xref: /freebsd/sys/kern/uipc_shm.c (revision 2e36db147e0edeb1b213f283f9b41a8d1a902a88)
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 (vm_page_active(m))
213 		vm_page_reference(m);
214 	else
215 		vm_page_activate(m);
216 	vm_page_unlock(m);
217 	VM_OBJECT_WUNLOCK(obj);
218 	error = uiomove_fromphys(&m, offset, tlen, uio);
219 	if (uio->uio_rw == UIO_WRITE && error == 0) {
220 		VM_OBJECT_WLOCK(obj);
221 		vm_page_dirty(m);
222 		vm_pager_page_unswapped(m);
223 		VM_OBJECT_WUNLOCK(obj);
224 	}
225 	vm_page_lock(m);
226 	vm_page_unhold(m);
227 	vm_page_unlock(m);
228 
229 	return (error);
230 }
231 
232 int
233 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
234 {
235 	ssize_t resid;
236 	size_t len;
237 	int error;
238 
239 	error = 0;
240 	while ((resid = uio->uio_resid) > 0) {
241 		if (obj_size <= uio->uio_offset)
242 			break;
243 		len = MIN(obj_size - uio->uio_offset, resid);
244 		if (len == 0)
245 			break;
246 		error = uiomove_object_page(obj, len, uio);
247 		if (error != 0 || resid == uio->uio_resid)
248 			break;
249 	}
250 	return (error);
251 }
252 
253 static int
254 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
255 {
256 	struct shmfd *shmfd;
257 	off_t foffset;
258 	int error;
259 
260 	shmfd = fp->f_data;
261 	foffset = foffset_lock(fp, 0);
262 	error = 0;
263 	switch (whence) {
264 	case L_INCR:
265 		if (foffset < 0 ||
266 		    (offset > 0 && foffset > OFF_MAX - offset)) {
267 			error = EOVERFLOW;
268 			break;
269 		}
270 		offset += foffset;
271 		break;
272 	case L_XTND:
273 		if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
274 			error = EOVERFLOW;
275 			break;
276 		}
277 		offset += shmfd->shm_size;
278 		break;
279 	case L_SET:
280 		break;
281 	default:
282 		error = EINVAL;
283 	}
284 	if (error == 0) {
285 		if (offset < 0 || offset > shmfd->shm_size)
286 			error = EINVAL;
287 		else
288 			td->td_uretoff.tdu_off = offset;
289 	}
290 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
291 	return (error);
292 }
293 
294 static int
295 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
296     int flags, struct thread *td)
297 {
298 	struct shmfd *shmfd;
299 	void *rl_cookie;
300 	int error;
301 
302 	shmfd = fp->f_data;
303 #ifdef MAC
304 	error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
305 	if (error)
306 		return (error);
307 #endif
308 	foffset_lock_uio(fp, uio, flags);
309 	rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset,
310 	    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
311 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
312 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
313 	foffset_unlock_uio(fp, uio, flags);
314 	return (error);
315 }
316 
317 static int
318 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
319     int flags, struct thread *td)
320 {
321 	struct shmfd *shmfd;
322 	void *rl_cookie;
323 	int error;
324 
325 	shmfd = fp->f_data;
326 #ifdef MAC
327 	error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
328 	if (error)
329 		return (error);
330 #endif
331 	foffset_lock_uio(fp, uio, flags);
332 	if ((flags & FOF_OFFSET) == 0) {
333 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
334 		    &shmfd->shm_mtx);
335 	} else {
336 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset,
337 		    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
338 	}
339 
340 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
341 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
342 	foffset_unlock_uio(fp, uio, flags);
343 	return (error);
344 }
345 
346 static int
347 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
348     struct thread *td)
349 {
350 	struct shmfd *shmfd;
351 #ifdef MAC
352 	int error;
353 #endif
354 
355 	shmfd = fp->f_data;
356 #ifdef MAC
357 	error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
358 	if (error)
359 		return (error);
360 #endif
361 	return (shm_dotruncate(shmfd, length));
362 }
363 
364 static int
365 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
366     struct thread *td)
367 {
368 	struct shmfd *shmfd;
369 #ifdef MAC
370 	int error;
371 #endif
372 
373 	shmfd = fp->f_data;
374 
375 #ifdef MAC
376 	error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
377 	if (error)
378 		return (error);
379 #endif
380 
381 	/*
382 	 * Attempt to return sanish values for fstat() on a memory file
383 	 * descriptor.
384 	 */
385 	bzero(sb, sizeof(*sb));
386 	sb->st_blksize = PAGE_SIZE;
387 	sb->st_size = shmfd->shm_size;
388 	sb->st_blocks = howmany(sb->st_size, sb->st_blksize);
389 	mtx_lock(&shm_timestamp_lock);
390 	sb->st_atim = shmfd->shm_atime;
391 	sb->st_ctim = shmfd->shm_ctime;
392 	sb->st_mtim = shmfd->shm_mtime;
393 	sb->st_birthtim = shmfd->shm_birthtime;
394 	sb->st_mode = S_IFREG | shmfd->shm_mode;		/* XXX */
395 	sb->st_uid = shmfd->shm_uid;
396 	sb->st_gid = shmfd->shm_gid;
397 	mtx_unlock(&shm_timestamp_lock);
398 	sb->st_dev = shm_dev_ino;
399 	sb->st_ino = shmfd->shm_ino;
400 
401 	return (0);
402 }
403 
404 static int
405 shm_close(struct file *fp, struct thread *td)
406 {
407 	struct shmfd *shmfd;
408 
409 	shmfd = fp->f_data;
410 	fp->f_data = NULL;
411 	shm_drop(shmfd);
412 
413 	return (0);
414 }
415 
416 int
417 shm_dotruncate(struct shmfd *shmfd, off_t length)
418 {
419 	vm_object_t object;
420 	vm_page_t m;
421 	vm_pindex_t idx, nobjsize;
422 	vm_ooffset_t delta;
423 	int base, rv;
424 
425 	KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
426 	object = shmfd->shm_object;
427 	VM_OBJECT_WLOCK(object);
428 	if (length == shmfd->shm_size) {
429 		VM_OBJECT_WUNLOCK(object);
430 		return (0);
431 	}
432 	nobjsize = OFF_TO_IDX(length + PAGE_MASK);
433 
434 	/* Are we shrinking?  If so, trim the end. */
435 	if (length < shmfd->shm_size) {
436 		/*
437 		 * Disallow any requests to shrink the size if this
438 		 * object is mapped into the kernel.
439 		 */
440 		if (shmfd->shm_kmappings > 0) {
441 			VM_OBJECT_WUNLOCK(object);
442 			return (EBUSY);
443 		}
444 
445 		/*
446 		 * Zero the truncated part of the last page.
447 		 */
448 		base = length & PAGE_MASK;
449 		if (base != 0) {
450 			idx = OFF_TO_IDX(length);
451 retry:
452 			m = vm_page_lookup(object, idx);
453 			if (m != NULL) {
454 				if (vm_page_sleep_if_busy(m, "shmtrc"))
455 					goto retry;
456 			} else if (vm_pager_has_page(object, idx, NULL, NULL)) {
457 				m = vm_page_alloc(object, idx,
458 				    VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL);
459 				if (m == NULL)
460 					goto retry;
461 				rv = vm_pager_get_pages(object, &m, 1, NULL,
462 				    NULL);
463 				vm_page_lock(m);
464 				if (rv == VM_PAGER_OK) {
465 					/*
466 					 * Since the page was not resident,
467 					 * and therefore not recently
468 					 * accessed, immediately enqueue it
469 					 * for asynchronous laundering.  The
470 					 * current operation is not regarded
471 					 * as an access.
472 					 */
473 					vm_page_launder(m);
474 					vm_page_unlock(m);
475 					vm_page_xunbusy(m);
476 				} else {
477 					vm_page_free(m);
478 					vm_page_unlock(m);
479 					VM_OBJECT_WUNLOCK(object);
480 					return (EIO);
481 				}
482 			}
483 			if (m != NULL) {
484 				pmap_zero_page_area(m, base, PAGE_SIZE - base);
485 				KASSERT(m->valid == VM_PAGE_BITS_ALL,
486 				    ("shm_dotruncate: page %p is invalid", m));
487 				vm_page_dirty(m);
488 				vm_pager_page_unswapped(m);
489 			}
490 		}
491 		delta = IDX_TO_OFF(object->size - nobjsize);
492 
493 		/* Toss in memory pages. */
494 		if (nobjsize < object->size)
495 			vm_object_page_remove(object, nobjsize, object->size,
496 			    0);
497 
498 		/* Toss pages from swap. */
499 		if (object->type == OBJT_SWAP)
500 			swap_pager_freespace(object, nobjsize, delta);
501 
502 		/* Free the swap accounted for shm */
503 		swap_release_by_cred(delta, object->cred);
504 		object->charge -= delta;
505 	} else {
506 		/* Try to reserve additional swap space. */
507 		delta = IDX_TO_OFF(nobjsize - object->size);
508 		if (!swap_reserve_by_cred(delta, object->cred)) {
509 			VM_OBJECT_WUNLOCK(object);
510 			return (ENOMEM);
511 		}
512 		object->charge += delta;
513 	}
514 	shmfd->shm_size = length;
515 	mtx_lock(&shm_timestamp_lock);
516 	vfs_timestamp(&shmfd->shm_ctime);
517 	shmfd->shm_mtime = shmfd->shm_ctime;
518 	mtx_unlock(&shm_timestamp_lock);
519 	object->size = nobjsize;
520 	VM_OBJECT_WUNLOCK(object);
521 	return (0);
522 }
523 
524 /*
525  * shmfd object management including creation and reference counting
526  * routines.
527  */
528 struct shmfd *
529 shm_alloc(struct ucred *ucred, mode_t mode)
530 {
531 	struct shmfd *shmfd;
532 	int ino;
533 
534 	shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
535 	shmfd->shm_size = 0;
536 	shmfd->shm_uid = ucred->cr_uid;
537 	shmfd->shm_gid = ucred->cr_gid;
538 	shmfd->shm_mode = mode;
539 	shmfd->shm_object = vm_pager_allocate(OBJT_DEFAULT, NULL,
540 	    shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
541 	KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
542 	shmfd->shm_object->pg_color = 0;
543 	VM_OBJECT_WLOCK(shmfd->shm_object);
544 	vm_object_clear_flag(shmfd->shm_object, OBJ_ONEMAPPING);
545 	vm_object_set_flag(shmfd->shm_object, OBJ_COLORED | OBJ_NOSPLIT);
546 	VM_OBJECT_WUNLOCK(shmfd->shm_object);
547 	vfs_timestamp(&shmfd->shm_birthtime);
548 	shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
549 	    shmfd->shm_birthtime;
550 	ino = alloc_unr(shm_ino_unr);
551 	if (ino == -1)
552 		shmfd->shm_ino = 0;
553 	else
554 		shmfd->shm_ino = ino;
555 	refcount_init(&shmfd->shm_refs, 1);
556 	mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
557 	rangelock_init(&shmfd->shm_rl);
558 #ifdef MAC
559 	mac_posixshm_init(shmfd);
560 	mac_posixshm_create(ucred, shmfd);
561 #endif
562 
563 	return (shmfd);
564 }
565 
566 struct shmfd *
567 shm_hold(struct shmfd *shmfd)
568 {
569 
570 	refcount_acquire(&shmfd->shm_refs);
571 	return (shmfd);
572 }
573 
574 void
575 shm_drop(struct shmfd *shmfd)
576 {
577 
578 	if (refcount_release(&shmfd->shm_refs)) {
579 #ifdef MAC
580 		mac_posixshm_destroy(shmfd);
581 #endif
582 		rangelock_destroy(&shmfd->shm_rl);
583 		mtx_destroy(&shmfd->shm_mtx);
584 		vm_object_deallocate(shmfd->shm_object);
585 		if (shmfd->shm_ino != 0)
586 			free_unr(shm_ino_unr, shmfd->shm_ino);
587 		free(shmfd, M_SHMFD);
588 	}
589 }
590 
591 /*
592  * Determine if the credentials have sufficient permissions for a
593  * specified combination of FREAD and FWRITE.
594  */
595 int
596 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
597 {
598 	accmode_t accmode;
599 	int error;
600 
601 	accmode = 0;
602 	if (flags & FREAD)
603 		accmode |= VREAD;
604 	if (flags & FWRITE)
605 		accmode |= VWRITE;
606 	mtx_lock(&shm_timestamp_lock);
607 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
608 	    accmode, ucred, NULL);
609 	mtx_unlock(&shm_timestamp_lock);
610 	return (error);
611 }
612 
613 /*
614  * Dictionary management.  We maintain an in-kernel dictionary to map
615  * paths to shmfd objects.  We use the FNV hash on the path to store
616  * the mappings in a hash table.
617  */
618 static void
619 shm_init(void *arg)
620 {
621 
622 	mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
623 	sx_init(&shm_dict_lock, "shm dictionary");
624 	shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
625 	shm_ino_unr = new_unrhdr(1, INT32_MAX, NULL);
626 	KASSERT(shm_ino_unr != NULL, ("shm fake inodes not initialized"));
627 	shm_dev_ino = devfs_alloc_cdp_inode();
628 	KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
629 }
630 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
631 
632 static struct shmfd *
633 shm_lookup(char *path, Fnv32_t fnv)
634 {
635 	struct shm_mapping *map;
636 
637 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
638 		if (map->sm_fnv != fnv)
639 			continue;
640 		if (strcmp(map->sm_path, path) == 0)
641 			return (map->sm_shmfd);
642 	}
643 
644 	return (NULL);
645 }
646 
647 static void
648 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
649 {
650 	struct shm_mapping *map;
651 
652 	map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
653 	map->sm_path = path;
654 	map->sm_fnv = fnv;
655 	map->sm_shmfd = shm_hold(shmfd);
656 	shmfd->shm_path = path;
657 	LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
658 }
659 
660 static int
661 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
662 {
663 	struct shm_mapping *map;
664 	int error;
665 
666 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
667 		if (map->sm_fnv != fnv)
668 			continue;
669 		if (strcmp(map->sm_path, path) == 0) {
670 #ifdef MAC
671 			error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
672 			if (error)
673 				return (error);
674 #endif
675 			error = shm_access(map->sm_shmfd, ucred,
676 			    FREAD | FWRITE);
677 			if (error)
678 				return (error);
679 			map->sm_shmfd->shm_path = NULL;
680 			LIST_REMOVE(map, sm_link);
681 			shm_drop(map->sm_shmfd);
682 			free(map->sm_path, M_SHMFD);
683 			free(map, M_SHMFD);
684 			return (0);
685 		}
686 	}
687 
688 	return (ENOENT);
689 }
690 
691 int
692 kern_shm_open(struct thread *td, const char *userpath, int flags, mode_t mode,
693     struct filecaps *fcaps)
694 {
695 	struct filedesc *fdp;
696 	struct shmfd *shmfd;
697 	struct file *fp;
698 	char *path;
699 	const char *pr_path;
700 	size_t pr_pathlen;
701 	Fnv32_t fnv;
702 	mode_t cmode;
703 	int fd, error;
704 
705 #ifdef CAPABILITY_MODE
706 	/*
707 	 * shm_open(2) is only allowed for anonymous objects.
708 	 */
709 	if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON))
710 		return (ECAPMODE);
711 #endif
712 
713 	AUDIT_ARG_FFLAGS(flags);
714 	AUDIT_ARG_MODE(mode);
715 
716 	if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
717 		return (EINVAL);
718 
719 	if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
720 		return (EINVAL);
721 
722 	fdp = td->td_proc->p_fd;
723 	cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS;
724 
725 	error = falloc_caps(td, &fp, &fd, O_CLOEXEC, fcaps);
726 	if (error)
727 		return (error);
728 
729 	/* A SHM_ANON path pointer creates an anonymous object. */
730 	if (userpath == SHM_ANON) {
731 		/* A read-only anonymous object is pointless. */
732 		if ((flags & O_ACCMODE) == O_RDONLY) {
733 			fdclose(td, fp, fd);
734 			fdrop(fp, td);
735 			return (EINVAL);
736 		}
737 		shmfd = shm_alloc(td->td_ucred, cmode);
738 	} else {
739 		path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
740 		pr_path = td->td_ucred->cr_prison->pr_path;
741 
742 		/* Construct a full pathname for jailed callers. */
743 		pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
744 		    : strlcpy(path, pr_path, MAXPATHLEN);
745 		error = copyinstr(userpath, path + pr_pathlen,
746 		    MAXPATHLEN - pr_pathlen, NULL);
747 #ifdef KTRACE
748 		if (error == 0 && KTRPOINT(curthread, KTR_NAMEI))
749 			ktrnamei(path);
750 #endif
751 		/* Require paths to start with a '/' character. */
752 		if (error == 0 && path[pr_pathlen] != '/')
753 			error = EINVAL;
754 		if (error) {
755 			fdclose(td, fp, fd);
756 			fdrop(fp, td);
757 			free(path, M_SHMFD);
758 			return (error);
759 		}
760 
761 		AUDIT_ARG_UPATH1_CANON(path);
762 		fnv = fnv_32_str(path, FNV1_32_INIT);
763 		sx_xlock(&shm_dict_lock);
764 		shmfd = shm_lookup(path, fnv);
765 		if (shmfd == NULL) {
766 			/* Object does not yet exist, create it if requested. */
767 			if (flags & O_CREAT) {
768 #ifdef MAC
769 				error = mac_posixshm_check_create(td->td_ucred,
770 				    path);
771 				if (error == 0) {
772 #endif
773 					shmfd = shm_alloc(td->td_ucred, cmode);
774 					shm_insert(path, fnv, shmfd);
775 #ifdef MAC
776 				}
777 #endif
778 			} else {
779 				free(path, M_SHMFD);
780 				error = ENOENT;
781 			}
782 		} else {
783 			/*
784 			 * Object already exists, obtain a new
785 			 * reference if requested and permitted.
786 			 */
787 			free(path, M_SHMFD);
788 			if ((flags & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
789 				error = EEXIST;
790 			else {
791 #ifdef MAC
792 				error = mac_posixshm_check_open(td->td_ucred,
793 				    shmfd, FFLAGS(flags & O_ACCMODE));
794 				if (error == 0)
795 #endif
796 				error = shm_access(shmfd, td->td_ucred,
797 				    FFLAGS(flags & O_ACCMODE));
798 			}
799 
800 			/*
801 			 * Truncate the file back to zero length if
802 			 * O_TRUNC was specified and the object was
803 			 * opened with read/write.
804 			 */
805 			if (error == 0 &&
806 			    (flags & (O_ACCMODE | O_TRUNC)) ==
807 			    (O_RDWR | O_TRUNC)) {
808 #ifdef MAC
809 				error = mac_posixshm_check_truncate(
810 					td->td_ucred, fp->f_cred, shmfd);
811 				if (error == 0)
812 #endif
813 					shm_dotruncate(shmfd, 0);
814 			}
815 			if (error == 0)
816 				shm_hold(shmfd);
817 		}
818 		sx_xunlock(&shm_dict_lock);
819 
820 		if (error) {
821 			fdclose(td, fp, fd);
822 			fdrop(fp, td);
823 			return (error);
824 		}
825 	}
826 
827 	finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
828 
829 	td->td_retval[0] = fd;
830 	fdrop(fp, td);
831 
832 	return (0);
833 }
834 
835 /* System calls. */
836 int
837 sys_shm_open(struct thread *td, struct shm_open_args *uap)
838 {
839 
840 	return (kern_shm_open(td, uap->path, uap->flags, uap->mode, NULL));
841 }
842 
843 int
844 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
845 {
846 	char *path;
847 	const char *pr_path;
848 	size_t pr_pathlen;
849 	Fnv32_t fnv;
850 	int error;
851 
852 	path = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
853 	pr_path = td->td_ucred->cr_prison->pr_path;
854 	pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
855 	    : strlcpy(path, pr_path, MAXPATHLEN);
856 	error = copyinstr(uap->path, path + pr_pathlen, MAXPATHLEN - pr_pathlen,
857 	    NULL);
858 	if (error) {
859 		free(path, M_TEMP);
860 		return (error);
861 	}
862 #ifdef KTRACE
863 	if (KTRPOINT(curthread, KTR_NAMEI))
864 		ktrnamei(path);
865 #endif
866 	AUDIT_ARG_UPATH1_CANON(path);
867 	fnv = fnv_32_str(path, FNV1_32_INIT);
868 	sx_xlock(&shm_dict_lock);
869 	error = shm_remove(path, fnv, td->td_ucred);
870 	sx_xunlock(&shm_dict_lock);
871 	free(path, M_TEMP);
872 
873 	return (error);
874 }
875 
876 int
877 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
878     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
879     vm_ooffset_t foff, struct thread *td)
880 {
881 	struct shmfd *shmfd;
882 	vm_prot_t maxprot;
883 	int error;
884 
885 	shmfd = fp->f_data;
886 	maxprot = VM_PROT_NONE;
887 
888 	/* FREAD should always be set. */
889 	if ((fp->f_flag & FREAD) != 0)
890 		maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
891 	if ((fp->f_flag & FWRITE) != 0)
892 		maxprot |= VM_PROT_WRITE;
893 
894 	/* Don't permit shared writable mappings on read-only descriptors. */
895 	if ((flags & MAP_SHARED) != 0 &&
896 	    (maxprot & VM_PROT_WRITE) == 0 &&
897 	    (prot & VM_PROT_WRITE) != 0)
898 		return (EACCES);
899 	maxprot &= cap_maxprot;
900 
901 	/* See comment in vn_mmap(). */
902 	if (
903 #ifdef _LP64
904 	    objsize > OFF_MAX ||
905 #endif
906 	    foff < 0 || foff > OFF_MAX - objsize)
907 		return (EINVAL);
908 
909 #ifdef MAC
910 	error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
911 	if (error != 0)
912 		return (error);
913 #endif
914 
915 	mtx_lock(&shm_timestamp_lock);
916 	vfs_timestamp(&shmfd->shm_atime);
917 	mtx_unlock(&shm_timestamp_lock);
918 	vm_object_reference(shmfd->shm_object);
919 
920 	error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
921 	    shmfd->shm_object, foff, FALSE, td);
922 	if (error != 0)
923 		vm_object_deallocate(shmfd->shm_object);
924 	return (error);
925 }
926 
927 static int
928 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
929     struct thread *td)
930 {
931 	struct shmfd *shmfd;
932 	int error;
933 
934 	error = 0;
935 	shmfd = fp->f_data;
936 	mtx_lock(&shm_timestamp_lock);
937 	/*
938 	 * SUSv4 says that x bits of permission need not be affected.
939 	 * Be consistent with our shm_open there.
940 	 */
941 #ifdef MAC
942 	error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
943 	if (error != 0)
944 		goto out;
945 #endif
946 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid,
947 	    shmfd->shm_gid, VADMIN, active_cred, NULL);
948 	if (error != 0)
949 		goto out;
950 	shmfd->shm_mode = mode & ACCESSPERMS;
951 out:
952 	mtx_unlock(&shm_timestamp_lock);
953 	return (error);
954 }
955 
956 static int
957 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
958     struct thread *td)
959 {
960 	struct shmfd *shmfd;
961 	int error;
962 
963 	error = 0;
964 	shmfd = fp->f_data;
965 	mtx_lock(&shm_timestamp_lock);
966 #ifdef MAC
967 	error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
968 	if (error != 0)
969 		goto out;
970 #endif
971 	if (uid == (uid_t)-1)
972 		uid = shmfd->shm_uid;
973 	if (gid == (gid_t)-1)
974                  gid = shmfd->shm_gid;
975 	if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
976 	    (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
977 	    (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN, 0)))
978 		goto out;
979 	shmfd->shm_uid = uid;
980 	shmfd->shm_gid = gid;
981 out:
982 	mtx_unlock(&shm_timestamp_lock);
983 	return (error);
984 }
985 
986 /*
987  * Helper routines to allow the backing object of a shared memory file
988  * descriptor to be mapped in the kernel.
989  */
990 int
991 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
992 {
993 	struct shmfd *shmfd;
994 	vm_offset_t kva, ofs;
995 	vm_object_t obj;
996 	int rv;
997 
998 	if (fp->f_type != DTYPE_SHM)
999 		return (EINVAL);
1000 	shmfd = fp->f_data;
1001 	obj = shmfd->shm_object;
1002 	VM_OBJECT_WLOCK(obj);
1003 	/*
1004 	 * XXXRW: This validation is probably insufficient, and subject to
1005 	 * sign errors.  It should be fixed.
1006 	 */
1007 	if (offset >= shmfd->shm_size ||
1008 	    offset + size > round_page(shmfd->shm_size)) {
1009 		VM_OBJECT_WUNLOCK(obj);
1010 		return (EINVAL);
1011 	}
1012 
1013 	shmfd->shm_kmappings++;
1014 	vm_object_reference_locked(obj);
1015 	VM_OBJECT_WUNLOCK(obj);
1016 
1017 	/* Map the object into the kernel_map and wire it. */
1018 	kva = vm_map_min(kernel_map);
1019 	ofs = offset & PAGE_MASK;
1020 	offset = trunc_page(offset);
1021 	size = round_page(size + ofs);
1022 	rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1023 	    VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1024 	    VM_PROT_READ | VM_PROT_WRITE, 0);
1025 	if (rv == KERN_SUCCESS) {
1026 		rv = vm_map_wire(kernel_map, kva, kva + size,
1027 		    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1028 		if (rv == KERN_SUCCESS) {
1029 			*memp = (void *)(kva + ofs);
1030 			return (0);
1031 		}
1032 		vm_map_remove(kernel_map, kva, kva + size);
1033 	} else
1034 		vm_object_deallocate(obj);
1035 
1036 	/* On failure, drop our mapping reference. */
1037 	VM_OBJECT_WLOCK(obj);
1038 	shmfd->shm_kmappings--;
1039 	VM_OBJECT_WUNLOCK(obj);
1040 
1041 	return (vm_mmap_to_errno(rv));
1042 }
1043 
1044 /*
1045  * We require the caller to unmap the entire entry.  This allows us to
1046  * safely decrement shm_kmappings when a mapping is removed.
1047  */
1048 int
1049 shm_unmap(struct file *fp, void *mem, size_t size)
1050 {
1051 	struct shmfd *shmfd;
1052 	vm_map_entry_t entry;
1053 	vm_offset_t kva, ofs;
1054 	vm_object_t obj;
1055 	vm_pindex_t pindex;
1056 	vm_prot_t prot;
1057 	boolean_t wired;
1058 	vm_map_t map;
1059 	int rv;
1060 
1061 	if (fp->f_type != DTYPE_SHM)
1062 		return (EINVAL);
1063 	shmfd = fp->f_data;
1064 	kva = (vm_offset_t)mem;
1065 	ofs = kva & PAGE_MASK;
1066 	kva = trunc_page(kva);
1067 	size = round_page(size + ofs);
1068 	map = kernel_map;
1069 	rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1070 	    &obj, &pindex, &prot, &wired);
1071 	if (rv != KERN_SUCCESS)
1072 		return (EINVAL);
1073 	if (entry->start != kva || entry->end != kva + size) {
1074 		vm_map_lookup_done(map, entry);
1075 		return (EINVAL);
1076 	}
1077 	vm_map_lookup_done(map, entry);
1078 	if (obj != shmfd->shm_object)
1079 		return (EINVAL);
1080 	vm_map_remove(map, kva, kva + size);
1081 	VM_OBJECT_WLOCK(obj);
1082 	KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1083 	shmfd->shm_kmappings--;
1084 	VM_OBJECT_WUNLOCK(obj);
1085 	return (0);
1086 }
1087 
1088 static int
1089 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1090 {
1091 	const char *path, *pr_path;
1092 	struct shmfd *shmfd;
1093 	size_t pr_pathlen;
1094 
1095 	kif->kf_type = KF_TYPE_SHM;
1096 	shmfd = fp->f_data;
1097 
1098 	mtx_lock(&shm_timestamp_lock);
1099 	kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;	/* XXX */
1100 	mtx_unlock(&shm_timestamp_lock);
1101 	kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1102 	if (shmfd->shm_path != NULL) {
1103 		sx_slock(&shm_dict_lock);
1104 		if (shmfd->shm_path != NULL) {
1105 			path = shmfd->shm_path;
1106 			pr_path = curthread->td_ucred->cr_prison->pr_path;
1107 			if (strcmp(pr_path, "/") != 0) {
1108 				/* Return the jail-rooted pathname. */
1109 				pr_pathlen = strlen(pr_path);
1110 				if (strncmp(path, pr_path, pr_pathlen) == 0 &&
1111 				    path[pr_pathlen] == '/')
1112 					path += pr_pathlen;
1113 			}
1114 			strlcpy(kif->kf_path, path, sizeof(kif->kf_path));
1115 		}
1116 		sx_sunlock(&shm_dict_lock);
1117 	}
1118 	return (0);
1119 }
1120