);
616 	mtx_unlock(&shm_timestamp_lock);
617 	return (error);
618 }
619 
620 /*
621  * Dictionary management.  We maintain an in-kernel dictionary to map
622  * paths to shmfd objects.  We use the FNV hash on the path to store
623  * the mappings in a hash table.
624  */
625 static void
626 shm_init(void *arg)
627 {
628 
629 	mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
630 	sx_init(&shm_dict_lock, "shm dictionary");
631 	shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
632 	new_unrhdr64(&shm_ino_unr, 1);
633 	shm_dev_ino = devfs_alloc_cdp_inode();
634 	KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
635 }
636 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
637 
638 static struct shmfd *
639 shm_lookup(char *path, Fnv32_t fnv)
640 {
641 	struct shm_mapping *map;
642 
643 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
644 		if (map->sm_fnv != fnv)
645 			continue;
646 		if (strcmp(map->sm_path, path) == 0)
647 			return (map->sm_shmfd);
648 	}
649 
650 	return (NULL);
651 }
652 
653 static void
654 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
655 {
656 	struct shm_mapping *map;
657 
658 	map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
659 	map->sm_path = path;
660 	map->sm_fnv = fnv;
661 	map->sm_shmfd = shm_hold(shmfd);
662 	shmfd->shm_path = path;
663 	LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
664 }
665 
666 static int
667 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
668 {
669 	struct shm_mapping *map;
670 	int error;
671 
672 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
673 		if (map->sm_fnv != fnv)
674 			continue;
675 		if (strcmp(map->sm_path, path) == 0) {
676 #ifdef MAC
677 			error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
678 			if (error)
679 				return (error);
680 #endif
681 			error = shm_access(map->sm_shmfd, ucred,
682 			    FREAD | FWRITE);
683 			if (error)
684 				return (error);
685 			map->sm_shmfd->shm_path = NULL;
686 			LIST_REMOVE(map, sm_link);
687 			shm_drop(map->sm_shmfd);
688 			free(map->sm_path, M_SHMFD);
689 			free(map, M_SHMFD);
690 			return (0);
691 		}
692 	}
693 
694 	return (ENOENT);
695 }
696 
697 int
698 kern_shm_open(struct thread *td, const char *userpath, int flags, mode_t mode,
699     struct filecaps *fcaps)
700 {
701 	struct filedesc *fdp;
702 	struct shmfd *shmfd;
703 	struct file *fp;
704 	char *path;
705 	const char *pr_path;
706 	size_t pr_pathlen;
707 	Fnv32_t fnv;
708 	mode_t cmode;
709 	int fd, error;
710 
711 #ifdef CAPABILITY_MODE
712 	/*
713 	 * shm_open(2) is only allowed for anonymous objects.
714 	 */
715 	if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON))
716 		return (ECAPMODE);
717 #endif
718 
719 	AUDIT_ARG_FFLAGS(flags);
720 	AUDIT_ARG_MODE(mode);
721 
722 	if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
723 		return (EINVAL);
724 
725 	if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
726 		return (EINVAL);
727 
728 	fdp = td->td_proc->p_fd;
729 	cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS;
730 
731 	error = falloc_caps(td, &fp, &fd, O_CLOEXEC, fcaps);
732 	if (error)
733 		return (error);
734 
735 	/* A SHM_ANON path pointer creates an anonymous object. */
736 	if (userpath == SHM_ANON) {
737 		/* A read-only anonymous object is pointless. */
738 		if ((flags & O_ACCMODE) == O_RDONLY) {
739 			fdclose(td, fp, fd);
740 			fdrop(fp, td);
741 			return (EINVAL);
742 		}
743 		shmfd = shm_alloc(td->td_ucred, cmode);
744 	} else {
745 		path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
746 		pr_path = td->td_ucred->cr_prison->pr_path;
747 
748 		/* Construct a full pathname for jailed callers. */
749 		pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
750 		    : strlcpy(path, pr_path, MAXPATHLEN);
751 		error = copyinstr(userpath, path + pr_pathlen,
752 		    MAXPATHLEN - pr_pathlen, NULL);
753 #ifdef KTRACE
754 		if (error == 0 && KTRPOINT(curthread, KTR_NAMEI))
755 			ktrnamei(path);
756 #endif
757 		/* Require paths to start with a '/' character. */
758 		if (error == 0 && path[pr_pathlen] != '/')
759 			error = EINVAL;
760 		if (error) {
761 			fdclose(td, fp, fd);
762 			fdrop(fp, td);
763 			free(path, M_SHMFD);
764 			return (error);
765 		}
766 
767 		AUDIT_ARG_UPATH1_CANON(path);
768 		fnv = fnv_32_str(path, FNV1_32_INIT);
769 		sx_xlock(&shm_dict_lock);
770 		shmfd = shm_lookup(path, fnv);
771 		if (shmfd == NULL) {
772 			/* Object does not yet exist, create it if requested. */
773 			if (flags & O_CREAT) {
774 #ifdef MAC
775 				error = mac_posixshm_check_create(td->td_ucred,
776 				    path);
777 				if (error == 0) {
778 #endif
779 					shmfd = shm_alloc(td->td_ucred, cmode);
780 					shm_insert(path, fnv, shmfd);
781 #ifdef MAC
782 				}
783 #endif
784 			} else {
785 				free(path, M_SHMFD);
786 				error = ENOENT;
787 			}
788 		} else {
789 			/*
790 			 * Object already exists, obtain a new
791 			 * reference if requested and permitted.
792 			 */
793 			free(path, M_SHMFD);
794 			if ((flags & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
795 				error = EEXIST;
796 			else {
797 #ifdef MAC
798 				error = mac_posixshm_check_open(td->td_ucred,
799 				    shmfd, FFLAGS(flags & O_ACCMODE));
800 				if (error == 0)
801 #endif
802 				error = shm_access(shmfd, td->td_ucred,
803 				    FFLAGS(flags & O_ACCMODE));
804 			}
805 
806 			/*
807 			 * Truncate the file back to zero length if
808 			 * O_TRUNC was specified and the object was
809 			 * opened with read/write.
810 			 */
811 			if (error == 0 &&
812 			    (flags & (O_ACCMODE | O_TRUNC)) ==
813 			    (O_RDWR | O_TRUNC)) {
814 #ifdef MAC
815 				error = mac_posixshm_check_truncate(
816 					td->td_ucred, fp->f_cred, shmfd);
817 				if (error == 0)
818 #endif
819 					shm_dotruncate(shmfd, 0);
820 			}
821 			if (error == 0)
822 				shm_hold(shmfd);
823 		}
824 		sx_xunlock(&shm_dict_lock);
825 
826 		if (error) {
827 			fdclose(td, fp, fd);
828 			fdrop(fp, td);
829 			return (error);
830 		}
831 	}
832 
833 	finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
834 
835 	td->td_retval[0] = fd;
836 	fdrop(fp, td);
837 
838 	return (0);
839 }
840 
841 /* System calls. */
842 int
843 sys_shm_open(struct thread *td, struct shm_open_args *uap)
844 {
845 
846 	return (kern_shm_open(td, uap->path, uap->flags, uap->mode, NULL));
847 }
848 
849 int
850 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
851 {
852 	char *path;
853 	const char *pr_path;
854 	size_t pr_pathlen;
855 	Fnv32_t fnv;
856 	int error;
857 
858 	path = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
859 	pr_path = td->td_ucred->cr_prison->pr_path;
860 	pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
861 	    : strlcpy(path, pr_path, MAXPATHLEN);
862 	error = copyinstr(uap->path, path + pr_pathlen, MAXPATHLEN - pr_pathlen,
863 	    NULL);
864 	if (error) {
865 		free(path, M_TEMP);
866 		return (error);
867 	}
868 #ifdef KTRACE
869 	if (KTRPOINT(curthread, KTR_NAMEI))
870 		ktrnamei(path);
871 #endif
872 	AUDIT_ARG_UPATH1_CANON(path);
873 	fnv = fnv_32_str(path, FNV1_32_INIT);
874 	sx_xlock(&shm_dict_lock);
875 	error = shm_remove(path, fnv, td->td_ucred);
876 	sx_xunlock(&shm_dict_lock);
877 	free(path, M_TEMP);
878 
879 	return (error);
880 }
881 
882 int
883 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
884     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
885     vm_ooffset_t foff, struct thread *td)
886 {
887 	struct shmfd *shmfd;
888 	vm_prot_t maxprot;
889 	int error;
890 
891 	shmfd = fp->f_data;
892 	maxprot = VM_PROT_NONE;
893 
894 	/* FREAD should always be set. */
895 	if ((fp->f_flag & FREAD) != 0)
896 		maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
897 	if ((fp->f_flag & FWRITE) != 0)
898 		maxprot |= VM_PROT_WRITE;
899 
900 	/* Don't permit shared writable mappings on read-only descriptors. */
901 	if ((flags & MAP_SHARED) != 0 &&
902 	    (maxprot & VM_PROT_WRITE) == 0 &&
903 	    (prot & VM_PROT_WRITE) != 0)
904 		return (EACCES);
905 	maxprot &= cap_maxprot;
906 
907 	/* See comment in vn_mmap(). */
908 	if (
909 #ifdef _LP64
910 	    objsize > OFF_MAX ||
911 #endif
912 	    foff < 0 || foff > OFF_MAX - objsize)
913 		return (EINVAL);
914 
915 #ifdef MAC
916 	error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
917 	if (error != 0)
918 		return (error);
919 #endif
920 
921 	mtx_lock(&shm_timestamp_lock);
922 	vfs_timestamp(&shmfd->shm_atime);
923 	mtx_unlock(&shm_timestamp_lock);
924 	vm_object_reference(shmfd->shm_object);
925 
926 	error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
927 	    shmfd->shm_object, foff, FALSE, td);
928 	if (error != 0)
929 		vm_object_deallocate(shmfd->shm_object);
930 	return (error);
931 }
932 
933 static int
934 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
935     struct thread *td)
936 {
937 	struct shmfd *shmfd;
938 	int error;
939 
940 	error = 0;
941 	shmfd = fp->f_data;
942 	mtx_lock(&shm_timestamp_lock);
943 	/*
944 	 * SUSv4 says that x bits of permission need not be affected.
945 	 * Be consistent with our shm_open there.
946 	 */
947 #ifdef MAC
948 	error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
949 	if (error != 0)
950 		goto out;
951 #endif
952 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid,
953 	    shmfd->shm_gid, VADMIN, active_cred, NULL);
954 	if (error != 0)
955 		goto out;
956 	shmfd->shm_mode = mode & ACCESSPERMS;
957 out:
958 	mtx_unlock(&shm_timestamp_lock);
959 	return (error);
960 }
961 
962 static int
963 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
964     struct thread *td)
965 {
966 	struct shmfd *shmfd;
967 	int error;
968 
969 	error = 0;
970 	shmfd = fp->f_data;
971 	mtx_lock(&shm_timestamp_lock);
972 #ifdef MAC
973 	error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
974 	if (error != 0)
975 		goto out;
976 #endif
977 	if (uid == (uid_t)-1)
978 		uid = shmfd->shm_uid;
979 	if (gid == (gid_t)-1)
980                  gid = shmfd->shm_gid;
981 	if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
982 	    (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
983 	    (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN)))
984 		goto out;
985 	shmfd->shm_uid = uid;
986 	shmfd->shm_gid = gid;
987 out:
988 	mtx_unlock(&shm_timestamp_lock);
989 	return (error);
990 }
991 
992 /*
993  * Helper routines to allow the backing object of a shared memory file
994  * descriptor to be mapped in the kernel.
995  */
996 int
997 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
998 {
999 	struct shmfd *shmfd;
1000 	vm_offset_t kva, ofs;
1001 	vm_object_t obj;
1002 	int rv;
1003 
1004 	if (fp->f_type != DTYPE_SHM)
1005 		return (EINVAL);
1006 	shmfd = fp->f_data;
1007 	obj = shmfd->shm_object;
1008 	VM_OBJECT_WLOCK(obj);
1009 	/*
1010 	 * XXXRW: This validation is probably insufficient, and subject to
1011 	 * sign errors.  It should be fixed.
1012 	 */
1013 	if (offset >= shmfd->shm_size ||
1014 	    offset + size > round_page(shmfd->shm_size)) {
1015 		VM_OBJECT_WUNLOCK(obj);
1016 		return (EINVAL);
1017 	}
1018 
1019 	shmfd->shm_kmappings++;
1020 	vm_object_reference_locked(obj);
1021 	VM_OBJECT_WUNLOCK(obj);
1022 
1023 	/* Map the object into the kernel_map and wire it. */
1024 	kva = vm_map_min(kernel_map);
1025 	ofs = offset & PAGE_MASK;
1026 	offset = trunc_page(offset);
1027 	size = round_page(size + ofs);
1028 	rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1029 	    VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1030 	    VM_PROT_READ | VM_PROT_WRITE, 0);
1031 	if (rv == KERN_SUCCESS) {
1032 		rv = vm_map_wire(kernel_map, kva, kva + size,
1033 		    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1034 		if (rv == KERN_SUCCESS) {
1035 			*memp = (void *)(kva + ofs);
1036 			return (0);
1037 		}
1038 		vm_map_remove(kernel_map, kva, kva + size);
1039 	} else
1040 		vm_object_deallocate(obj);
1041 
1042 	/* On failure, drop our mapping reference. */
1043 	VM_OBJECT_WLOCK(obj);
1044 	shmfd->shm_kmappings--;
1045 	VM_OBJECT_WUNLOCK(obj);
1046 
1047 	return (vm_mmap_to_errno(rv));
1048 }
1049 
1050 /*
1051  * We require the caller to unmap the entire entry.  This allows us to
1052  * safely decrement shm_kmappings when a mapping is removed.
1053  */
1054 int
1055 shm_unmap(struct file *fp, void *mem, size_t size)
1056 {
1057 	struct shmfd *shmfd;
1058 	vm_map_entry_t entry;
1059 	vm_offset_t kva, ofs;
1060 	vm_object_t obj;
1061 	vm_pindex_t pindex;
1062 	vm_prot_t prot;
1063 	boolean_t wired;
1064 	vm_map_t map;
1065 	int rv;
1066 
1067 	if (fp->f_type != DTYPE_SHM)
1068 		return (EINVAL);
1069 	shmfd = fp->f_data;
1070 	kva = (vm_offset_t)mem;
1071 	ofs = kva & PAGE_MASK;
1072 	kva = trunc_page(kva);
1073 	size = round_page(size + ofs);
1074 	map = kernel_map;
1075 	rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1076 	    &obj, &pindex, &prot, &wired);
1077 	if (rv != KERN_SUCCESS)
1078 		return (EINVAL);
1079 	if (entry->start != kva || entry->end != kva + size) {
1080 		vm_map_lookup_done(map, entry);
1081 		return (EINVAL);
1082 	}
1083 	vm_map_lookup_done(map, entry);
1084 	if (obj != shmfd->shm_object)
1085 		return (EINVAL);
1086 	vm_map_remove(map, kva, kva + size);
1087 	VM_OBJECT_WLOCK(obj);
1088 	KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1089 	shmfd->shm_kmappings--;
1090 	VM_OBJECT_WUNLOCK(obj);
1091 	return (0);
1092 }
1093 
1094 static int
1095 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list)
1096 {
1097 	const char *path, *pr_path;
1098 	size_t pr_pathlen;
1099 	bool visible;
1100 
1101 	sx_assert(&shm_dict_lock, SA_LOCKED);
1102 	kif->kf_type = KF_TYPE_SHM;
1103 	kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;
1104 	kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1105 	if (shmfd->shm_path != NULL) {
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 				visible = strncmp(path, pr_path, pr_pathlen)
1113 				    == 0 && path[pr_pathlen] == '/';
1114 				if (list && !visible)
1115 					return (EPERM);
1116 				if (visible)
1117 					path += pr_pathlen;
1118 			}
1119 			strlcpy(kif->kf_path, path, sizeof(kif->kf_path));
1120 		}
1121 	}
1122 	return (0);
1123 }
1124 
1125 static int
1126 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif,
1127     struct filedesc *fdp __unused)
1128 {
1129 	int res;
1130 
1131 	sx_slock(&shm_dict_lock);
1132 	res = shm_fill_kinfo_locked(fp->f_data, kif, false);
1133 	sx_sunlock(&shm_dict_lock);
1134 	return (res);
1135 }
1136 
1137 static int
1138 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS)
1139 {
1140 	struct shm_mapping *shmm;
1141 	struct sbuf sb;
1142 	struct kinfo_file kif;
1143 	u_long i;
1144 	ssize_t curlen;
1145 	int error, error2;
1146 
1147 	sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req);
1148 	sbuf_clear_flags(&sb, SBUF_INCLUDENUL);
1149 	curlen = 0;
1150 	error = 0;
1151 	sx_slock(&shm_dict_lock);
1152 	for (i = 0; i < shm_hash + 1; i++) {
1153 		LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) {
1154 			error = shm_fill_kinfo_locked(shmm->sm_shmfd,
1155 			    &kif, true);
1156 			if (error == EPERM)
1157 				continue;
1158 			if (error != 0)
1159 				break;
1160 			pack_kinfo(&kif);
1161 			if (req->oldptr != NULL &&
1162 			    kif.kf_structsize + curlen > req->oldlen)
1163 				break;
1164 			error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ?
1165 			    0 : ENOMEM;
1166 			if (error != 0)
1167 				break;
1168 			curlen += kif.kf_structsize;
1169 		}
1170 	}
1171 	sx_sunlock(&shm_dict_lock);
1172 	error2 = sbuf_finish(&sb);
1173 	sbuf_delete(&sb);
1174 	return (error != 0 ? error : error2);
1175 }
1176 
1177 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list,
1178     CTLFLAG_RD | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE,
1179     NULL, 0, sysctl_posix_shm_list, "",
1180     "POSIX SHM list");
1181