xref: /freebsd/sys/kern/uipc_shm.c (revision f4b37ed0f8b307b1f3f0f630ca725d68f1dff30d)
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/sysctl.h>
72 #include <sys/sysproto.h>
73 #include <sys/systm.h>
74 #include <sys/sx.h>
75 #include <sys/time.h>
76 #include <sys/vnode.h>
77 #include <sys/unistd.h>
78 #include <sys/user.h>
79 
80 #include <security/mac/mac_framework.h>
81 
82 #include <vm/vm.h>
83 #include <vm/vm_param.h>
84 #include <vm/pmap.h>
85 #include <vm/vm_extern.h>
86 #include <vm/vm_map.h>
87 #include <vm/vm_kern.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_pageout.h>
91 #include <vm/vm_pager.h>
92 #include <vm/swap_pager.h>
93 
94 struct shm_mapping {
95 	char		*sm_path;
96 	Fnv32_t		sm_fnv;
97 	struct shmfd	*sm_shmfd;
98 	LIST_ENTRY(shm_mapping) sm_link;
99 };
100 
101 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor");
102 static LIST_HEAD(, shm_mapping) *shm_dictionary;
103 static struct sx shm_dict_lock;
104 static struct mtx shm_timestamp_lock;
105 static u_long shm_hash;
106 static struct unrhdr *shm_ino_unr;
107 static dev_t shm_dev_ino;
108 
109 #define	SHM_HASH(fnv)	(&shm_dictionary[(fnv) & shm_hash])
110 
111 static int	shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags);
112 static struct shmfd *shm_alloc(struct ucred *ucred, mode_t mode);
113 static void	shm_init(void *arg);
114 static void	shm_drop(struct shmfd *shmfd);
115 static struct shmfd *shm_hold(struct shmfd *shmfd);
116 static void	shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd);
117 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv);
118 static int	shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred);
119 static int	shm_dotruncate(struct shmfd *shmfd, off_t length);
120 
121 static fo_rdwr_t	shm_read;
122 static fo_rdwr_t	shm_write;
123 static fo_truncate_t	shm_truncate;
124 static fo_stat_t	shm_stat;
125 static fo_close_t	shm_close;
126 static fo_chmod_t	shm_chmod;
127 static fo_chown_t	shm_chown;
128 static fo_seek_t	shm_seek;
129 static fo_fill_kinfo_t	shm_fill_kinfo;
130 static fo_mmap_t	shm_mmap;
131 
132 /* File descriptor operations. */
133 static struct fileops shm_ops = {
134 	.fo_read = shm_read,
135 	.fo_write = shm_write,
136 	.fo_truncate = shm_truncate,
137 	.fo_ioctl = invfo_ioctl,
138 	.fo_poll = invfo_poll,
139 	.fo_kqfilter = invfo_kqfilter,
140 	.fo_stat = shm_stat,
141 	.fo_close = shm_close,
142 	.fo_chmod = shm_chmod,
143 	.fo_chown = shm_chown,
144 	.fo_sendfile = vn_sendfile,
145 	.fo_seek = shm_seek,
146 	.fo_fill_kinfo = shm_fill_kinfo,
147 	.fo_mmap = shm_mmap,
148 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
149 };
150 
151 FEATURE(posix_shm, "POSIX shared memory");
152 
153 static int
154 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio)
155 {
156 	vm_page_t m;
157 	vm_pindex_t idx;
158 	size_t tlen;
159 	int error, offset, rv;
160 
161 	idx = OFF_TO_IDX(uio->uio_offset);
162 	offset = uio->uio_offset & PAGE_MASK;
163 	tlen = MIN(PAGE_SIZE - offset, len);
164 
165 	VM_OBJECT_WLOCK(obj);
166 
167 	/*
168 	 * Read I/O without either a corresponding resident page or swap
169 	 * page: use zero_region.  This is intended to avoid instantiating
170 	 * pages on read from a sparse region.
171 	 */
172 	if (uio->uio_rw == UIO_READ && vm_page_lookup(obj, idx) == NULL &&
173 	    !vm_pager_has_page(obj, idx, NULL, NULL)) {
174 		VM_OBJECT_WUNLOCK(obj);
175 		return (uiomove(__DECONST(void *, zero_region), tlen, uio));
176 	}
177 
178 	/*
179 	 * Parallel reads of the page content from disk are prevented
180 	 * by exclusive busy.
181 	 *
182 	 * Although the tmpfs vnode lock is held here, it is
183 	 * nonetheless safe to sleep waiting for a free page.  The
184 	 * pageout daemon does not need to acquire the tmpfs vnode
185 	 * lock to page out tobj's pages because tobj is a OBJT_SWAP
186 	 * type object.
187 	 */
188 	m = vm_page_grab(obj, idx, VM_ALLOC_NORMAL);
189 	if (m->valid != VM_PAGE_BITS_ALL) {
190 		if (vm_pager_has_page(obj, idx, NULL, NULL)) {
191 			rv = vm_pager_get_pages(obj, &m, 1, 0);
192 			if (rv != VM_PAGER_OK) {
193 				printf(
194 	    "uiomove_object: vm_obj %p idx %jd valid %x pager error %d\n",
195 				    obj, idx, m->valid, rv);
196 				vm_page_lock(m);
197 				vm_page_free(m);
198 				vm_page_unlock(m);
199 				VM_OBJECT_WUNLOCK(obj);
200 				return (EIO);
201 			}
202 		} else
203 			vm_page_zero_invalid(m, TRUE);
204 	}
205 	vm_page_xunbusy(m);
206 	vm_page_lock(m);
207 	vm_page_hold(m);
208 	if (m->queue == PQ_NONE) {
209 		vm_page_deactivate(m);
210 	} else {
211 		/* Requeue to maintain LRU ordering. */
212 		vm_page_requeue(m);
213 	}
214 	vm_page_unlock(m);
215 	VM_OBJECT_WUNLOCK(obj);
216 	error = uiomove_fromphys(&m, offset, tlen, uio);
217 	if (uio->uio_rw == UIO_WRITE && error == 0) {
218 		VM_OBJECT_WLOCK(obj);
219 		vm_page_dirty(m);
220 		vm_pager_page_unswapped(m);
221 		VM_OBJECT_WUNLOCK(obj);
222 	}
223 	vm_page_lock(m);
224 	vm_page_unhold(m);
225 	vm_page_unlock(m);
226 
227 	return (error);
228 }
229 
230 int
231 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
232 {
233 	ssize_t resid;
234 	size_t len;
235 	int error;
236 
237 	error = 0;
238 	while ((resid = uio->uio_resid) > 0) {
239 		if (obj_size <= uio->uio_offset)
240 			break;
241 		len = MIN(obj_size - uio->uio_offset, resid);
242 		if (len == 0)
243 			break;
244 		error = uiomove_object_page(obj, len, uio);
245 		if (error != 0 || resid == uio->uio_resid)
246 			break;
247 	}
248 	return (error);
249 }
250 
251 static int
252 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
253 {
254 	struct shmfd *shmfd;
255 	off_t foffset;
256 	int error;
257 
258 	shmfd = fp->f_data;
259 	foffset = foffset_lock(fp, 0);
260 	error = 0;
261 	switch (whence) {
262 	case L_INCR:
263 		if (foffset < 0 ||
264 		    (offset > 0 && foffset > OFF_MAX - offset)) {
265 			error = EOVERFLOW;
266 			break;
267 		}
268 		offset += foffset;
269 		break;
270 	case L_XTND:
271 		if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
272 			error = EOVERFLOW;
273 			break;
274 		}
275 		offset += shmfd->shm_size;
276 		break;
277 	case L_SET:
278 		break;
279 	default:
280 		error = EINVAL;
281 	}
282 	if (error == 0) {
283 		if (offset < 0 || offset > shmfd->shm_size)
284 			error = EINVAL;
285 		else
286 			td->td_uretoff.tdu_off = offset;
287 	}
288 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
289 	return (error);
290 }
291 
292 static int
293 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
294     int flags, struct thread *td)
295 {
296 	struct shmfd *shmfd;
297 	void *rl_cookie;
298 	int error;
299 
300 	shmfd = fp->f_data;
301 	foffset_lock_uio(fp, uio, flags);
302 	rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset,
303 	    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
304 #ifdef MAC
305 	error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
306 	if (error)
307 		return (error);
308 #endif
309 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
310 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
311 	foffset_unlock_uio(fp, uio, flags);
312 	return (error);
313 }
314 
315 static int
316 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
317     int flags, struct thread *td)
318 {
319 	struct shmfd *shmfd;
320 	void *rl_cookie;
321 	int error;
322 
323 	shmfd = fp->f_data;
324 #ifdef MAC
325 	error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
326 	if (error)
327 		return (error);
328 #endif
329 	foffset_lock_uio(fp, uio, flags);
330 	if ((flags & FOF_OFFSET) == 0) {
331 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
332 		    &shmfd->shm_mtx);
333 	} else {
334 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset,
335 		    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
336 	}
337 
338 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
339 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
340 	foffset_unlock_uio(fp, uio, flags);
341 	return (error);
342 }
343 
344 static int
345 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
346     struct thread *td)
347 {
348 	struct shmfd *shmfd;
349 #ifdef MAC
350 	int error;
351 #endif
352 
353 	shmfd = fp->f_data;
354 #ifdef MAC
355 	error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
356 	if (error)
357 		return (error);
358 #endif
359 	return (shm_dotruncate(shmfd, length));
360 }
361 
362 static int
363 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
364     struct thread *td)
365 {
366 	struct shmfd *shmfd;
367 #ifdef MAC
368 	int error;
369 #endif
370 
371 	shmfd = fp->f_data;
372 
373 #ifdef MAC
374 	error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
375 	if (error)
376 		return (error);
377 #endif
378 
379 	/*
380 	 * Attempt to return sanish values for fstat() on a memory file
381 	 * descriptor.
382 	 */
383 	bzero(sb, sizeof(*sb));
384 	sb->st_blksize = PAGE_SIZE;
385 	sb->st_size = shmfd->shm_size;
386 	sb->st_blocks = (sb->st_size + sb->st_blksize - 1) / sb->st_blksize;
387 	mtx_lock(&shm_timestamp_lock);
388 	sb->st_atim = shmfd->shm_atime;
389 	sb->st_ctim = shmfd->shm_ctime;
390 	sb->st_mtim = shmfd->shm_mtime;
391 	sb->st_birthtim = shmfd->shm_birthtime;
392 	sb->st_mode = S_IFREG | shmfd->shm_mode;		/* XXX */
393 	sb->st_uid = shmfd->shm_uid;
394 	sb->st_gid = shmfd->shm_gid;
395 	mtx_unlock(&shm_timestamp_lock);
396 	sb->st_dev = shm_dev_ino;
397 	sb->st_ino = shmfd->shm_ino;
398 
399 	return (0);
400 }
401 
402 static int
403 shm_close(struct file *fp, struct thread *td)
404 {
405 	struct shmfd *shmfd;
406 
407 	shmfd = fp->f_data;
408 	fp->f_data = NULL;
409 	shm_drop(shmfd);
410 
411 	return (0);
412 }
413 
414 static int
415 shm_dotruncate(struct shmfd *shmfd, off_t length)
416 {
417 	vm_object_t object;
418 	vm_page_t m;
419 	vm_pindex_t idx, nobjsize;
420 	vm_ooffset_t delta;
421 	int base, rv;
422 
423 	object = shmfd->shm_object;
424 	VM_OBJECT_WLOCK(object);
425 	if (length == shmfd->shm_size) {
426 		VM_OBJECT_WUNLOCK(object);
427 		return (0);
428 	}
429 	nobjsize = OFF_TO_IDX(length + PAGE_MASK);
430 
431 	/* Are we shrinking?  If so, trim the end. */
432 	if (length < shmfd->shm_size) {
433 		/*
434 		 * Disallow any requests to shrink the size if this
435 		 * object is mapped into the kernel.
436 		 */
437 		if (shmfd->shm_kmappings > 0) {
438 			VM_OBJECT_WUNLOCK(object);
439 			return (EBUSY);
440 		}
441 
442 		/*
443 		 * Zero the truncated part of the last page.
444 		 */
445 		base = length & PAGE_MASK;
446 		if (base != 0) {
447 			idx = OFF_TO_IDX(length);
448 retry:
449 			m = vm_page_lookup(object, idx);
450 			if (m != NULL) {
451 				if (vm_page_sleep_if_busy(m, "shmtrc"))
452 					goto retry;
453 			} else if (vm_pager_has_page(object, idx, NULL, NULL)) {
454 				m = vm_page_alloc(object, idx, VM_ALLOC_NORMAL);
455 				if (m == NULL) {
456 					VM_OBJECT_WUNLOCK(object);
457 					VM_WAIT;
458 					VM_OBJECT_WLOCK(object);
459 					goto retry;
460 				} else if (m->valid != VM_PAGE_BITS_ALL)
461 					rv = vm_pager_get_pages(object, &m, 1,
462 					    0);
463 				else
464 					/* A cached page was reactivated. */
465 					rv = VM_PAGER_OK;
466 				vm_page_lock(m);
467 				if (rv == VM_PAGER_OK) {
468 					vm_page_deactivate(m);
469 					vm_page_unlock(m);
470 					vm_page_xunbusy(m);
471 				} else {
472 					vm_page_free(m);
473 					vm_page_unlock(m);
474 					VM_OBJECT_WUNLOCK(object);
475 					return (EIO);
476 				}
477 			}
478 			if (m != NULL) {
479 				pmap_zero_page_area(m, base, PAGE_SIZE - base);
480 				KASSERT(m->valid == VM_PAGE_BITS_ALL,
481 				    ("shm_dotruncate: page %p is invalid", m));
482 				vm_page_dirty(m);
483 				vm_pager_page_unswapped(m);
484 			}
485 		}
486 		delta = ptoa(object->size - nobjsize);
487 
488 		/* Toss in memory pages. */
489 		if (nobjsize < object->size)
490 			vm_object_page_remove(object, nobjsize, object->size,
491 			    0);
492 
493 		/* Toss pages from swap. */
494 		if (object->type == OBJT_SWAP)
495 			swap_pager_freespace(object, nobjsize, delta);
496 
497 		/* Free the swap accounted for shm */
498 		swap_release_by_cred(delta, object->cred);
499 		object->charge -= delta;
500 	} else {
501 		/* Attempt to reserve the swap */
502 		delta = ptoa(nobjsize - object->size);
503 		if (!swap_reserve_by_cred(delta, object->cred)) {
504 			VM_OBJECT_WUNLOCK(object);
505 			return (ENOMEM);
506 		}
507 		object->charge += delta;
508 	}
509 	shmfd->shm_size = length;
510 	mtx_lock(&shm_timestamp_lock);
511 	vfs_timestamp(&shmfd->shm_ctime);
512 	shmfd->shm_mtime = shmfd->shm_ctime;
513 	mtx_unlock(&shm_timestamp_lock);
514 	object->size = nobjsize;
515 	VM_OBJECT_WUNLOCK(object);
516 	return (0);
517 }
518 
519 /*
520  * shmfd object management including creation and reference counting
521  * routines.
522  */
523 static struct shmfd *
524 shm_alloc(struct ucred *ucred, mode_t mode)
525 {
526 	struct shmfd *shmfd;
527 	int ino;
528 
529 	shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
530 	shmfd->shm_size = 0;
531 	shmfd->shm_uid = ucred->cr_uid;
532 	shmfd->shm_gid = ucred->cr_gid;
533 	shmfd->shm_mode = mode;
534 	shmfd->shm_object = vm_pager_allocate(OBJT_DEFAULT, NULL,
535 	    shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
536 	KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
537 	shmfd->shm_object->pg_color = 0;
538 	VM_OBJECT_WLOCK(shmfd->shm_object);
539 	vm_object_clear_flag(shmfd->shm_object, OBJ_ONEMAPPING);
540 	vm_object_set_flag(shmfd->shm_object, OBJ_COLORED | OBJ_NOSPLIT);
541 	VM_OBJECT_WUNLOCK(shmfd->shm_object);
542 	vfs_timestamp(&shmfd->shm_birthtime);
543 	shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
544 	    shmfd->shm_birthtime;
545 	ino = alloc_unr(shm_ino_unr);
546 	if (ino == -1)
547 		shmfd->shm_ino = 0;
548 	else
549 		shmfd->shm_ino = ino;
550 	refcount_init(&shmfd->shm_refs, 1);
551 	mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
552 	rangelock_init(&shmfd->shm_rl);
553 #ifdef MAC
554 	mac_posixshm_init(shmfd);
555 	mac_posixshm_create(ucred, shmfd);
556 #endif
557 
558 	return (shmfd);
559 }
560 
561 static struct shmfd *
562 shm_hold(struct shmfd *shmfd)
563 {
564 
565 	refcount_acquire(&shmfd->shm_refs);
566 	return (shmfd);
567 }
568 
569 static void
570 shm_drop(struct shmfd *shmfd)
571 {
572 
573 	if (refcount_release(&shmfd->shm_refs)) {
574 #ifdef MAC
575 		mac_posixshm_destroy(shmfd);
576 #endif
577 		rangelock_destroy(&shmfd->shm_rl);
578 		mtx_destroy(&shmfd->shm_mtx);
579 		vm_object_deallocate(shmfd->shm_object);
580 		if (shmfd->shm_ino != 0)
581 			free_unr(shm_ino_unr, shmfd->shm_ino);
582 		free(shmfd, M_SHMFD);
583 	}
584 }
585 
586 /*
587  * Determine if the credentials have sufficient permissions for a
588  * specified combination of FREAD and FWRITE.
589  */
590 static int
591 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
592 {
593 	accmode_t accmode;
594 	int error;
595 
596 	accmode = 0;
597 	if (flags & FREAD)
598 		accmode |= VREAD;
599 	if (flags & FWRITE)
600 		accmode |= VWRITE;
601 	mtx_lock(&shm_timestamp_lock);
602 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
603 	    accmode, ucred, NULL);
604 	mtx_unlock(&shm_timestamp_lock);
605 	return (error);
606 }
607 
608 /*
609  * Dictionary management.  We maintain an in-kernel dictionary to map
610  * paths to shmfd objects.  We use the FNV hash on the path to store
611  * the mappings in a hash table.
612  */
613 static void
614 shm_init(void *arg)
615 {
616 
617 	mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
618 	sx_init(&shm_dict_lock, "shm dictionary");
619 	shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
620 	shm_ino_unr = new_unrhdr(1, INT32_MAX, NULL);
621 	KASSERT(shm_ino_unr != NULL, ("shm fake inodes not initialized"));
622 	shm_dev_ino = devfs_alloc_cdp_inode();
623 	KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
624 }
625 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
626 
627 static struct shmfd *
628 shm_lookup(char *path, Fnv32_t fnv)
629 {
630 	struct shm_mapping *map;
631 
632 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
633 		if (map->sm_fnv != fnv)
634 			continue;
635 		if (strcmp(map->sm_path, path) == 0)
636 			return (map->sm_shmfd);
637 	}
638 
639 	return (NULL);
640 }
641 
642 static void
643 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
644 {
645 	struct shm_mapping *map;
646 
647 	map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
648 	map->sm_path = path;
649 	map->sm_fnv = fnv;
650 	map->sm_shmfd = shm_hold(shmfd);
651 	shmfd->shm_path = path;
652 	LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
653 }
654 
655 static int
656 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
657 {
658 	struct shm_mapping *map;
659 	int error;
660 
661 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
662 		if (map->sm_fnv != fnv)
663 			continue;
664 		if (strcmp(map->sm_path, path) == 0) {
665 #ifdef MAC
666 			error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
667 			if (error)
668 				return (error);
669 #endif
670 			error = shm_access(map->sm_shmfd, ucred,
671 			    FREAD | FWRITE);
672 			if (error)
673 				return (error);
674 			map->sm_shmfd->shm_path = NULL;
675 			LIST_REMOVE(map, sm_link);
676 			shm_drop(map->sm_shmfd);
677 			free(map->sm_path, M_SHMFD);
678 			free(map, M_SHMFD);
679 			return (0);
680 		}
681 	}
682 
683 	return (ENOENT);
684 }
685 
686 /* System calls. */
687 int
688 sys_shm_open(struct thread *td, struct shm_open_args *uap)
689 {
690 	struct filedesc *fdp;
691 	struct shmfd *shmfd;
692 	struct file *fp;
693 	char *path;
694 	Fnv32_t fnv;
695 	mode_t cmode;
696 	int fd, error;
697 
698 #ifdef CAPABILITY_MODE
699 	/*
700 	 * shm_open(2) is only allowed for anonymous objects.
701 	 */
702 	if (IN_CAPABILITY_MODE(td) && (uap->path != SHM_ANON))
703 		return (ECAPMODE);
704 #endif
705 
706 	if ((uap->flags & O_ACCMODE) != O_RDONLY &&
707 	    (uap->flags & O_ACCMODE) != O_RDWR)
708 		return (EINVAL);
709 
710 	if ((uap->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 = (uap->mode & ~fdp->fd_cmask) & ACCESSPERMS;
715 
716 	error = falloc(td, &fp, &fd, O_CLOEXEC);
717 	if (error)
718 		return (error);
719 
720 	/* A SHM_ANON path pointer creates an anonymous object. */
721 	if (uap->path == SHM_ANON) {
722 		/* A read-only anonymous object is pointless. */
723 		if ((uap->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(uap->path, 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 (uap->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 ((uap->flags & (O_CREAT | O_EXCL)) ==
773 			    (O_CREAT | O_EXCL))
774 				error = EEXIST;
775 			else {
776 #ifdef MAC
777 				error = mac_posixshm_check_open(td->td_ucred,
778 				    shmfd, FFLAGS(uap->flags & O_ACCMODE));
779 				if (error == 0)
780 #endif
781 				error = shm_access(shmfd, td->td_ucred,
782 				    FFLAGS(uap->flags & O_ACCMODE));
783 			}
784 
785 			/*
786 			 * Truncate the file back to zero length if
787 			 * O_TRUNC was specified and the object was
788 			 * opened with read/write.
789 			 */
790 			if (error == 0 &&
791 			    (uap->flags & (O_ACCMODE | O_TRUNC)) ==
792 			    (O_RDWR | O_TRUNC)) {
793 #ifdef MAC
794 				error = mac_posixshm_check_truncate(
795 					td->td_ucred, fp->f_cred, shmfd);
796 				if (error == 0)
797 #endif
798 					shm_dotruncate(shmfd, 0);
799 			}
800 			if (error == 0)
801 				shm_hold(shmfd);
802 		}
803 		sx_xunlock(&shm_dict_lock);
804 
805 		if (error) {
806 			fdclose(td, fp, fd);
807 			fdrop(fp, td);
808 			return (error);
809 		}
810 	}
811 
812 	finit(fp, FFLAGS(uap->flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
813 
814 	td->td_retval[0] = fd;
815 	fdrop(fp, td);
816 
817 	return (0);
818 }
819 
820 int
821 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
822 {
823 	char *path;
824 	Fnv32_t fnv;
825 	int error;
826 
827 	path = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
828 	error = copyinstr(uap->path, path, MAXPATHLEN, NULL);
829 	if (error) {
830 		free(path, M_TEMP);
831 		return (error);
832 	}
833 #ifdef KTRACE
834 	if (KTRPOINT(curthread, KTR_NAMEI))
835 		ktrnamei(path);
836 #endif
837 	fnv = fnv_32_str(path, FNV1_32_INIT);
838 	sx_xlock(&shm_dict_lock);
839 	error = shm_remove(path, fnv, td->td_ucred);
840 	sx_xunlock(&shm_dict_lock);
841 	free(path, M_TEMP);
842 
843 	return (error);
844 }
845 
846 int
847 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
848     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
849     vm_ooffset_t foff, struct thread *td)
850 {
851 	struct shmfd *shmfd;
852 	vm_prot_t maxprot;
853 	int error;
854 
855 	shmfd = fp->f_data;
856 	maxprot = VM_PROT_NONE;
857 
858 	/* FREAD should always be set. */
859 	if ((fp->f_flag & FREAD) != 0)
860 		maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
861 	if ((fp->f_flag & FWRITE) != 0)
862 		maxprot |= VM_PROT_WRITE;
863 
864 	/* Don't permit shared writable mappings on read-only descriptors. */
865 	if ((flags & MAP_SHARED) != 0 &&
866 	    (maxprot & VM_PROT_WRITE) == 0 &&
867 	    (prot & VM_PROT_WRITE) != 0)
868 		return (EACCES);
869 	maxprot &= cap_maxprot;
870 
871 #ifdef MAC
872 	error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
873 	if (error != 0)
874 		return (error);
875 #endif
876 
877 	/*
878 	 * XXXRW: This validation is probably insufficient, and subject to
879 	 * sign errors.  It should be fixed.
880 	 */
881 	if (foff >= shmfd->shm_size ||
882 	    foff + objsize > round_page(shmfd->shm_size))
883 		return (EINVAL);
884 
885 	mtx_lock(&shm_timestamp_lock);
886 	vfs_timestamp(&shmfd->shm_atime);
887 	mtx_unlock(&shm_timestamp_lock);
888 	vm_object_reference(shmfd->shm_object);
889 
890 	error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
891 	    shmfd->shm_object, foff, FALSE, td);
892 	if (error != 0)
893 		vm_object_deallocate(shmfd->shm_object);
894 	return (0);
895 }
896 
897 static int
898 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
899     struct thread *td)
900 {
901 	struct shmfd *shmfd;
902 	int error;
903 
904 	error = 0;
905 	shmfd = fp->f_data;
906 	mtx_lock(&shm_timestamp_lock);
907 	/*
908 	 * SUSv4 says that x bits of permission need not be affected.
909 	 * Be consistent with our shm_open there.
910 	 */
911 #ifdef MAC
912 	error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
913 	if (error != 0)
914 		goto out;
915 #endif
916 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid,
917 	    shmfd->shm_gid, VADMIN, active_cred, NULL);
918 	if (error != 0)
919 		goto out;
920 	shmfd->shm_mode = mode & ACCESSPERMS;
921 out:
922 	mtx_unlock(&shm_timestamp_lock);
923 	return (error);
924 }
925 
926 static int
927 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
928     struct thread *td)
929 {
930 	struct shmfd *shmfd;
931 	int error;
932 
933 	error = 0;
934 	shmfd = fp->f_data;
935 	mtx_lock(&shm_timestamp_lock);
936 #ifdef MAC
937 	error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
938 	if (error != 0)
939 		goto out;
940 #endif
941 	if (uid == (uid_t)-1)
942 		uid = shmfd->shm_uid;
943 	if (gid == (gid_t)-1)
944                  gid = shmfd->shm_gid;
945 	if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
946 	    (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
947 	    (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN, 0)))
948 		goto out;
949 	shmfd->shm_uid = uid;
950 	shmfd->shm_gid = gid;
951 out:
952 	mtx_unlock(&shm_timestamp_lock);
953 	return (error);
954 }
955 
956 /*
957  * Helper routines to allow the backing object of a shared memory file
958  * descriptor to be mapped in the kernel.
959  */
960 int
961 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
962 {
963 	struct shmfd *shmfd;
964 	vm_offset_t kva, ofs;
965 	vm_object_t obj;
966 	int rv;
967 
968 	if (fp->f_type != DTYPE_SHM)
969 		return (EINVAL);
970 	shmfd = fp->f_data;
971 	obj = shmfd->shm_object;
972 	VM_OBJECT_WLOCK(obj);
973 	/*
974 	 * XXXRW: This validation is probably insufficient, and subject to
975 	 * sign errors.  It should be fixed.
976 	 */
977 	if (offset >= shmfd->shm_size ||
978 	    offset + size > round_page(shmfd->shm_size)) {
979 		VM_OBJECT_WUNLOCK(obj);
980 		return (EINVAL);
981 	}
982 
983 	shmfd->shm_kmappings++;
984 	vm_object_reference_locked(obj);
985 	VM_OBJECT_WUNLOCK(obj);
986 
987 	/* Map the object into the kernel_map and wire it. */
988 	kva = vm_map_min(kernel_map);
989 	ofs = offset & PAGE_MASK;
990 	offset = trunc_page(offset);
991 	size = round_page(size + ofs);
992 	rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
993 	    VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
994 	    VM_PROT_READ | VM_PROT_WRITE, 0);
995 	if (rv == KERN_SUCCESS) {
996 		rv = vm_map_wire(kernel_map, kva, kva + size,
997 		    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
998 		if (rv == KERN_SUCCESS) {
999 			*memp = (void *)(kva + ofs);
1000 			return (0);
1001 		}
1002 		vm_map_remove(kernel_map, kva, kva + size);
1003 	} else
1004 		vm_object_deallocate(obj);
1005 
1006 	/* On failure, drop our mapping reference. */
1007 	VM_OBJECT_WLOCK(obj);
1008 	shmfd->shm_kmappings--;
1009 	VM_OBJECT_WUNLOCK(obj);
1010 
1011 	return (vm_mmap_to_errno(rv));
1012 }
1013 
1014 /*
1015  * We require the caller to unmap the entire entry.  This allows us to
1016  * safely decrement shm_kmappings when a mapping is removed.
1017  */
1018 int
1019 shm_unmap(struct file *fp, void *mem, size_t size)
1020 {
1021 	struct shmfd *shmfd;
1022 	vm_map_entry_t entry;
1023 	vm_offset_t kva, ofs;
1024 	vm_object_t obj;
1025 	vm_pindex_t pindex;
1026 	vm_prot_t prot;
1027 	boolean_t wired;
1028 	vm_map_t map;
1029 	int rv;
1030 
1031 	if (fp->f_type != DTYPE_SHM)
1032 		return (EINVAL);
1033 	shmfd = fp->f_data;
1034 	kva = (vm_offset_t)mem;
1035 	ofs = kva & PAGE_MASK;
1036 	kva = trunc_page(kva);
1037 	size = round_page(size + ofs);
1038 	map = kernel_map;
1039 	rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1040 	    &obj, &pindex, &prot, &wired);
1041 	if (rv != KERN_SUCCESS)
1042 		return (EINVAL);
1043 	if (entry->start != kva || entry->end != kva + size) {
1044 		vm_map_lookup_done(map, entry);
1045 		return (EINVAL);
1046 	}
1047 	vm_map_lookup_done(map, entry);
1048 	if (obj != shmfd->shm_object)
1049 		return (EINVAL);
1050 	vm_map_remove(map, kva, kva + size);
1051 	VM_OBJECT_WLOCK(obj);
1052 	KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1053 	shmfd->shm_kmappings--;
1054 	VM_OBJECT_WUNLOCK(obj);
1055 	return (0);
1056 }
1057 
1058 static int
1059 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1060 {
1061 	struct shmfd *shmfd;
1062 
1063 	kif->kf_type = KF_TYPE_SHM;
1064 	shmfd = fp->f_data;
1065 
1066 	mtx_lock(&shm_timestamp_lock);
1067 	kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;	/* XXX */
1068 	mtx_unlock(&shm_timestamp_lock);
1069 	kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1070 	if (shmfd->shm_path != NULL) {
1071 		sx_slock(&shm_dict_lock);
1072 		if (shmfd->shm_path != NULL)
1073 			strlcpy(kif->kf_path, shmfd->shm_path,
1074 			    sizeof(kif->kf_path));
1075 		sx_sunlock(&shm_dict_lock);
1076 	}
1077 	return (0);
1078 }
1079