xref: /freebsd/sys/kern/uipc_shm.c (revision b2d48be1bc7df45ddd13b143a160d0acb5a383c5)
1 /*-
2  * Copyright (c) 2006, 2011 Robert N. M. Watson
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26 
27 /*
28  * Support for shared swap-backed anonymous memory objects via
29  * shm_open(2) and shm_unlink(2).  While most of the implementation is
30  * here, vm_mmap.c contains mapping logic changes.
31  *
32  * TODO:
33  *
34  * (1) Need to export data to a userland tool via a sysctl.  Should ipcs(1)
35  *     and ipcrm(1) be expanded or should new tools to manage both POSIX
36  *     kernel semaphores and POSIX shared memory be written?
37  *
38  * (2) Add support for this file type to fstat(1).
39  *
40  * (3) Resource limits?  Does this need its own resource limits or are the
41  *     existing limits in mmap(2) sufficient?
42  */
43 
44 #include <sys/cdefs.h>
45 __FBSDID("$FreeBSD$");
46 
47 #include "opt_capsicum.h"
48 #include "opt_ktrace.h"
49 
50 #include <sys/param.h>
51 #include <sys/capsicum.h>
52 #include <sys/conf.h>
53 #include <sys/fcntl.h>
54 #include <sys/file.h>
55 #include <sys/filedesc.h>
56 #include <sys/fnv_hash.h>
57 #include <sys/kernel.h>
58 #include <sys/uio.h>
59 #include <sys/signal.h>
60 #include <sys/ktrace.h>
61 #include <sys/lock.h>
62 #include <sys/malloc.h>
63 #include <sys/mman.h>
64 #include <sys/mutex.h>
65 #include <sys/priv.h>
66 #include <sys/proc.h>
67 #include <sys/refcount.h>
68 #include <sys/resourcevar.h>
69 #include <sys/rwlock.h>
70 #include <sys/stat.h>
71 #include <sys/syscallsubr.h>
72 #include <sys/sysctl.h>
73 #include <sys/sysproto.h>
74 #include <sys/systm.h>
75 #include <sys/sx.h>
76 #include <sys/time.h>
77 #include <sys/vnode.h>
78 #include <sys/unistd.h>
79 #include <sys/user.h>
80 
81 #include <security/mac/mac_framework.h>
82 
83 #include <vm/vm.h>
84 #include <vm/vm_param.h>
85 #include <vm/pmap.h>
86 #include <vm/vm_extern.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_kern.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pageout.h>
92 #include <vm/vm_pager.h>
93 #include <vm/swap_pager.h>
94 
95 struct shm_mapping {
96 	char		*sm_path;
97 	Fnv32_t		sm_fnv;
98 	struct shmfd	*sm_shmfd;
99 	LIST_ENTRY(shm_mapping) sm_link;
100 };
101 
102 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor");
103 static LIST_HEAD(, shm_mapping) *shm_dictionary;
104 static struct sx shm_dict_lock;
105 static struct mtx shm_timestamp_lock;
106 static u_long shm_hash;
107 static struct unrhdr *shm_ino_unr;
108 static dev_t shm_dev_ino;
109 
110 #define	SHM_HASH(fnv)	(&shm_dictionary[(fnv) & shm_hash])
111 
112 static int	shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags);
113 static struct shmfd *shm_alloc(struct ucred *ucred, mode_t mode);
114 static void	shm_init(void *arg);
115 static void	shm_drop(struct shmfd *shmfd);
116 static struct shmfd *shm_hold(struct shmfd *shmfd);
117 static void	shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd);
118 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv);
119 static int	shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred);
120 static int	shm_dotruncate(struct shmfd *shmfd, off_t length);
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_stat_t	shm_stat;
126 static fo_close_t	shm_close;
127 static fo_chmod_t	shm_chmod;
128 static fo_chown_t	shm_chown;
129 static fo_seek_t	shm_seek;
130 static fo_fill_kinfo_t	shm_fill_kinfo;
131 static fo_mmap_t	shm_mmap;
132 
133 /* File descriptor operations. */
134 static struct fileops shm_ops = {
135 	.fo_read = shm_read,
136 	.fo_write = shm_write,
137 	.fo_truncate = shm_truncate,
138 	.fo_ioctl = invfo_ioctl,
139 	.fo_poll = invfo_poll,
140 	.fo_kqfilter = invfo_kqfilter,
141 	.fo_stat = shm_stat,
142 	.fo_close = shm_close,
143 	.fo_chmod = shm_chmod,
144 	.fo_chown = shm_chown,
145 	.fo_sendfile = vn_sendfile,
146 	.fo_seek = shm_seek,
147 	.fo_fill_kinfo = shm_fill_kinfo,
148 	.fo_mmap = shm_mmap,
149 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
150 };
151 
152 FEATURE(posix_shm, "POSIX shared memory");
153 
154 static int
155 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio)
156 {
157 	vm_page_t m;
158 	vm_pindex_t idx;
159 	size_t tlen;
160 	int error, offset, rv;
161 
162 	idx = OFF_TO_IDX(uio->uio_offset);
163 	offset = uio->uio_offset & PAGE_MASK;
164 	tlen = MIN(PAGE_SIZE - offset, len);
165 
166 	VM_OBJECT_WLOCK(obj);
167 
168 	/*
169 	 * Read I/O without either a corresponding resident page or swap
170 	 * page: use zero_region.  This is intended to avoid instantiating
171 	 * pages on read from a sparse region.
172 	 */
173 	if (uio->uio_rw == UIO_READ && vm_page_lookup(obj, idx) == NULL &&
174 	    !vm_pager_has_page(obj, idx, NULL, NULL)) {
175 		VM_OBJECT_WUNLOCK(obj);
176 		return (uiomove(__DECONST(void *, zero_region), tlen, uio));
177 	}
178 
179 	/*
180 	 * Parallel reads of the page content from disk are prevented
181 	 * by exclusive busy.
182 	 *
183 	 * Although the tmpfs vnode lock is held here, it is
184 	 * nonetheless safe to sleep waiting for a free page.  The
185 	 * pageout daemon does not need to acquire the tmpfs vnode
186 	 * lock to page out tobj's pages because tobj is a OBJT_SWAP
187 	 * type object.
188 	 */
189 	m = vm_page_grab(obj, idx, VM_ALLOC_NORMAL);
190 	if (m->valid != VM_PAGE_BITS_ALL) {
191 		if (vm_pager_has_page(obj, idx, NULL, NULL)) {
192 			rv = vm_pager_get_pages(obj, &m, 1, 0);
193 			if (rv != VM_PAGER_OK) {
194 				printf(
195 	    "uiomove_object: vm_obj %p idx %jd valid %x pager error %d\n",
196 				    obj, idx, m->valid, rv);
197 				vm_page_lock(m);
198 				vm_page_free(m);
199 				vm_page_unlock(m);
200 				VM_OBJECT_WUNLOCK(obj);
201 				return (EIO);
202 			}
203 		} else
204 			vm_page_zero_invalid(m, TRUE);
205 	}
206 	vm_page_xunbusy(m);
207 	vm_page_lock(m);
208 	vm_page_hold(m);
209 	if (m->queue == PQ_NONE) {
210 		vm_page_deactivate(m);
211 	} else {
212 		/* Requeue to maintain LRU ordering. */
213 		vm_page_requeue(m);
214 	}
215 	vm_page_unlock(m);
216 	VM_OBJECT_WUNLOCK(obj);
217 	error = uiomove_fromphys(&m, offset, tlen, uio);
218 	if (uio->uio_rw == UIO_WRITE && error == 0) {
219 		VM_OBJECT_WLOCK(obj);
220 		vm_page_dirty(m);
221 		vm_pager_page_unswapped(m);
222 		VM_OBJECT_WUNLOCK(obj);
223 	}
224 	vm_page_lock(m);
225 	vm_page_unhold(m);
226 	vm_page_unlock(m);
227 
228 	return (error);
229 }
230 
231 int
232 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
233 {
234 	ssize_t resid;
235 	size_t len;
236 	int error;
237 
238 	error = 0;
239 	while ((resid = uio->uio_resid) > 0) {
240 		if (obj_size <= uio->uio_offset)
241 			break;
242 		len = MIN(obj_size - uio->uio_offset, resid);
243 		if (len == 0)
244 			break;
245 		error = uiomove_object_page(obj, len, uio);
246 		if (error != 0 || resid == uio->uio_resid)
247 			break;
248 	}
249 	return (error);
250 }
251 
252 static int
253 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
254 {
255 	struct shmfd *shmfd;
256 	off_t foffset;
257 	int error;
258 
259 	shmfd = fp->f_data;
260 	foffset = foffset_lock(fp, 0);
261 	error = 0;
262 	switch (whence) {
263 	case L_INCR:
264 		if (foffset < 0 ||
265 		    (offset > 0 && foffset > OFF_MAX - offset)) {
266 			error = EOVERFLOW;
267 			break;
268 		}
269 		offset += foffset;
270 		break;
271 	case L_XTND:
272 		if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
273 			error = EOVERFLOW;
274 			break;
275 		}
276 		offset += shmfd->shm_size;
277 		break;
278 	case L_SET:
279 		break;
280 	default:
281 		error = EINVAL;
282 	}
283 	if (error == 0) {
284 		if (offset < 0 || offset > shmfd->shm_size)
285 			error = EINVAL;
286 		else
287 			td->td_uretoff.tdu_off = offset;
288 	}
289 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
290 	return (error);
291 }
292 
293 static int
294 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
295     int flags, struct thread *td)
296 {
297 	struct shmfd *shmfd;
298 	void *rl_cookie;
299 	int error;
300 
301 	shmfd = fp->f_data;
302 	foffset_lock_uio(fp, uio, flags);
303 	rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset,
304 	    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
305 #ifdef MAC
306 	error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
307 	if (error)
308 		return (error);
309 #endif
310 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
311 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
312 	foffset_unlock_uio(fp, uio, flags);
313 	return (error);
314 }
315 
316 static int
317 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
318     int flags, struct thread *td)
319 {
320 	struct shmfd *shmfd;
321 	void *rl_cookie;
322 	int error;
323 
324 	shmfd = fp->f_data;
325 #ifdef MAC
326 	error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
327 	if (error)
328 		return (error);
329 #endif
330 	foffset_lock_uio(fp, uio, flags);
331 	if ((flags & FOF_OFFSET) == 0) {
332 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
333 		    &shmfd->shm_mtx);
334 	} else {
335 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset,
336 		    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
337 	}
338 
339 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
340 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
341 	foffset_unlock_uio(fp, uio, flags);
342 	return (error);
343 }
344 
345 static int
346 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
347     struct thread *td)
348 {
349 	struct shmfd *shmfd;
350 #ifdef MAC
351 	int error;
352 #endif
353 
354 	shmfd = fp->f_data;
355 #ifdef MAC
356 	error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
357 	if (error)
358 		return (error);
359 #endif
360 	return (shm_dotruncate(shmfd, length));
361 }
362 
363 static int
364 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
365     struct thread *td)
366 {
367 	struct shmfd *shmfd;
368 #ifdef MAC
369 	int error;
370 #endif
371 
372 	shmfd = fp->f_data;
373 
374 #ifdef MAC
375 	error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
376 	if (error)
377 		return (error);
378 #endif
379 
380 	/*
381 	 * Attempt to return sanish values for fstat() on a memory file
382 	 * descriptor.
383 	 */
384 	bzero(sb, sizeof(*sb));
385 	sb->st_blksize = PAGE_SIZE;
386 	sb->st_size = shmfd->shm_size;
387 	sb->st_blocks = (sb->st_size + sb->st_blksize - 1) / sb->st_blksize;
388 	mtx_lock(&shm_timestamp_lock);
389 	sb->st_atim = shmfd->shm_atime;
390 	sb->st_ctim = shmfd->shm_ctime;
391 	sb->st_mtim = shmfd->shm_mtime;
392 	sb->st_birthtim = shmfd->shm_birthtime;
393 	sb->st_mode = S_IFREG | shmfd->shm_mode;		/* XXX */
394 	sb->st_uid = shmfd->shm_uid;
395 	sb->st_gid = shmfd->shm_gid;
396 	mtx_unlock(&shm_timestamp_lock);
397 	sb->st_dev = shm_dev_ino;
398 	sb->st_ino = shmfd->shm_ino;
399 
400 	return (0);
401 }
402 
403 static int
404 shm_close(struct file *fp, struct thread *td)
405 {
406 	struct shmfd *shmfd;
407 
408 	shmfd = fp->f_data;
409 	fp->f_data = NULL;
410 	shm_drop(shmfd);
411 
412 	return (0);
413 }
414 
415 static int
416 shm_dotruncate(struct shmfd *shmfd, off_t length)
417 {
418 	vm_object_t object;
419 	vm_page_t m;
420 	vm_pindex_t idx, nobjsize;
421 	vm_ooffset_t delta;
422 	int base, rv;
423 
424 	object = shmfd->shm_object;
425 	VM_OBJECT_WLOCK(object);
426 	if (length == shmfd->shm_size) {
427 		VM_OBJECT_WUNLOCK(object);
428 		return (0);
429 	}
430 	nobjsize = OFF_TO_IDX(length + PAGE_MASK);
431 
432 	/* Are we shrinking?  If so, trim the end. */
433 	if (length < shmfd->shm_size) {
434 		/*
435 		 * Disallow any requests to shrink the size if this
436 		 * object is mapped into the kernel.
437 		 */
438 		if (shmfd->shm_kmappings > 0) {
439 			VM_OBJECT_WUNLOCK(object);
440 			return (EBUSY);
441 		}
442 
443 		/*
444 		 * Zero the truncated part of the last page.
445 		 */
446 		base = length & PAGE_MASK;
447 		if (base != 0) {
448 			idx = OFF_TO_IDX(length);
449 retry:
450 			m = vm_page_lookup(object, idx);
451 			if (m != NULL) {
452 				if (vm_page_sleep_if_busy(m, "shmtrc"))
453 					goto retry;
454 			} else if (vm_pager_has_page(object, idx, NULL, NULL)) {
455 				m = vm_page_alloc(object, idx, VM_ALLOC_NORMAL);
456 				if (m == NULL) {
457 					VM_OBJECT_WUNLOCK(object);
458 					VM_WAIT;
459 					VM_OBJECT_WLOCK(object);
460 					goto retry;
461 				} else if (m->valid != VM_PAGE_BITS_ALL)
462 					rv = vm_pager_get_pages(object, &m, 1,
463 					    0);
464 				else
465 					/* A cached page was reactivated. */
466 					rv = VM_PAGER_OK;
467 				vm_page_lock(m);
468 				if (rv == VM_PAGER_OK) {
469 					vm_page_deactivate(m);
470 					vm_page_unlock(m);
471 					vm_page_xunbusy(m);
472 				} else {
473 					vm_page_free(m);
474 					vm_page_unlock(m);
475 					VM_OBJECT_WUNLOCK(object);
476 					return (EIO);
477 				}
478 			}
479 			if (m != NULL) {
480 				pmap_zero_page_area(m, base, PAGE_SIZE - base);
481 				KASSERT(m->valid == VM_PAGE_BITS_ALL,
482 				    ("shm_dotruncate: page %p is invalid", m));
483 				vm_page_dirty(m);
484 				vm_pager_page_unswapped(m);
485 			}
486 		}
487 		delta = ptoa(object->size - nobjsize);
488 
489 		/* Toss in memory pages. */
490 		if (nobjsize < object->size)
491 			vm_object_page_remove(object, nobjsize, object->size,
492 			    0);
493 
494 		/* Toss pages from swap. */
495 		if (object->type == OBJT_SWAP)
496 			swap_pager_freespace(object, nobjsize, delta);
497 
498 		/* Free the swap accounted for shm */
499 		swap_release_by_cred(delta, object->cred);
500 		object->charge -= delta;
501 	} else {
502 		/* Attempt to reserve the swap */
503 		delta = ptoa(nobjsize - object->size);
504 		if (!swap_reserve_by_cred(delta, object->cred)) {
505 			VM_OBJECT_WUNLOCK(object);
506 			return (ENOMEM);
507 		}
508 		object->charge += delta;
509 	}
510 	shmfd->shm_size = length;
511 	mtx_lock(&shm_timestamp_lock);
512 	vfs_timestamp(&shmfd->shm_ctime);
513 	shmfd->shm_mtime = shmfd->shm_ctime;
514 	mtx_unlock(&shm_timestamp_lock);
515 	object->size = nobjsize;
516 	VM_OBJECT_WUNLOCK(object);
517 	return (0);
518 }
519 
520 /*
521  * shmfd object management including creation and reference counting
522  * routines.
523  */
524 static struct shmfd *
525 shm_alloc(struct ucred *ucred, mode_t mode)
526 {
527 	struct shmfd *shmfd;
528 	int ino;
529 
530 	shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
531 	shmfd->shm_size = 0;
532 	shmfd->shm_uid = ucred->cr_uid;
533 	shmfd->shm_gid = ucred->cr_gid;
534 	shmfd->shm_mode = mode;
535 	shmfd->shm_object = vm_pager_allocate(OBJT_DEFAULT, NULL,
536 	    shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
537 	KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
538 	shmfd->shm_object->pg_color = 0;
539 	VM_OBJECT_WLOCK(shmfd->shm_object);
540 	vm_object_clear_flag(shmfd->shm_object, OBJ_ONEMAPPING);
541 	vm_object_set_flag(shmfd->shm_object, OBJ_COLORED | OBJ_NOSPLIT);
542 	VM_OBJECT_WUNLOCK(shmfd->shm_object);
543 	vfs_timestamp(&shmfd->shm_birthtime);
544 	shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
545 	    shmfd->shm_birthtime;
546 	ino = alloc_unr(shm_ino_unr);
547 	if (ino == -1)
548 		shmfd->shm_ino = 0;
549 	else
550 		shmfd->shm_ino = ino;
551 	refcount_init(&shmfd->shm_refs, 1);
552 	mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
553 	rangelock_init(&shmfd->shm_rl);
554 #ifdef MAC
555 	mac_posixshm_init(shmfd);
556 	mac_posixshm_create(ucred, shmfd);
557 #endif
558 
559 	return (shmfd);
560 }
561 
562 static struct shmfd *
563 shm_hold(struct shmfd *shmfd)
564 {
565 
566 	refcount_acquire(&shmfd->shm_refs);
567 	return (shmfd);
568 }
569 
570 static void
571 shm_drop(struct shmfd *shmfd)
572 {
573 
574 	if (refcount_release(&shmfd->shm_refs)) {
575 #ifdef MAC
576 		mac_posixshm_destroy(shmfd);
577 #endif
578 		rangelock_destroy(&shmfd->shm_rl);
579 		mtx_destroy(&shmfd->shm_mtx);
580 		vm_object_deallocate(shmfd->shm_object);
581 		if (shmfd->shm_ino != 0)
582 			free_unr(shm_ino_unr, shmfd->shm_ino);
583 		free(shmfd, M_SHMFD);
584 	}
585 }
586 
587 /*
588  * Determine if the credentials have sufficient permissions for a
589  * specified combination of FREAD and FWRITE.
590  */
591 static int
592 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
593 {
594 	accmode_t accmode;
595 	int error;
596 
597 	accmode = 0;
598 	if (flags & FREAD)
599 		accmode |= VREAD;
600 	if (flags & FWRITE)
601 		accmode |= VWRITE;
602 	mtx_lock(&shm_timestamp_lock);
603 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
604 	    accmode, ucred, NULL);
605 	mtx_unlock(&shm_timestamp_lock);
606 	return (error);
607 }
608 
609 /*
610  * Dictionary management.  We maintain an in-kernel dictionary to map
611  * paths to shmfd objects.  We use the FNV hash on the path to store
612  * the mappings in a hash table.
613  */
614 static void
615 shm_init(void *arg)
616 {
617 
618 	mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
619 	sx_init(&shm_dict_lock, "shm dictionary");
620 	shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
621 	shm_ino_unr = new_unrhdr(1, INT32_MAX, NULL);
622 	KASSERT(shm_ino_unr != NULL, ("shm fake inodes not initialized"));
623 	shm_dev_ino = devfs_alloc_cdp_inode();
624 	KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
625 }
626 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
627 
628 static struct shmfd *
629 shm_lookup(char *path, Fnv32_t fnv)
630 {
631 	struct shm_mapping *map;
632 
633 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
634 		if (map->sm_fnv != fnv)
635 			continue;
636 		if (strcmp(map->sm_path, path) == 0)
637 			return (map->sm_shmfd);
638 	}
639 
640 	return (NULL);
641 }
642 
643 static void
644 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
645 {
646 	struct shm_mapping *map;
647 
648 	map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
649 	map->sm_path = path;
650 	map->sm_fnv = fnv;
651 	map->sm_shmfd = shm_hold(shmfd);
652 	shmfd->shm_path = path;
653 	LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
654 }
655 
656 static int
657 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
658 {
659 	struct shm_mapping *map;
660 	int error;
661 
662 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
663 		if (map->sm_fnv != fnv)
664 			continue;
665 		if (strcmp(map->sm_path, path) == 0) {
666 #ifdef MAC
667 			error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
668 			if (error)
669 				return (error);
670 #endif
671 			error = shm_access(map->sm_shmfd, ucred,
672 			    FREAD | FWRITE);
673 			if (error)
674 				return (error);
675 			map->sm_shmfd->shm_path = NULL;
676 			LIST_REMOVE(map, sm_link);
677 			shm_drop(map->sm_shmfd);
678 			free(map->sm_path, M_SHMFD);
679 			free(map, M_SHMFD);
680 			return (0);
681 		}
682 	}
683 
684 	return (ENOENT);
685 }
686 
687 int
688 kern_shm_open(struct thread *td, const char *userpath, int flags, mode_t mode,
689     struct filecaps *fcaps)
690 {
691 	struct filedesc *fdp;
692 	struct shmfd *shmfd;
693 	struct file *fp;
694 	char *path;
695 	Fnv32_t fnv;
696 	mode_t cmode;
697 	int fd, error;
698 
699 #ifdef CAPABILITY_MODE
700 	/*
701 	 * shm_open(2) is only allowed for anonymous objects.
702 	 */
703 	if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON))
704 		return (ECAPMODE);
705 #endif
706 
707 	if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
708 		return (EINVAL);
709 
710 	if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
711 		return (EINVAL);
712 
713 	fdp = td->td_proc->p_fd;
714 	cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS;
715 
716 	error = falloc_caps(td, &fp, &fd, O_CLOEXEC, fcaps);
717 	if (error)
718 		return (error);
719 
720 	/* A SHM_ANON path pointer creates an anonymous object. */
721 	if (userpath == SHM_ANON) {
722 		/* A read-only anonymous object is pointless. */
723 		if ((flags & O_ACCMODE) == O_RDONLY) {
724 			fdclose(td, fp, fd);
725 			fdrop(fp, td);
726 			return (EINVAL);
727 		}
728 		shmfd = shm_alloc(td->td_ucred, cmode);
729 	} else {
730 		path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
731 		error = copyinstr(userpath, path, MAXPATHLEN, NULL);
732 #ifdef KTRACE
733 		if (error == 0 && KTRPOINT(curthread, KTR_NAMEI))
734 			ktrnamei(path);
735 #endif
736 		/* Require paths to start with a '/' character. */
737 		if (error == 0 && path[0] != '/')
738 			error = EINVAL;
739 		if (error) {
740 			fdclose(td, fp, fd);
741 			fdrop(fp, td);
742 			free(path, M_SHMFD);
743 			return (error);
744 		}
745 
746 		fnv = fnv_32_str(path, FNV1_32_INIT);
747 		sx_xlock(&shm_dict_lock);
748 		shmfd = shm_lookup(path, fnv);
749 		if (shmfd == NULL) {
750 			/* Object does not yet exist, create it if requested. */
751 			if (flags & O_CREAT) {
752 #ifdef MAC
753 				error = mac_posixshm_check_create(td->td_ucred,
754 				    path);
755 				if (error == 0) {
756 #endif
757 					shmfd = shm_alloc(td->td_ucred, cmode);
758 					shm_insert(path, fnv, shmfd);
759 #ifdef MAC
760 				}
761 #endif
762 			} else {
763 				free(path, M_SHMFD);
764 				error = ENOENT;
765 			}
766 		} else {
767 			/*
768 			 * Object already exists, obtain a new
769 			 * reference if requested and permitted.
770 			 */
771 			free(path, M_SHMFD);
772 			if ((flags & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
773 				error = EEXIST;
774 			else {
775 #ifdef MAC
776 				error = mac_posixshm_check_open(td->td_ucred,
777 				    shmfd, FFLAGS(flags & O_ACCMODE));
778 				if (error == 0)
779 #endif
780 				error = shm_access(shmfd, td->td_ucred,
781 				    FFLAGS(flags & O_ACCMODE));
782 			}
783 
784 			/*
785 			 * Truncate the file back to zero length if
786 			 * O_TRUNC was specified and the object was
787 			 * opened with read/write.
788 			 */
789 			if (error == 0 &&
790 			    (flags & (O_ACCMODE | O_TRUNC)) ==
791 			    (O_RDWR | O_TRUNC)) {
792 #ifdef MAC
793 				error = mac_posixshm_check_truncate(
794 					td->td_ucred, fp->f_cred, shmfd);
795 				if (error == 0)
796 #endif
797 					shm_dotruncate(shmfd, 0);
798 			}
799 			if (error == 0)
800 				shm_hold(shmfd);
801 		}
802 		sx_xunlock(&shm_dict_lock);
803 
804 		if (error) {
805 			fdclose(td, fp, fd);
806 			fdrop(fp, td);
807 			return (error);
808 		}
809 	}
810 
811 	finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
812 
813 	td->td_retval[0] = fd;
814 	fdrop(fp, td);
815 
816 	return (0);
817 }
818 
819 /* System calls. */
820 int
821 sys_shm_open(struct thread *td, struct shm_open_args *uap)
822 {
823 
824 	return (kern_shm_open(td, uap->path, uap->flags, uap->mode, NULL));
825 }
826 
827 int
828 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
829 {
830 	char *path;
831 	Fnv32_t fnv;
832 	int error;
833 
834 	path = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
835 	error = copyinstr(uap->path, path, MAXPATHLEN, NULL);
836 	if (error) {
837 		free(path, M_TEMP);
838 		return (error);
839 	}
840 #ifdef KTRACE
841 	if (KTRPOINT(curthread, KTR_NAMEI))
842 		ktrnamei(path);
843 #endif
844 	fnv = fnv_32_str(path, FNV1_32_INIT);
845 	sx_xlock(&shm_dict_lock);
846 	error = shm_remove(path, fnv, td->td_ucred);
847 	sx_xunlock(&shm_dict_lock);
848 	free(path, M_TEMP);
849 
850 	return (error);
851 }
852 
853 int
854 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
855     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
856     vm_ooffset_t foff, struct thread *td)
857 {
858 	struct shmfd *shmfd;
859 	vm_prot_t maxprot;
860 	int error;
861 
862 	shmfd = fp->f_data;
863 	maxprot = VM_PROT_NONE;
864 
865 	/* FREAD should always be set. */
866 	if ((fp->f_flag & FREAD) != 0)
867 		maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
868 	if ((fp->f_flag & FWRITE) != 0)
869 		maxprot |= VM_PROT_WRITE;
870 
871 	/* Don't permit shared writable mappings on read-only descriptors. */
872 	if ((flags & MAP_SHARED) != 0 &&
873 	    (maxprot & VM_PROT_WRITE) == 0 &&
874 	    (prot & VM_PROT_WRITE) != 0)
875 		return (EACCES);
876 	maxprot &= cap_maxprot;
877 
878 #ifdef MAC
879 	error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
880 	if (error != 0)
881 		return (error);
882 #endif
883 
884 	/*
885 	 * XXXRW: This validation is probably insufficient, and subject to
886 	 * sign errors.  It should be fixed.
887 	 */
888 	if (foff >= shmfd->shm_size ||
889 	    foff + objsize > round_page(shmfd->shm_size))
890 		return (EINVAL);
891 
892 	mtx_lock(&shm_timestamp_lock);
893 	vfs_timestamp(&shmfd->shm_atime);
894 	mtx_unlock(&shm_timestamp_lock);
895 	vm_object_reference(shmfd->shm_object);
896 
897 	error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
898 	    shmfd->shm_object, foff, FALSE, td);
899 	if (error != 0)
900 		vm_object_deallocate(shmfd->shm_object);
901 	return (0);
902 }
903 
904 static int
905 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
906     struct thread *td)
907 {
908 	struct shmfd *shmfd;
909 	int error;
910 
911 	error = 0;
912 	shmfd = fp->f_data;
913 	mtx_lock(&shm_timestamp_lock);
914 	/*
915 	 * SUSv4 says that x bits of permission need not be affected.
916 	 * Be consistent with our shm_open there.
917 	 */
918 #ifdef MAC
919 	error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
920 	if (error != 0)
921 		goto out;
922 #endif
923 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid,
924 	    shmfd->shm_gid, VADMIN, active_cred, NULL);
925 	if (error != 0)
926 		goto out;
927 	shmfd->shm_mode = mode & ACCESSPERMS;
928 out:
929 	mtx_unlock(&shm_timestamp_lock);
930 	return (error);
931 }
932 
933 static int
934 shm_chown(struct file *fp, uid_t uid, gid_t gid, 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 #ifdef MAC
944 	error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
945 	if (error != 0)
946 		goto out;
947 #endif
948 	if (uid == (uid_t)-1)
949 		uid = shmfd->shm_uid;
950 	if (gid == (gid_t)-1)
951                  gid = shmfd->shm_gid;
952 	if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
953 	    (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
954 	    (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN, 0)))
955 		goto out;
956 	shmfd->shm_uid = uid;
957 	shmfd->shm_gid = gid;
958 out:
959 	mtx_unlock(&shm_timestamp_lock);
960 	return (error);
961 }
962 
963 /*
964  * Helper routines to allow the backing object of a shared memory file
965  * descriptor to be mapped in the kernel.
966  */
967 int
968 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
969 {
970 	struct shmfd *shmfd;
971 	vm_offset_t kva, ofs;
972 	vm_object_t obj;
973 	int rv;
974 
975 	if (fp->f_type != DTYPE_SHM)
976 		return (EINVAL);
977 	shmfd = fp->f_data;
978 	obj = shmfd->shm_object;
979 	VM_OBJECT_WLOCK(obj);
980 	/*
981 	 * XXXRW: This validation is probably insufficient, and subject to
982 	 * sign errors.  It should be fixed.
983 	 */
984 	if (offset >= shmfd->shm_size ||
985 	    offset + size > round_page(shmfd->shm_size)) {
986 		VM_OBJECT_WUNLOCK(obj);
987 		return (EINVAL);
988 	}
989 
990 	shmfd->shm_kmappings++;
991 	vm_object_reference_locked(obj);
992 	VM_OBJECT_WUNLOCK(obj);
993 
994 	/* Map the object into the kernel_map and wire it. */
995 	kva = vm_map_min(kernel_map);
996 	ofs = offset & PAGE_MASK;
997 	offset = trunc_page(offset);
998 	size = round_page(size + ofs);
999 	rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1000 	    VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1001 	    VM_PROT_READ | VM_PROT_WRITE, 0);
1002 	if (rv == KERN_SUCCESS) {
1003 		rv = vm_map_wire(kernel_map, kva, kva + size,
1004 		    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1005 		if (rv == KERN_SUCCESS) {
1006 			*memp = (void *)(kva + ofs);
1007 			return (0);
1008 		}
1009 		vm_map_remove(kernel_map, kva, kva + size);
1010 	} else
1011 		vm_object_deallocate(obj);
1012 
1013 	/* On failure, drop our mapping reference. */
1014 	VM_OBJECT_WLOCK(obj);
1015 	shmfd->shm_kmappings--;
1016 	VM_OBJECT_WUNLOCK(obj);
1017 
1018 	return (vm_mmap_to_errno(rv));
1019 }
1020 
1021 /*
1022  * We require the caller to unmap the entire entry.  This allows us to
1023  * safely decrement shm_kmappings when a mapping is removed.
1024  */
1025 int
1026 shm_unmap(struct file *fp, void *mem, size_t size)
1027 {
1028 	struct shmfd *shmfd;
1029 	vm_map_entry_t entry;
1030 	vm_offset_t kva, ofs;
1031 	vm_object_t obj;
1032 	vm_pindex_t pindex;
1033 	vm_prot_t prot;
1034 	boolean_t wired;
1035 	vm_map_t map;
1036 	int rv;
1037 
1038 	if (fp->f_type != DTYPE_SHM)
1039 		return (EINVAL);
1040 	shmfd = fp->f_data;
1041 	kva = (vm_offset_t)mem;
1042 	ofs = kva & PAGE_MASK;
1043 	kva = trunc_page(kva);
1044 	size = round_page(size + ofs);
1045 	map = kernel_map;
1046 	rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1047 	    &obj, &pindex, &prot, &wired);
1048 	if (rv != KERN_SUCCESS)
1049 		return (EINVAL);
1050 	if (entry->start != kva || entry->end != kva + size) {
1051 		vm_map_lookup_done(map, entry);
1052 		return (EINVAL);
1053 	}
1054 	vm_map_lookup_done(map, entry);
1055 	if (obj != shmfd->shm_object)
1056 		return (EINVAL);
1057 	vm_map_remove(map, kva, kva + size);
1058 	VM_OBJECT_WLOCK(obj);
1059 	KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1060 	shmfd->shm_kmappings--;
1061 	VM_OBJECT_WUNLOCK(obj);
1062 	return (0);
1063 }
1064 
1065 static int
1066 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1067 {
1068 	struct shmfd *shmfd;
1069 
1070 	kif->kf_type = KF_TYPE_SHM;
1071 	shmfd = fp->f_data;
1072 
1073 	mtx_lock(&shm_timestamp_lock);
1074 	kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;	/* XXX */
1075 	mtx_unlock(&shm_timestamp_lock);
1076 	kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1077 	if (shmfd->shm_path != NULL) {
1078 		sx_slock(&shm_dict_lock);
1079 		if (shmfd->shm_path != NULL)
1080 			strlcpy(kif->kf_path, shmfd->shm_path,
1081 			    sizeof(kif->kf_path));
1082 		sx_sunlock(&shm_dict_lock);
1083 	}
1084 	return (0);
1085 }
1086