xref: /freebsd/sys/kern/uipc_shm.c (revision 86aa9539fef591a363b06a0ebd3aa7a07f4c1579)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2006, 2011, 2016-2017 Robert N. M. Watson
5  * All rights reserved.
6  *
7  * Portions of this software were developed by BAE Systems, the University of
8  * Cambridge Computer Laboratory, and Memorial University under DARPA/AFRL
9  * contract FA8650-15-C-7558 ("CADETS"), as part of the DARPA Transparent
10  * Computing (TC) research program.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 /*
35  * Support for shared swap-backed anonymous memory objects via
36  * shm_open(2) and shm_unlink(2).  While most of the implementation is
37  * here, vm_mmap.c contains mapping logic changes.
38  *
39  * posixshmcontrol(1) allows users to inspect the state of the memory
40  * objects.  Per-uid swap resource limit controls total amount of
41  * memory that user can consume for anonymous objects, including
42  * shared.
43  */
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include "opt_capsicum.h"
49 #include "opt_ktrace.h"
50 
51 #include <sys/param.h>
52 #include <sys/capsicum.h>
53 #include <sys/conf.h>
54 #include <sys/fcntl.h>
55 #include <sys/file.h>
56 #include <sys/filedesc.h>
57 #include <sys/filio.h>
58 #include <sys/fnv_hash.h>
59 #include <sys/kernel.h>
60 #include <sys/limits.h>
61 #include <sys/uio.h>
62 #include <sys/signal.h>
63 #include <sys/jail.h>
64 #include <sys/ktrace.h>
65 #include <sys/lock.h>
66 #include <sys/malloc.h>
67 #include <sys/mman.h>
68 #include <sys/mutex.h>
69 #include <sys/priv.h>
70 #include <sys/proc.h>
71 #include <sys/refcount.h>
72 #include <sys/resourcevar.h>
73 #include <sys/rwlock.h>
74 #include <sys/sbuf.h>
75 #include <sys/stat.h>
76 #include <sys/syscallsubr.h>
77 #include <sys/sysctl.h>
78 #include <sys/sysproto.h>
79 #include <sys/systm.h>
80 #include <sys/sx.h>
81 #include <sys/time.h>
82 #include <sys/vnode.h>
83 #include <sys/unistd.h>
84 #include <sys/user.h>
85 
86 #include <security/audit/audit.h>
87 #include <security/mac/mac_framework.h>
88 
89 #include <vm/vm.h>
90 #include <vm/vm_param.h>
91 #include <vm/pmap.h>
92 #include <vm/vm_extern.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_kern.h>
95 #include <vm/vm_object.h>
96 #include <vm/vm_page.h>
97 #include <vm/vm_pageout.h>
98 #include <vm/vm_pager.h>
99 #include <vm/swap_pager.h>
100 
101 struct shm_mapping {
102 	char		*sm_path;
103 	Fnv32_t		sm_fnv;
104 	struct shmfd	*sm_shmfd;
105 	LIST_ENTRY(shm_mapping) sm_link;
106 };
107 
108 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor");
109 static LIST_HEAD(, shm_mapping) *shm_dictionary;
110 static struct sx shm_dict_lock;
111 static struct mtx shm_timestamp_lock;
112 static u_long shm_hash;
113 static struct unrhdr64 shm_ino_unr;
114 static dev_t shm_dev_ino;
115 
116 #define	SHM_HASH(fnv)	(&shm_dictionary[(fnv) & shm_hash])
117 
118 static void	shm_init(void *arg);
119 static void	shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd);
120 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv);
121 static int	shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred);
122 
123 static fo_rdwr_t	shm_read;
124 static fo_rdwr_t	shm_write;
125 static fo_truncate_t	shm_truncate;
126 static fo_ioctl_t	shm_ioctl;
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 = shm_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_wire(m);
212 	vm_page_unlock(m);
213 	VM_OBJECT_WUNLOCK(obj);
214 	error = uiomove_fromphys(&m, offset, tlen, uio);
215 	if (uio->uio_rw == UIO_WRITE && error == 0) {
216 		VM_OBJECT_WLOCK(obj);
217 		vm_page_dirty(m);
218 		vm_pager_page_unswapped(m);
219 		VM_OBJECT_WUNLOCK(obj);
220 	}
221 	vm_page_lock(m);
222 	vm_page_unwire(m, PQ_ACTIVE);
223 	vm_page_unlock(m);
224 
225 	return (error);
226 }
227 
228 int
229 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
230 {
231 	ssize_t resid;
232 	size_t len;
233 	int error;
234 
235 	error = 0;
236 	while ((resid = uio->uio_resid) > 0) {
237 		if (obj_size <= uio->uio_offset)
238 			break;
239 		len = MIN(obj_size - uio->uio_offset, resid);
240 		if (len == 0)
241 			break;
242 		error = uiomove_object_page(obj, len, uio);
243 		if (error != 0 || resid == uio->uio_resid)
244 			break;
245 	}
246 	return (error);
247 }
248 
249 static int
250 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
251 {
252 	struct shmfd *shmfd;
253 	off_t foffset;
254 	int error;
255 
256 	shmfd = fp->f_data;
257 	foffset = foffset_lock(fp, 0);
258 	error = 0;
259 	switch (whence) {
260 	case L_INCR:
261 		if (foffset < 0 ||
262 		    (offset > 0 && foffset > OFF_MAX - offset)) {
263 			error = EOVERFLOW;
264 			break;
265 		}
266 		offset += foffset;
267 		break;
268 	case L_XTND:
269 		if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
270 			error = EOVERFLOW;
271 			break;
272 		}
273 		offset += shmfd->shm_size;
274 		break;
275 	case L_SET:
276 		break;
277 	default:
278 		error = EINVAL;
279 	}
280 	if (error == 0) {
281 		if (offset < 0 || offset > shmfd->shm_size)
282 			error = EINVAL;
283 		else
284 			td->td_uretoff.tdu_off = offset;
285 	}
286 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
287 	return (error);
288 }
289 
290 static int
291 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
292     int flags, struct thread *td)
293 {
294 	struct shmfd *shmfd;
295 	void *rl_cookie;
296 	int error;
297 
298 	shmfd = fp->f_data;
299 #ifdef MAC
300 	error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
301 	if (error)
302 		return (error);
303 #endif
304 	foffset_lock_uio(fp, uio, flags);
305 	rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset,
306 	    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
307 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
308 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
309 	foffset_unlock_uio(fp, uio, flags);
310 	return (error);
311 }
312 
313 static int
314 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
315     int flags, struct thread *td)
316 {
317 	struct shmfd *shmfd;
318 	void *rl_cookie;
319 	int error;
320 
321 	shmfd = fp->f_data;
322 #ifdef MAC
323 	error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
324 	if (error)
325 		return (error);
326 #endif
327 	foffset_lock_uio(fp, uio, flags);
328 	if ((flags & FOF_OFFSET) == 0) {
329 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
330 		    &shmfd->shm_mtx);
331 	} else {
332 		rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset,
333 		    uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
334 	}
335 
336 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
337 	rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
338 	foffset_unlock_uio(fp, uio, flags);
339 	return (error);
340 }
341 
342 static int
343 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
344     struct thread *td)
345 {
346 	struct shmfd *shmfd;
347 #ifdef MAC
348 	int error;
349 #endif
350 
351 	shmfd = fp->f_data;
352 #ifdef MAC
353 	error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
354 	if (error)
355 		return (error);
356 #endif
357 	return (shm_dotruncate(shmfd, length));
358 }
359 
360 int
361 shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
362     struct thread *td)
363 {
364 
365 	switch (com) {
366 	case FIONBIO:
367 	case FIOASYNC:
368 		/*
369 		 * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work,
370 		 * just like it would on an unlinked regular file
371 		 */
372 		return (0);
373 	default:
374 		return (ENOTTY);
375 	}
376 }
377 
378 static int
379 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
380     struct thread *td)
381 {
382 	struct shmfd *shmfd;
383 #ifdef MAC
384 	int error;
385 #endif
386 
387 	shmfd = fp->f_data;
388 
389 #ifdef MAC
390 	error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
391 	if (error)
392 		return (error);
393 #endif
394 
395 	/*
396 	 * Attempt to return sanish values for fstat() on a memory file
397 	 * descriptor.
398 	 */
399 	bzero(sb, sizeof(*sb));
400 	sb->st_blksize = PAGE_SIZE;
401 	sb->st_size = shmfd->shm_size;
402 	sb->st_blocks = howmany(sb->st_size, sb->st_blksize);
403 	mtx_lock(&shm_timestamp_lock);
404 	sb->st_atim = shmfd->shm_atime;
405 	sb->st_ctim = shmfd->shm_ctime;
406 	sb->st_mtim = shmfd->shm_mtime;
407 	sb->st_birthtim = shmfd->shm_birthtime;
408 	sb->st_mode = S_IFREG | shmfd->shm_mode;		/* XXX */
409 	sb->st_uid = shmfd->shm_uid;
410 	sb->st_gid = shmfd->shm_gid;
411 	mtx_unlock(&shm_timestamp_lock);
412 	sb->st_dev = shm_dev_ino;
413 	sb->st_ino = shmfd->shm_ino;
414 	sb->st_nlink = shmfd->shm_object->ref_count;
415 
416 	return (0);
417 }
418 
419 static int
420 shm_close(struct file *fp, struct thread *td)
421 {
422 	struct shmfd *shmfd;
423 
424 	shmfd = fp->f_data;
425 	fp->f_data = NULL;
426 	shm_drop(shmfd);
427 
428 	return (0);
429 }
430 
431 int
432 shm_dotruncate(struct shmfd *shmfd, off_t length)
433 {
434 	vm_object_t object;
435 	vm_page_t m;
436 	vm_pindex_t idx, nobjsize;
437 	vm_ooffset_t delta;
438 	int base, rv;
439 
440 	KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
441 	object = shmfd->shm_object;
442 	VM_OBJECT_WLOCK(object);
443 	if (length == shmfd->shm_size) {
444 		VM_OBJECT_WUNLOCK(object);
445 		return (0);
446 	}
447 	nobjsize = OFF_TO_IDX(length + PAGE_MASK);
448 
449 	/* Are we shrinking?  If so, trim the end. */
450 	if (length < shmfd->shm_size) {
451 		/*
452 		 * Disallow any requests to shrink the size if this
453 		 * object is mapped into the kernel.
454 		 */
455 		if (shmfd->shm_kmappings > 0) {
456 			VM_OBJECT_WUNLOCK(object);
457 			return (EBUSY);
458 		}
459 
460 		/*
461 		 * Zero the truncated part of the last page.
462 		 */
463 		base = length & PAGE_MASK;
464 		if (base != 0) {
465 			idx = OFF_TO_IDX(length);
466 retry:
467 			m = vm_page_lookup(object, idx);
468 			if (m != NULL) {
469 				if (vm_page_sleep_if_busy(m, "shmtrc"))
470 					goto retry;
471 			} else if (vm_pager_has_page(object, idx, NULL, NULL)) {
472 				m = vm_page_alloc(object, idx,
473 				    VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL);
474 				if (m == NULL)
475 					goto retry;
476 				rv = vm_pager_get_pages(object, &m, 1, NULL,
477 				    NULL);
478 				vm_page_lock(m);
479 				if (rv == VM_PAGER_OK) {
480 					/*
481 					 * Since the page was not resident,
482 					 * and therefore not recently
483 					 * accessed, immediately enqueue it
484 					 * for asynchronous laundering.  The
485 					 * current operation is not regarded
486 					 * as an access.
487 					 */
488 					vm_page_launder(m);
489 					vm_page_unlock(m);
490 					vm_page_xunbusy(m);
491 				} else {
492 					vm_page_free(m);
493 					vm_page_unlock(m);
494 					VM_OBJECT_WUNLOCK(object);
495 					return (EIO);
496 				}
497 			}
498 			if (m != NULL) {
499 				pmap_zero_page_area(m, base, PAGE_SIZE - base);
500 				KASSERT(m->valid == VM_PAGE_BITS_ALL,
501 				    ("shm_dotruncate: page %p is invalid", m));
502 				vm_page_dirty(m);
503 				vm_pager_page_unswapped(m);
504 			}
505 		}
506 		delta = IDX_TO_OFF(object->size - nobjsize);
507 
508 		/* Toss in memory pages. */
509 		if (nobjsize < object->size)
510 			vm_object_page_remove(object, nobjsize, object->size,
511 			    0);
512 
513 		/* Toss pages from swap. */
514 		if (object->type == OBJT_SWAP)
515 			swap_pager_freespace(object, nobjsize, delta);
516 
517 		/* Free the swap accounted for shm */
518 		swap_release_by_cred(delta, object->cred);
519 		object->charge -= delta;
520 	} else {
521 		/* Try to reserve additional swap space. */
522 		delta = IDX_TO_OFF(nobjsize - object->size);
523 		if (!swap_reserve_by_cred(delta, object->cred)) {
524 			VM_OBJECT_WUNLOCK(object);
525 			return (ENOMEM);
526 		}
527 		object->charge += delta;
528 	}
529 	shmfd->shm_size = length;
530 	mtx_lock(&shm_timestamp_lock);
531 	vfs_timestamp(&shmfd->shm_ctime);
532 	shmfd->shm_mtime = shmfd->shm_ctime;
533 	mtx_unlock(&shm_timestamp_lock);
534 	object->size = nobjsize;
535 	VM_OBJECT_WUNLOCK(object);
536 	return (0);
537 }
538 
539 /*
540  * shmfd object management including creation and reference counting
541  * routines.
542  */
543 struct shmfd *
544 shm_alloc(struct ucred *ucred, mode_t mode)
545 {
546 	struct shmfd *shmfd;
547 
548 	shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
549 	shmfd->shm_size = 0;
550 	shmfd->shm_uid = ucred->cr_uid;
551 	shmfd->shm_gid = ucred->cr_gid;
552 	shmfd->shm_mode = mode;
553 	shmfd->shm_object = vm_pager_allocate(OBJT_DEFAULT, NULL,
554 	    shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
555 	KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
556 	shmfd->shm_object->pg_color = 0;
557 	VM_OBJECT_WLOCK(shmfd->shm_object);
558 	vm_object_clear_flag(shmfd->shm_object, OBJ_ONEMAPPING);
559 	vm_object_set_flag(shmfd->shm_object, OBJ_COLORED | OBJ_NOSPLIT);
560 	VM_OBJECT_WUNLOCK(shmfd->shm_object);
561 	vfs_timestamp(&shmfd->shm_birthtime);
562 	shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
563 	    shmfd->shm_birthtime;
564 	shmfd->shm_ino = alloc_unr64(&shm_ino_unr);
565 	refcount_init(&shmfd->shm_refs, 1);
566 	mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
567 	rangelock_init(&shmfd->shm_rl);
568 #ifdef MAC
569 	mac_posixshm_init(shmfd);
570 	mac_posixshm_create(ucred, shmfd);
571 #endif
572 
573 	return (shmfd);
574 }
575 
576 struct shmfd *
577 shm_hold(struct shmfd *shmfd)
578 {
579 
580 	refcount_acquire(&shmfd->shm_refs);
581 	return (shmfd);
582 }
583 
584 void
585 shm_drop(struct shmfd *shmfd)
586 {
587 
588 	if (refcount_release(&shmfd->shm_refs)) {
589 #ifdef MAC
590 		mac_posixshm_destroy(shmfd);
591 #endif
592 		rangelock_destroy(&shmfd->shm_rl);
593 		mtx_destroy(&shmfd->shm_mtx);
594 		vm_object_deallocate(shmfd->shm_object);
595 		free(shmfd, M_SHMFD);
596 	}
597 }
598 
599 /*
600  * Determine if the credentials have sufficient permissions for a
601  * specified combination of FREAD and FWRITE.
602  */
603 int
604 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
605 {
606 	accmode_t accmode;
607 	int error;
608 
609 	accmode = 0;
610 	if (flags & FREAD)
611 		accmode |= VREAD;
612 	if (flags & FWRITE)
613 		accmode |= VWRITE;
614 	mtx_lock(&shm_timestamp_lock);
615 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
616 	    accmode, ucred, NULL);
617 	mtx_unlock(&shm_timestamp_lock);
618 	return (error);
619 }
620 
621 /*
622  * Dictionary management.  We maintain an in-kernel dictionary to map
623  * paths to shmfd objects.  We use the FNV hash on the path to store
624  * the mappings in a hash table.
625  */
626 static void
627 shm_init(void *arg)
628 {
629 
630 	mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
631 	sx_init(&shm_dict_lock, "shm dictionary");
632 	shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
633 	new_unrhdr64(&shm_ino_unr, 1);
634 	shm_dev_ino = devfs_alloc_cdp_inode();
635 	KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
636 }
637 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
638 
639 static struct shmfd *
640 shm_lookup(char *path, Fnv32_t fnv)
641 {
642 	struct shm_mapping *map;
643 
644 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
645 		if (map->sm_fnv != fnv)
646 			continue;
647 		if (strcmp(map->sm_path, path) == 0)
648 			return (map->sm_shmfd);
649 	}
650 
651 	return (NULL);
652 }
653 
654 static void
655 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
656 {
657 	struct shm_mapping *map;
658 
659 	map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
660 	map->sm_path = path;
661 	map->sm_fnv = fnv;
662 	map->sm_shmfd = shm_hold(shmfd);
663 	shmfd->shm_path = path;
664 	LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
665 }
666 
667 static int
668 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
669 {
670 	struct shm_mapping *map;
671 	int error;
672 
673 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
674 		if (map->sm_fnv != fnv)
675 			continue;
676 		if (strcmp(map->sm_path, path) == 0) {
677 #ifdef MAC
678 			error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
679 			if (error)
680 				return (error);
681 #endif
682 			error = shm_access(map->sm_shmfd, ucred,
683 			    FREAD | FWRITE);
684 			if (error)
685 				return (error);
686 			map->sm_shmfd->shm_path = NULL;
687 			LIST_REMOVE(map, sm_link);
688 			shm_drop(map->sm_shmfd);
689 			free(map->sm_path, M_SHMFD);
690 			free(map, M_SHMFD);
691 			return (0);
692 		}
693 	}
694 
695 	return (ENOENT);
696 }
697 
698 int
699 kern_shm_open(struct thread *td, const char *userpath, int flags, mode_t mode,
700     struct filecaps *fcaps)
701 {
702 	struct filedesc *fdp;
703 	struct shmfd *shmfd;
704 	struct file *fp;
705 	char *path;
706 	const char *pr_path;
707 	size_t pr_pathlen;
708 	Fnv32_t fnv;
709 	mode_t cmode;
710 	int fd, error;
711 
712 #ifdef CAPABILITY_MODE
713 	/*
714 	 * shm_open(2) is only allowed for anonymous objects.
715 	 */
716 	if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON))
717 		return (ECAPMODE);
718 #endif
719 
720 	AUDIT_ARG_FFLAGS(flags);
721 	AUDIT_ARG_MODE(mode);
722 
723 	if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
724 		return (EINVAL);
725 
726 	if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
727 		return (EINVAL);
728 
729 	fdp = td->td_proc->p_fd;
730 	cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS;
731 
732 	/*
733 	 * shm_open(2) created shm should always have O_CLOEXEC set, as mandated
734 	 * by POSIX.  We allow it to be unset here so that an in-kernel
735 	 * interface may be written as a thin layer around shm, optionally not
736 	 * setting CLOEXEC.  For shm_open(2), O_CLOEXEC is set unconditionally
737 	 * in sys_shm_open() to keep this implementation compliant.
738 	 */
739 	error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps);
740 	if (error)
741 		return (error);
742 
743 	/* A SHM_ANON path pointer creates an anonymous object. */
744 	if (userpath == SHM_ANON) {
745 		/* A read-only anonymous object is pointless. */
746 		if ((flags & O_ACCMODE) == O_RDONLY) {
747 			fdclose(td, fp, fd);
748 			fdrop(fp, td);
749 			return (EINVAL);
750 		}
751 		shmfd = shm_alloc(td->td_ucred, cmode);
752 	} else {
753 		path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
754 		pr_path = td->td_ucred->cr_prison->pr_path;
755 
756 		/* Construct a full pathname for jailed callers. */
757 		pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
758 		    : strlcpy(path, pr_path, MAXPATHLEN);
759 		error = copyinstr(userpath, path + pr_pathlen,
760 		    MAXPATHLEN - pr_pathlen, NULL);
761 #ifdef KTRACE
762 		if (error == 0 && KTRPOINT(curthread, KTR_NAMEI))
763 			ktrnamei(path);
764 #endif
765 		/* Require paths to start with a '/' character. */
766 		if (error == 0 && path[pr_pathlen] != '/')
767 			error = EINVAL;
768 		if (error) {
769 			fdclose(td, fp, fd);
770 			fdrop(fp, td);
771 			free(path, M_SHMFD);
772 			return (error);
773 		}
774 
775 		AUDIT_ARG_UPATH1_CANON(path);
776 		fnv = fnv_32_str(path, FNV1_32_INIT);
777 		sx_xlock(&shm_dict_lock);
778 		shmfd = shm_lookup(path, fnv);
779 		if (shmfd == NULL) {
780 			/* Object does not yet exist, create it if requested. */
781 			if (flags & O_CREAT) {
782 #ifdef MAC
783 				error = mac_posixshm_check_create(td->td_ucred,
784 				    path);
785 				if (error == 0) {
786 #endif
787 					shmfd = shm_alloc(td->td_ucred, cmode);
788 					shm_insert(path, fnv, shmfd);
789 #ifdef MAC
790 				}
791 #endif
792 			} else {
793 				free(path, M_SHMFD);
794 				error = ENOENT;
795 			}
796 		} else {
797 			/*
798 			 * Object already exists, obtain a new
799 			 * reference if requested and permitted.
800 			 */
801 			free(path, M_SHMFD);
802 			if ((flags & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
803 				error = EEXIST;
804 			else {
805 #ifdef MAC
806 				error = mac_posixshm_check_open(td->td_ucred,
807 				    shmfd, FFLAGS(flags & O_ACCMODE));
808 				if (error == 0)
809 #endif
810 				error = shm_access(shmfd, td->td_ucred,
811 				    FFLAGS(flags & O_ACCMODE));
812 			}
813 
814 			/*
815 			 * Truncate the file back to zero length if
816 			 * O_TRUNC was specified and the object was
817 			 * opened with read/write.
818 			 */
819 			if (error == 0 &&
820 			    (flags & (O_ACCMODE | O_TRUNC)) ==
821 			    (O_RDWR | O_TRUNC)) {
822 #ifdef MAC
823 				error = mac_posixshm_check_truncate(
824 					td->td_ucred, fp->f_cred, shmfd);
825 				if (error == 0)
826 #endif
827 					shm_dotruncate(shmfd, 0);
828 			}
829 			if (error == 0)
830 				shm_hold(shmfd);
831 		}
832 		sx_xunlock(&shm_dict_lock);
833 
834 		if (error) {
835 			fdclose(td, fp, fd);
836 			fdrop(fp, td);
837 			return (error);
838 		}
839 	}
840 
841 	finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
842 
843 	td->td_retval[0] = fd;
844 	fdrop(fp, td);
845 
846 	return (0);
847 }
848 
849 /* System calls. */
850 int
851 sys_shm_open(struct thread *td, struct shm_open_args *uap)
852 {
853 
854 	return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC, uap->mode,
855 	    NULL));
856 }
857 
858 int
859 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
860 {
861 	char *path;
862 	const char *pr_path;
863 	size_t pr_pathlen;
864 	Fnv32_t fnv;
865 	int error;
866 
867 	path = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
868 	pr_path = td->td_ucred->cr_prison->pr_path;
869 	pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
870 	    : strlcpy(path, pr_path, MAXPATHLEN);
871 	error = copyinstr(uap->path, path + pr_pathlen, MAXPATHLEN - pr_pathlen,
872 	    NULL);
873 	if (error) {
874 		free(path, M_TEMP);
875 		return (error);
876 	}
877 #ifdef KTRACE
878 	if (KTRPOINT(curthread, KTR_NAMEI))
879 		ktrnamei(path);
880 #endif
881 	AUDIT_ARG_UPATH1_CANON(path);
882 	fnv = fnv_32_str(path, FNV1_32_INIT);
883 	sx_xlock(&shm_dict_lock);
884 	error = shm_remove(path, fnv, td->td_ucred);
885 	sx_xunlock(&shm_dict_lock);
886 	free(path, M_TEMP);
887 
888 	return (error);
889 }
890 
891 int
892 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
893     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
894     vm_ooffset_t foff, struct thread *td)
895 {
896 	struct shmfd *shmfd;
897 	vm_prot_t maxprot;
898 	int error;
899 
900 	shmfd = fp->f_data;
901 	maxprot = VM_PROT_NONE;
902 
903 	/* FREAD should always be set. */
904 	if ((fp->f_flag & FREAD) != 0)
905 		maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
906 	if ((fp->f_flag & FWRITE) != 0)
907 		maxprot |= VM_PROT_WRITE;
908 
909 	/* Don't permit shared writable mappings on read-only descriptors. */
910 	if ((flags & MAP_SHARED) != 0 &&
911 	    (maxprot & VM_PROT_WRITE) == 0 &&
912 	    (prot & VM_PROT_WRITE) != 0)
913 		return (EACCES);
914 	maxprot &= cap_maxprot;
915 
916 	/* See comment in vn_mmap(). */
917 	if (
918 #ifdef _LP64
919 	    objsize > OFF_MAX ||
920 #endif
921 	    foff < 0 || foff > OFF_MAX - objsize)
922 		return (EINVAL);
923 
924 #ifdef MAC
925 	error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
926 	if (error != 0)
927 		return (error);
928 #endif
929 
930 	mtx_lock(&shm_timestamp_lock);
931 	vfs_timestamp(&shmfd->shm_atime);
932 	mtx_unlock(&shm_timestamp_lock);
933 	vm_object_reference(shmfd->shm_object);
934 
935 	error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
936 	    shmfd->shm_object, foff, FALSE, td);
937 	if (error != 0)
938 		vm_object_deallocate(shmfd->shm_object);
939 	return (error);
940 }
941 
942 static int
943 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
944     struct thread *td)
945 {
946 	struct shmfd *shmfd;
947 	int error;
948 
949 	error = 0;
950 	shmfd = fp->f_data;
951 	mtx_lock(&shm_timestamp_lock);
952 	/*
953 	 * SUSv4 says that x bits of permission need not be affected.
954 	 * Be consistent with our shm_open there.
955 	 */
956 #ifdef MAC
957 	error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
958 	if (error != 0)
959 		goto out;
960 #endif
961 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid,
962 	    shmfd->shm_gid, VADMIN, active_cred, NULL);
963 	if (error != 0)
964 		goto out;
965 	shmfd->shm_mode = mode & ACCESSPERMS;
966 out:
967 	mtx_unlock(&shm_timestamp_lock);
968 	return (error);
969 }
970 
971 static int
972 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
973     struct thread *td)
974 {
975 	struct shmfd *shmfd;
976 	int error;
977 
978 	error = 0;
979 	shmfd = fp->f_data;
980 	mtx_lock(&shm_timestamp_lock);
981 #ifdef MAC
982 	error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
983 	if (error != 0)
984 		goto out;
985 #endif
986 	if (uid == (uid_t)-1)
987 		uid = shmfd->shm_uid;
988 	if (gid == (gid_t)-1)
989                  gid = shmfd->shm_gid;
990 	if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
991 	    (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
992 	    (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN)))
993 		goto out;
994 	shmfd->shm_uid = uid;
995 	shmfd->shm_gid = gid;
996 out:
997 	mtx_unlock(&shm_timestamp_lock);
998 	return (error);
999 }
1000 
1001 /*
1002  * Helper routines to allow the backing object of a shared memory file
1003  * descriptor to be mapped in the kernel.
1004  */
1005 int
1006 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
1007 {
1008 	struct shmfd *shmfd;
1009 	vm_offset_t kva, ofs;
1010 	vm_object_t obj;
1011 	int rv;
1012 
1013 	if (fp->f_type != DTYPE_SHM)
1014 		return (EINVAL);
1015 	shmfd = fp->f_data;
1016 	obj = shmfd->shm_object;
1017 	VM_OBJECT_WLOCK(obj);
1018 	/*
1019 	 * XXXRW: This validation is probably insufficient, and subject to
1020 	 * sign errors.  It should be fixed.
1021 	 */
1022 	if (offset >= shmfd->shm_size ||
1023 	    offset + size > round_page(shmfd->shm_size)) {
1024 		VM_OBJECT_WUNLOCK(obj);
1025 		return (EINVAL);
1026 	}
1027 
1028 	shmfd->shm_kmappings++;
1029 	vm_object_reference_locked(obj);
1030 	VM_OBJECT_WUNLOCK(obj);
1031 
1032 	/* Map the object into the kernel_map and wire it. */
1033 	kva = vm_map_min(kernel_map);
1034 	ofs = offset & PAGE_MASK;
1035 	offset = trunc_page(offset);
1036 	size = round_page(size + ofs);
1037 	rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1038 	    VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1039 	    VM_PROT_READ | VM_PROT_WRITE, 0);
1040 	if (rv == KERN_SUCCESS) {
1041 		rv = vm_map_wire(kernel_map, kva, kva + size,
1042 		    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1043 		if (rv == KERN_SUCCESS) {
1044 			*memp = (void *)(kva + ofs);
1045 			return (0);
1046 		}
1047 		vm_map_remove(kernel_map, kva, kva + size);
1048 	} else
1049 		vm_object_deallocate(obj);
1050 
1051 	/* On failure, drop our mapping reference. */
1052 	VM_OBJECT_WLOCK(obj);
1053 	shmfd->shm_kmappings--;
1054 	VM_OBJECT_WUNLOCK(obj);
1055 
1056 	return (vm_mmap_to_errno(rv));
1057 }
1058 
1059 /*
1060  * We require the caller to unmap the entire entry.  This allows us to
1061  * safely decrement shm_kmappings when a mapping is removed.
1062  */
1063 int
1064 shm_unmap(struct file *fp, void *mem, size_t size)
1065 {
1066 	struct shmfd *shmfd;
1067 	vm_map_entry_t entry;
1068 	vm_offset_t kva, ofs;
1069 	vm_object_t obj;
1070 	vm_pindex_t pindex;
1071 	vm_prot_t prot;
1072 	boolean_t wired;
1073 	vm_map_t map;
1074 	int rv;
1075 
1076 	if (fp->f_type != DTYPE_SHM)
1077 		return (EINVAL);
1078 	shmfd = fp->f_data;
1079 	kva = (vm_offset_t)mem;
1080 	ofs = kva & PAGE_MASK;
1081 	kva = trunc_page(kva);
1082 	size = round_page(size + ofs);
1083 	map = kernel_map;
1084 	rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1085 	    &obj, &pindex, &prot, &wired);
1086 	if (rv != KERN_SUCCESS)
1087 		return (EINVAL);
1088 	if (entry->start != kva || entry->end != kva + size) {
1089 		vm_map_lookup_done(map, entry);
1090 		return (EINVAL);
1091 	}
1092 	vm_map_lookup_done(map, entry);
1093 	if (obj != shmfd->shm_object)
1094 		return (EINVAL);
1095 	vm_map_remove(map, kva, kva + size);
1096 	VM_OBJECT_WLOCK(obj);
1097 	KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1098 	shmfd->shm_kmappings--;
1099 	VM_OBJECT_WUNLOCK(obj);
1100 	return (0);
1101 }
1102 
1103 static int
1104 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list)
1105 {
1106 	const char *path, *pr_path;
1107 	size_t pr_pathlen;
1108 	bool visible;
1109 
1110 	sx_assert(&shm_dict_lock, SA_LOCKED);
1111 	kif->kf_type = KF_TYPE_SHM;
1112 	kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;
1113 	kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1114 	if (shmfd->shm_path != NULL) {
1115 		if (shmfd->shm_path != NULL) {
1116 			path = shmfd->shm_path;
1117 			pr_path = curthread->td_ucred->cr_prison->pr_path;
1118 			if (strcmp(pr_path, "/") != 0) {
1119 				/* Return the jail-rooted pathname. */
1120 				pr_pathlen = strlen(pr_path);
1121 				visible = strncmp(path, pr_path, pr_pathlen)
1122 				    == 0 && path[pr_pathlen] == '/';
1123 				if (list && !visible)
1124 					return (EPERM);
1125 				if (visible)
1126 					path += pr_pathlen;
1127 			}
1128 			strlcpy(kif->kf_path, path, sizeof(kif->kf_path));
1129 		}
1130 	}
1131 	return (0);
1132 }
1133 
1134 static int
1135 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif,
1136     struct filedesc *fdp __unused)
1137 {
1138 	int res;
1139 
1140 	sx_slock(&shm_dict_lock);
1141 	res = shm_fill_kinfo_locked(fp->f_data, kif, false);
1142 	sx_sunlock(&shm_dict_lock);
1143 	return (res);
1144 }
1145 
1146 static int
1147 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS)
1148 {
1149 	struct shm_mapping *shmm;
1150 	struct sbuf sb;
1151 	struct kinfo_file kif;
1152 	u_long i;
1153 	ssize_t curlen;
1154 	int error, error2;
1155 
1156 	sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req);
1157 	sbuf_clear_flags(&sb, SBUF_INCLUDENUL);
1158 	curlen = 0;
1159 	error = 0;
1160 	sx_slock(&shm_dict_lock);
1161 	for (i = 0; i < shm_hash + 1; i++) {
1162 		LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) {
1163 			error = shm_fill_kinfo_locked(shmm->sm_shmfd,
1164 			    &kif, true);
1165 			if (error == EPERM)
1166 				continue;
1167 			if (error != 0)
1168 				break;
1169 			pack_kinfo(&kif);
1170 			if (req->oldptr != NULL &&
1171 			    kif.kf_structsize + curlen > req->oldlen)
1172 				break;
1173 			error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ?
1174 			    0 : ENOMEM;
1175 			if (error != 0)
1176 				break;
1177 			curlen += kif.kf_structsize;
1178 		}
1179 	}
1180 	sx_sunlock(&shm_dict_lock);
1181 	error2 = sbuf_finish(&sb);
1182 	sbuf_delete(&sb);
1183 	return (error != 0 ? error : error2);
1184 }
1185 
1186 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list,
1187     CTLFLAG_RD | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE,
1188     NULL, 0, sysctl_posix_shm_list, "",
1189     "POSIX SHM list");
1190