xref: /freebsd/sys/kern/uipc_shm.c (revision 41059135ce931c0f1014a999ffabc6bc470ce856)
1 /*-
2  * Copyright (c) 2006, 2011, 2016-2017 Robert N. M. Watson
3  * All rights reserved.
4  *
5  * Portions of this software were developed by BAE Systems, the University of
6  * Cambridge Computer Laboratory, and Memorial University under DARPA/AFRL
7  * contract FA8650-15-C-7558 ("CADETS"), as part of the DARPA Transparent
8  * Computing (TC) research program.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 /*
33  * Support for shared swap-backed anonymous memory objects via
34  * shm_open(2) and shm_unlink(2).  While most of the implementation is
35  * here, vm_mmap.c contains mapping logic changes.
36  *
37  * TODO:
38  *
39  * (1) Need to export data to a userland tool via a sysctl.  Should ipcs(1)
40  *     and ipcrm(1) be expanded or should new tools to manage both POSIX
41  *     kernel semaphores and POSIX shared memory be written?
42  *
43  * (2) Add support for this file type to fstat(1).
44  *
45  * (3) Resource limits?  Does this need its own resource limits or are the
46  *     existing limits in mmap(2) sufficient?
47  */
48 
49 #include <sys/cdefs.h>
50 __FBSDID("$FreeBSD$");
51 
52 #include "opt_capsicum.h"
53 #include "opt_ktrace.h"
54 
55 #include <sys/param.h>
56 #include <sys/capsicum.h>
57 #include <sys/conf.h>
58 #include <sys/fcntl.h>
59 #include <sys/file.h>
60 #include <sys/filedesc.h>
61 #include <sys/fnv_hash.h>
62 #include <sys/kernel.h>
63 #include <sys/uio.h>
64 #include <sys/signal.h>
65 #include <sys/jail.h>
66 #include <sys/ktrace.h>
67 #include <sys/lock.h>
68 #include <sys/malloc.h>
69 #include <sys/mman.h>
70 #include <sys/mutex.h>
71 #include <sys/priv.h>
72 #include <sys/proc.h>
73 #include <sys/refcount.h>
74 #include <sys/resourcevar.h>
75 #include <sys/rwlock.h>
76 #include <sys/stat.h>
77 #include <sys/syscallsubr.h>
78 #include <sys/sysctl.h>
79 #include <sys/sysproto.h>
80 #include <sys/systm.h>
81 #include <sys/sx.h>
82 #include <sys/time.h>
83 #include <sys/vnode.h>
84 #include <sys/unistd.h>
85 #include <sys/user.h>
86 
87 #include <security/audit/audit.h>
88 #include <security/mac/mac_framework.h>
89 
90 #include <vm/vm.h>
91 #include <vm/vm_param.h>
92 #include <vm/pmap.h>
93 #include <vm/vm_extern.h>
94 #include <vm/vm_map.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_pageout.h>
99 #include <vm/vm_pager.h>
100 #include <vm/swap_pager.h>
101 
102 struct shm_mapping {
103 	char		*sm_path;
104 	Fnv32_t		sm_fnv;
105 	struct shmfd	*sm_shmfd;
106 	LIST_ENTRY(shm_mapping) sm_link;
107 };
108 
109 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor");
110 static LIST_HEAD(, shm_mapping) *shm_dictionary;
111 static struct sx shm_dict_lock;
112 static struct mtx shm_timestamp_lock;
113 static u_long shm_hash;
114 static struct unrhdr *shm_ino_unr;
115 static dev_t shm_dev_ino;
116 
117 #define	SHM_HASH(fnv)	(&shm_dictionary[(fnv) & shm_hash])
118 
119 static void	shm_init(void *arg);
120 static void	shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd);
121 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv);
122 static int	shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred);
123 
124 static fo_rdwr_t	shm_read;
125 static fo_rdwr_t	shm_write;
126 static fo_truncate_t	shm_truncate;
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 = invfo_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 	m = vm_page_grab(obj, idx, VM_ALLOC_NORMAL | VM_ALLOC_NOBUSY);
192 	if (m->valid != VM_PAGE_BITS_ALL) {
193 		vm_page_xbusy(m);
194 		if (vm_pager_has_page(obj, idx, NULL, NULL)) {
195 			rv = vm_pager_get_pages(obj, &m, 1, NULL, NULL);
196 			if (rv != VM_PAGER_OK) {
197 				printf(
198 	    "uiomove_object: vm_obj %p idx %jd valid %x pager error %d\n",
199 				    obj, idx, m->valid, rv);
200 				vm_page_lock(m);
201 				vm_page_free(m);
202 				vm_page_unlock(m);
203 				VM_OBJECT_WUNLOCK(obj);
204 				return (EIO);
205 			}
206 		} else
207 			vm_page_zero_invalid(m, TRUE);
208 		vm_page_xunbusy(m);
209 	}
210 	vm_page_lock(m);
211 	vm_page_hold(m);
212 	if (m->queue == PQ_NONE) {
213 		vm_page_deactivate(m);
214 	} else {
215 		/* Requeue to maintain LRU ordering. */
216 		vm_page_requeue(m);
217 	}
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, VM_ALLOC_NORMAL);
460 				if (m == NULL) {
461 					VM_OBJECT_WUNLOCK(object);
462 					VM_WAIT;
463 					VM_OBJECT_WLOCK(object);
464 					goto retry;
465 				}
466 				rv = vm_pager_get_pages(object, &m, 1, NULL,
467 				    NULL);
468 				vm_page_lock(m);
469 				if (rv == VM_PAGER_OK) {
470 					/*
471 					 * Since the page was not resident,
472 					 * and therefore not recently
473 					 * accessed, immediately enqueue it
474 					 * for asynchronous laundering.  The
475 					 * current operation is not regarded
476 					 * as an access.
477 					 */
478 					vm_page_launder(m);
479 					vm_page_unlock(m);
480 					vm_page_xunbusy(m);
481 				} else {
482 					vm_page_free(m);
483 					vm_page_unlock(m);
484 					VM_OBJECT_WUNLOCK(object);
485 					return (EIO);
486 				}
487 			}
488 			if (m != NULL) {
489 				pmap_zero_page_area(m, base, PAGE_SIZE - base);
490 				KASSERT(m->valid == VM_PAGE_BITS_ALL,
491 				    ("shm_dotruncate: page %p is invalid", m));
492 				vm_page_dirty(m);
493 				vm_pager_page_unswapped(m);
494 			}
495 		}
496 		delta = IDX_TO_OFF(object->size - nobjsize);
497 
498 		/* Toss in memory pages. */
499 		if (nobjsize < object->size)
500 			vm_object_page_remove(object, nobjsize, object->size,
501 			    0);
502 
503 		/* Toss pages from swap. */
504 		if (object->type == OBJT_SWAP)
505 			swap_pager_freespace(object, nobjsize, delta);
506 
507 		/* Free the swap accounted for shm */
508 		swap_release_by_cred(delta, object->cred);
509 		object->charge -= delta;
510 	} else {
511 		/* Try to reserve additional swap space. */
512 		delta = IDX_TO_OFF(nobjsize - object->size);
513 		if (!swap_reserve_by_cred(delta, object->cred)) {
514 			VM_OBJECT_WUNLOCK(object);
515 			return (ENOMEM);
516 		}
517 		object->charge += delta;
518 	}
519 	shmfd->shm_size = length;
520 	mtx_lock(&shm_timestamp_lock);
521 	vfs_timestamp(&shmfd->shm_ctime);
522 	shmfd->shm_mtime = shmfd->shm_ctime;
523 	mtx_unlock(&shm_timestamp_lock);
524 	object->size = nobjsize;
525 	VM_OBJECT_WUNLOCK(object);
526 	return (0);
527 }
528 
529 /*
530  * shmfd object management including creation and reference counting
531  * routines.
532  */
533 struct shmfd *
534 shm_alloc(struct ucred *ucred, mode_t mode)
535 {
536 	struct shmfd *shmfd;
537 	int ino;
538 
539 	shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
540 	shmfd->shm_size = 0;
541 	shmfd->shm_uid = ucred->cr_uid;
542 	shmfd->shm_gid = ucred->cr_gid;
543 	shmfd->shm_mode = mode;
544 	shmfd->shm_object = vm_pager_allocate(OBJT_DEFAULT, NULL,
545 	    shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
546 	KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
547 	shmfd->shm_object->pg_color = 0;
548 	VM_OBJECT_WLOCK(shmfd->shm_object);
549 	vm_object_clear_flag(shmfd->shm_object, OBJ_ONEMAPPING);
550 	vm_object_set_flag(shmfd->shm_object, OBJ_COLORED | OBJ_NOSPLIT);
551 	VM_OBJECT_WUNLOCK(shmfd->shm_object);
552 	vfs_timestamp(&shmfd->shm_birthtime);
553 	shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
554 	    shmfd->shm_birthtime;
555 	ino = alloc_unr(shm_ino_unr);
556 	if (ino == -1)
557 		shmfd->shm_ino = 0;
558 	else
559 		shmfd->shm_ino = ino;
560 	refcount_init(&shmfd->shm_refs, 1);
561 	mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
562 	rangelock_init(&shmfd->shm_rl);
563 #ifdef MAC
564 	mac_posixshm_init(shmfd);
565 	mac_posixshm_create(ucred, shmfd);
566 #endif
567 
568 	return (shmfd);
569 }
570 
571 struct shmfd *
572 shm_hold(struct shmfd *shmfd)
573 {
574 
575 	refcount_acquire(&shmfd->shm_refs);
576 	return (shmfd);
577 }
578 
579 void
580 shm_drop(struct shmfd *shmfd)
581 {
582 
583 	if (refcount_release(&shmfd->shm_refs)) {
584 #ifdef MAC
585 		mac_posixshm_destroy(shmfd);
586 #endif
587 		rangelock_destroy(&shmfd->shm_rl);
588 		mtx_destroy(&shmfd->shm_mtx);
589 		vm_object_deallocate(shmfd->shm_object);
590 		if (shmfd->shm_ino != 0)
591 			free_unr(shm_ino_unr, shmfd->shm_ino);
592 		free(shmfd, M_SHMFD);
593 	}
594 }
595 
596 /*
597  * Determine if the credentials have sufficient permissions for a
598  * specified combination of FREAD and FWRITE.
599  */
600 int
601 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
602 {
603 	accmode_t accmode;
604 	int error;
605 
606 	accmode = 0;
607 	if (flags & FREAD)
608 		accmode |= VREAD;
609 	if (flags & FWRITE)
610 		accmode |= VWRITE;
611 	mtx_lock(&shm_timestamp_lock);
612 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
613 	    accmode, ucred, NULL);
614 	mtx_unlock(&shm_timestamp_lock);
615 	return (error);
616 }
617 
618 /*
619  * Dictionary management.  We maintain an in-kernel dictionary to map
620  * paths to shmfd objects.  We use the FNV hash on the path to store
621  * the mappings in a hash table.
622  */
623 static void
624 shm_init(void *arg)
625 {
626 
627 	mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
628 	sx_init(&shm_dict_lock, "shm dictionary");
629 	shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
630 	shm_ino_unr = new_unrhdr(1, INT32_MAX, NULL);
631 	KASSERT(shm_ino_unr != NULL, ("shm fake inodes not initialized"));
632 	shm_dev_ino = devfs_alloc_cdp_inode();
633 	KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
634 }
635 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
636 
637 static struct shmfd *
638 shm_lookup(char *path, Fnv32_t fnv)
639 {
640 	struct shm_mapping *map;
641 
642 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
643 		if (map->sm_fnv != fnv)
644 			continue;
645 		if (strcmp(map->sm_path, path) == 0)
646 			return (map->sm_shmfd);
647 	}
648 
649 	return (NULL);
650 }
651 
652 static void
653 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
654 {
655 	struct shm_mapping *map;
656 
657 	map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
658 	map->sm_path = path;
659 	map->sm_fnv = fnv;
660 	map->sm_shmfd = shm_hold(shmfd);
661 	shmfd->shm_path = path;
662 	LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
663 }
664 
665 static int
666 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
667 {
668 	struct shm_mapping *map;
669 	int error;
670 
671 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
672 		if (map->sm_fnv != fnv)
673 			continue;
674 		if (strcmp(map->sm_path, path) == 0) {
675 #ifdef MAC
676 			error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
677 			if (error)
678 				return (error);
679 #endif
680 			error = shm_access(map->sm_shmfd, ucred,
681 			    FREAD | FWRITE);
682 			if (error)
683 				return (error);
684 			map->sm_shmfd->shm_path = NULL;
685 			LIST_REMOVE(map, sm_link);
686 			shm_drop(map->sm_shmfd);
687 			free(map->sm_path, M_SHMFD);
688 			free(map, M_SHMFD);
689 			return (0);
690 		}
691 	}
692 
693 	return (ENOENT);
694 }
695 
696 int
697 kern_shm_open(struct thread *td, const char *userpath, int flags, mode_t mode,
698     struct filecaps *fcaps)
699 {
700 	struct filedesc *fdp;
701 	struct shmfd *shmfd;
702 	struct file *fp;
703 	char *path;
704 	const char *pr_path;
705 	size_t pr_pathlen;
706 	Fnv32_t fnv;
707 	mode_t cmode;
708 	int fd, error;
709 
710 #ifdef CAPABILITY_MODE
711 	/*
712 	 * shm_open(2) is only allowed for anonymous objects.
713 	 */
714 	if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON))
715 		return (ECAPMODE);
716 #endif
717 
718 	AUDIT_ARG_FFLAGS(flags);
719 	AUDIT_ARG_MODE(mode);
720 
721 	if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
722 		return (EINVAL);
723 
724 	if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
725 		return (EINVAL);
726 
727 	fdp = td->td_proc->p_fd;
728 	cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS;
729 
730 	error = falloc_caps(td, &fp, &fd, O_CLOEXEC, fcaps);
731 	if (error)
732 		return (error);
733 
734 	/* A SHM_ANON path pointer creates an anonymous object. */
735 	if (userpath == SHM_ANON) {
736 		/* A read-only anonymous object is pointless. */
737 		if ((flags & O_ACCMODE) == O_RDONLY) {
738 			fdclose(td, fp, fd);
739 			fdrop(fp, td);
740 			return (EINVAL);
741 		}
742 		shmfd = shm_alloc(td->td_ucred, cmode);
743 	} else {
744 		path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
745 		pr_path = td->td_ucred->cr_prison->pr_path;
746 
747 		/* Construct a full pathname for jailed callers. */
748 		pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
749 		    : strlcpy(path, pr_path, MAXPATHLEN);
750 		error = copyinstr(userpath, path + pr_pathlen,
751 		    MAXPATHLEN - pr_pathlen, NULL);
752 #ifdef KTRACE
753 		if (error == 0 && KTRPOINT(curthread, KTR_NAMEI))
754 			ktrnamei(path);
755 #endif
756 		/* Require paths to start with a '/' character. */
757 		if (error == 0 && path[pr_pathlen] != '/')
758 			error = EINVAL;
759 		if (error) {
760 			fdclose(td, fp, fd);
761 			fdrop(fp, td);
762 			free(path, M_SHMFD);
763 			return (error);
764 		}
765 
766 		AUDIT_ARG_UPATH1_CANON(path);
767 		fnv = fnv_32_str(path, FNV1_32_INIT);
768 		sx_xlock(&shm_dict_lock);
769 		shmfd = shm_lookup(path, fnv);
770 		if (shmfd == NULL) {
771 			/* Object does not yet exist, create it if requested. */
772 			if (flags & O_CREAT) {
773 #ifdef MAC
774 				error = mac_posixshm_check_create(td->td_ucred,
775 				    path);
776 				if (error == 0) {
777 #endif
778 					shmfd = shm_alloc(td->td_ucred, cmode);
779 					shm_insert(path, fnv, shmfd);
780 #ifdef MAC
781 				}
782 #endif
783 			} else {
784 				free(path, M_SHMFD);
785 				error = ENOENT;
786 			}
787 		} else {
788 			/*
789 			 * Object already exists, obtain a new
790 			 * reference if requested and permitted.
791 			 */
792 			free(path, M_SHMFD);
793 			if ((flags & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
794 				error = EEXIST;
795 			else {
796 #ifdef MAC
797 				error = mac_posixshm_check_open(td->td_ucred,
798 				    shmfd, FFLAGS(flags & O_ACCMODE));
799 				if (error == 0)
800 #endif
801 				error = shm_access(shmfd, td->td_ucred,
802 				    FFLAGS(flags & O_ACCMODE));
803 			}
804 
805 			/*
806 			 * Truncate the file back to zero length if
807 			 * O_TRUNC was specified and the object was
808 			 * opened with read/write.
809 			 */
810 			if (error == 0 &&
811 			    (flags & (O_ACCMODE | O_TRUNC)) ==
812 			    (O_RDWR | O_TRUNC)) {
813 #ifdef MAC
814 				error = mac_posixshm_check_truncate(
815 					td->td_ucred, fp->f_cred, shmfd);
816 				if (error == 0)
817 #endif
818 					shm_dotruncate(shmfd, 0);
819 			}
820 			if (error == 0)
821 				shm_hold(shmfd);
822 		}
823 		sx_xunlock(&shm_dict_lock);
824 
825 		if (error) {
826 			fdclose(td, fp, fd);
827 			fdrop(fp, td);
828 			return (error);
829 		}
830 	}
831 
832 	finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
833 
834 	td->td_retval[0] = fd;
835 	fdrop(fp, td);
836 
837 	return (0);
838 }
839 
840 /* System calls. */
841 int
842 sys_shm_open(struct thread *td, struct shm_open_args *uap)
843 {
844 
845 	return (kern_shm_open(td, uap->path, uap->flags, uap->mode, NULL));
846 }
847 
848 int
849 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
850 {
851 	char *path;
852 	const char *pr_path;
853 	size_t pr_pathlen;
854 	Fnv32_t fnv;
855 	int error;
856 
857 	path = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
858 	pr_path = td->td_ucred->cr_prison->pr_path;
859 	pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
860 	    : strlcpy(path, pr_path, MAXPATHLEN);
861 	error = copyinstr(uap->path, path + pr_pathlen, MAXPATHLEN - pr_pathlen,
862 	    NULL);
863 	if (error) {
864 		free(path, M_TEMP);
865 		return (error);
866 	}
867 #ifdef KTRACE
868 	if (KTRPOINT(curthread, KTR_NAMEI))
869 		ktrnamei(path);
870 #endif
871 	AUDIT_ARG_UPATH1_CANON(path);
872 	fnv = fnv_32_str(path, FNV1_32_INIT);
873 	sx_xlock(&shm_dict_lock);
874 	error = shm_remove(path, fnv, td->td_ucred);
875 	sx_xunlock(&shm_dict_lock);
876 	free(path, M_TEMP);
877 
878 	return (error);
879 }
880 
881 int
882 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
883     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
884     vm_ooffset_t foff, struct thread *td)
885 {
886 	struct shmfd *shmfd;
887 	vm_prot_t maxprot;
888 	int error;
889 
890 	shmfd = fp->f_data;
891 	maxprot = VM_PROT_NONE;
892 
893 	/* FREAD should always be set. */
894 	if ((fp->f_flag & FREAD) != 0)
895 		maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
896 	if ((fp->f_flag & FWRITE) != 0)
897 		maxprot |= VM_PROT_WRITE;
898 
899 	/* Don't permit shared writable mappings on read-only descriptors. */
900 	if ((flags & MAP_SHARED) != 0 &&
901 	    (maxprot & VM_PROT_WRITE) == 0 &&
902 	    (prot & VM_PROT_WRITE) != 0)
903 		return (EACCES);
904 	maxprot &= cap_maxprot;
905 
906 	/* See comment in vn_mmap(). */
907 	if (
908 #ifdef _LP64
909 	    objsize > OFF_MAX ||
910 #endif
911 	    foff < 0 || foff > OFF_MAX - objsize)
912 		return (EINVAL);
913 
914 #ifdef MAC
915 	error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
916 	if (error != 0)
917 		return (error);
918 #endif
919 
920 	mtx_lock(&shm_timestamp_lock);
921 	vfs_timestamp(&shmfd->shm_atime);
922 	mtx_unlock(&shm_timestamp_lock);
923 	vm_object_reference(shmfd->shm_object);
924 
925 	error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
926 	    shmfd->shm_object, foff, FALSE, td);
927 	if (error != 0)
928 		vm_object_deallocate(shmfd->shm_object);
929 	return (0);
930 }
931 
932 static int
933 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
934     struct thread *td)
935 {
936 	struct shmfd *shmfd;
937 	int error;
938 
939 	error = 0;
940 	shmfd = fp->f_data;
941 	mtx_lock(&shm_timestamp_lock);
942 	/*
943 	 * SUSv4 says that x bits of permission need not be affected.
944 	 * Be consistent with our shm_open there.
945 	 */
946 #ifdef MAC
947 	error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
948 	if (error != 0)
949 		goto out;
950 #endif
951 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid,
952 	    shmfd->shm_gid, VADMIN, active_cred, NULL);
953 	if (error != 0)
954 		goto out;
955 	shmfd->shm_mode = mode & ACCESSPERMS;
956 out:
957 	mtx_unlock(&shm_timestamp_lock);
958 	return (error);
959 }
960 
961 static int
962 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
963     struct thread *td)
964 {
965 	struct shmfd *shmfd;
966 	int error;
967 
968 	error = 0;
969 	shmfd = fp->f_data;
970 	mtx_lock(&shm_timestamp_lock);
971 #ifdef MAC
972 	error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
973 	if (error != 0)
974 		goto out;
975 #endif
976 	if (uid == (uid_t)-1)
977 		uid = shmfd->shm_uid;
978 	if (gid == (gid_t)-1)
979                  gid = shmfd->shm_gid;
980 	if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
981 	    (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
982 	    (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN, 0)))
983 		goto out;
984 	shmfd->shm_uid = uid;
985 	shmfd->shm_gid = gid;
986 out:
987 	mtx_unlock(&shm_timestamp_lock);
988 	return (error);
989 }
990 
991 /*
992  * Helper routines to allow the backing object of a shared memory file
993  * descriptor to be mapped in the kernel.
994  */
995 int
996 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
997 {
998 	struct shmfd *shmfd;
999 	vm_offset_t kva, ofs;
1000 	vm_object_t obj;
1001 	int rv;
1002 
1003 	if (fp->f_type != DTYPE_SHM)
1004 		return (EINVAL);
1005 	shmfd = fp->f_data;
1006 	obj = shmfd->shm_object;
1007 	VM_OBJECT_WLOCK(obj);
1008 	/*
1009 	 * XXXRW: This validation is probably insufficient, and subject to
1010 	 * sign errors.  It should be fixed.
1011 	 */
1012 	if (offset >= shmfd->shm_size ||
1013 	    offset + size > round_page(shmfd->shm_size)) {
1014 		VM_OBJECT_WUNLOCK(obj);
1015 		return (EINVAL);
1016 	}
1017 
1018 	shmfd->shm_kmappings++;
1019 	vm_object_reference_locked(obj);
1020 	VM_OBJECT_WUNLOCK(obj);
1021 
1022 	/* Map the object into the kernel_map and wire it. */
1023 	kva = vm_map_min(kernel_map);
1024 	ofs = offset & PAGE_MASK;
1025 	offset = trunc_page(offset);
1026 	size = round_page(size + ofs);
1027 	rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1028 	    VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1029 	    VM_PROT_READ | VM_PROT_WRITE, 0);
1030 	if (rv == KERN_SUCCESS) {
1031 		rv = vm_map_wire(kernel_map, kva, kva + size,
1032 		    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1033 		if (rv == KERN_SUCCESS) {
1034 			*memp = (void *)(kva + ofs);
1035 			return (0);
1036 		}
1037 		vm_map_remove(kernel_map, kva, kva + size);
1038 	} else
1039 		vm_object_deallocate(obj);
1040 
1041 	/* On failure, drop our mapping reference. */
1042 	VM_OBJECT_WLOCK(obj);
1043 	shmfd->shm_kmappings--;
1044 	VM_OBJECT_WUNLOCK(obj);
1045 
1046 	return (vm_mmap_to_errno(rv));
1047 }
1048 
1049 /*
1050  * We require the caller to unmap the entire entry.  This allows us to
1051  * safely decrement shm_kmappings when a mapping is removed.
1052  */
1053 int
1054 shm_unmap(struct file *fp, void *mem, size_t size)
1055 {
1056 	struct shmfd *shmfd;
1057 	vm_map_entry_t entry;
1058 	vm_offset_t kva, ofs;
1059 	vm_object_t obj;
1060 	vm_pindex_t pindex;
1061 	vm_prot_t prot;
1062 	boolean_t wired;
1063 	vm_map_t map;
1064 	int rv;
1065 
1066 	if (fp->f_type != DTYPE_SHM)
1067 		return (EINVAL);
1068 	shmfd = fp->f_data;
1069 	kva = (vm_offset_t)mem;
1070 	ofs = kva & PAGE_MASK;
1071 	kva = trunc_page(kva);
1072 	size = round_page(size + ofs);
1073 	map = kernel_map;
1074 	rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1075 	    &obj, &pindex, &prot, &wired);
1076 	if (rv != KERN_SUCCESS)
1077 		return (EINVAL);
1078 	if (entry->start != kva || entry->end != kva + size) {
1079 		vm_map_lookup_done(map, entry);
1080 		return (EINVAL);
1081 	}
1082 	vm_map_lookup_done(map, entry);
1083 	if (obj != shmfd->shm_object)
1084 		return (EINVAL);
1085 	vm_map_remove(map, kva, kva + size);
1086 	VM_OBJECT_WLOCK(obj);
1087 	KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1088 	shmfd->shm_kmappings--;
1089 	VM_OBJECT_WUNLOCK(obj);
1090 	return (0);
1091 }
1092 
1093 static int
1094 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1095 {
1096 	const char *path, *pr_path;
1097 	struct shmfd *shmfd;
1098 	size_t pr_pathlen;
1099 
1100 	kif->kf_type = KF_TYPE_SHM;
1101 	shmfd = fp->f_data;
1102 
1103 	mtx_lock(&shm_timestamp_lock);
1104 	kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;	/* XXX */
1105 	mtx_unlock(&shm_timestamp_lock);
1106 	kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1107 	if (shmfd->shm_path != NULL) {
1108 		sx_slock(&shm_dict_lock);
1109 		if (shmfd->shm_path != NULL) {
1110 			path = shmfd->shm_path;
1111 			pr_path = curthread->td_ucred->cr_prison->pr_path;
1112 			if (strcmp(pr_path, "/") != 0) {
1113 				/* Return the jail-rooted pathname. */
1114 				pr_pathlen = strlen(pr_path);
1115 				if (strncmp(path, pr_path, pr_pathlen) == 0 &&
1116 				    path[pr_pathlen] == '/')
1117 					path += pr_pathlen;
1118 			}
1119 			strlcpy(kif->kf_path, path, sizeof(kif->kf_path));
1120 		}
1121 		sx_sunlock(&shm_dict_lock);
1122 	}
1123 	return (0);
1124 }
1125