xref: /freebsd-13-stable/sys/kern/uipc_shm.c (revision 7d2b98e645ce9454b84f7d1d193d98d0880c4627)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2006, 2011, 2016-2017 Robert N. M. Watson
5  * Copyright 2020 The FreeBSD Foundation
6  * All rights reserved.
7  *
8  * Portions of this software were developed by BAE Systems, the University of
9  * Cambridge Computer Laboratory, and Memorial University under DARPA/AFRL
10  * contract FA8650-15-C-7558 ("CADETS"), as part of the DARPA Transparent
11  * Computing (TC) research program.
12  *
13  * Portions of this software were developed by Konstantin Belousov
14  * under sponsorship from the FreeBSD Foundation.
15  *
16  * Redistribution and use in source and binary forms, with or without
17  * modification, are permitted provided that the following conditions
18  * are met:
19  * 1. Redistributions of source code must retain the above copyright
20  *    notice, this list of conditions and the following disclaimer.
21  * 2. Redistributions in binary form must reproduce the above copyright
22  *    notice, this list of conditions and the following disclaimer in the
23  *    documentation and/or other materials provided with the distribution.
24  *
25  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
26  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
27  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
28  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
29  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
30  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
31  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
32  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
33  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
34  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35  * SUCH DAMAGE.
36  */
37 
38 /*
39  * Support for shared swap-backed anonymous memory objects via
40  * shm_open(2), shm_rename(2), and shm_unlink(2).
41  * While most of the implementation is here, vm_mmap.c contains
42  * mapping logic changes.
43  *
44  * posixshmcontrol(1) allows users to inspect the state of the memory
45  * objects.  Per-uid swap resource limit controls total amount of
46  * memory that user can consume for anonymous objects, including
47  * shared.
48  */
49 
50 #include <sys/cdefs.h>
51 #include "opt_capsicum.h"
52 #include "opt_ktrace.h"
53 
54 #include <sys/param.h>
55 #include <sys/capsicum.h>
56 #include <sys/conf.h>
57 #include <sys/fcntl.h>
58 #include <sys/file.h>
59 #include <sys/filedesc.h>
60 #include <sys/filio.h>
61 #include <sys/fnv_hash.h>
62 #include <sys/kernel.h>
63 #include <sys/limits.h>
64 #include <sys/uio.h>
65 #include <sys/signal.h>
66 #include <sys/jail.h>
67 #include <sys/ktrace.h>
68 #include <sys/lock.h>
69 #include <sys/malloc.h>
70 #include <sys/mman.h>
71 #include <sys/mutex.h>
72 #include <sys/priv.h>
73 #include <sys/proc.h>
74 #include <sys/refcount.h>
75 #include <sys/resourcevar.h>
76 #include <sys/rwlock.h>
77 #include <sys/sbuf.h>
78 #include <sys/stat.h>
79 #include <sys/syscallsubr.h>
80 #include <sys/sysctl.h>
81 #include <sys/sysproto.h>
82 #include <sys/systm.h>
83 #include <sys/sx.h>
84 #include <sys/time.h>
85 #include <sys/vmmeter.h>
86 #include <sys/vnode.h>
87 #include <sys/unistd.h>
88 #include <sys/user.h>
89 
90 #include <security/audit/audit.h>
91 #include <security/mac/mac_framework.h>
92 
93 #include <vm/vm.h>
94 #include <vm/vm_param.h>
95 #include <vm/pmap.h>
96 #include <vm/vm_extern.h>
97 #include <vm/vm_map.h>
98 #include <vm/vm_kern.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_page.h>
101 #include <vm/vm_pageout.h>
102 #include <vm/vm_pager.h>
103 #include <vm/swap_pager.h>
104 
105 struct shm_mapping {
106 	char		*sm_path;
107 	Fnv32_t		sm_fnv;
108 	struct shmfd	*sm_shmfd;
109 	LIST_ENTRY(shm_mapping) sm_link;
110 };
111 
112 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor");
113 static LIST_HEAD(, shm_mapping) *shm_dictionary;
114 static struct sx shm_dict_lock;
115 static struct mtx shm_timestamp_lock;
116 static u_long shm_hash;
117 static struct unrhdr64 shm_ino_unr;
118 static dev_t shm_dev_ino;
119 
120 #define	SHM_HASH(fnv)	(&shm_dictionary[(fnv) & shm_hash])
121 
122 static void	shm_init(void *arg);
123 static void	shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd);
124 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv);
125 static int	shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred);
126 static void	shm_doremove(struct shm_mapping *map);
127 static int	shm_dotruncate_cookie(struct shmfd *shmfd, off_t length,
128     void *rl_cookie);
129 static int	shm_dotruncate_locked(struct shmfd *shmfd, off_t length,
130     void *rl_cookie);
131 static int	shm_copyin_path(struct thread *td, const char *userpath_in,
132     char **path_out);
133 
134 static fo_rdwr_t	shm_read;
135 static fo_rdwr_t	shm_write;
136 static fo_truncate_t	shm_truncate;
137 static fo_ioctl_t	shm_ioctl;
138 static fo_stat_t	shm_stat;
139 static fo_close_t	shm_close;
140 static fo_chmod_t	shm_chmod;
141 static fo_chown_t	shm_chown;
142 static fo_seek_t	shm_seek;
143 static fo_fill_kinfo_t	shm_fill_kinfo;
144 static fo_mmap_t	shm_mmap;
145 static fo_get_seals_t	shm_get_seals;
146 static fo_add_seals_t	shm_add_seals;
147 static fo_fallocate_t	shm_fallocate;
148 
149 /* File descriptor operations. */
150 struct fileops shm_ops = {
151 	.fo_read = shm_read,
152 	.fo_write = shm_write,
153 	.fo_truncate = shm_truncate,
154 	.fo_ioctl = shm_ioctl,
155 	.fo_poll = invfo_poll,
156 	.fo_kqfilter = invfo_kqfilter,
157 	.fo_stat = shm_stat,
158 	.fo_close = shm_close,
159 	.fo_chmod = shm_chmod,
160 	.fo_chown = shm_chown,
161 	.fo_sendfile = vn_sendfile,
162 	.fo_seek = shm_seek,
163 	.fo_fill_kinfo = shm_fill_kinfo,
164 	.fo_mmap = shm_mmap,
165 	.fo_get_seals = shm_get_seals,
166 	.fo_add_seals = shm_add_seals,
167 	.fo_fallocate = shm_fallocate,
168 	.fo_cmp = file_kcmp_generic,
169 	.fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE,
170 };
171 
172 FEATURE(posix_shm, "POSIX shared memory");
173 
174 static SYSCTL_NODE(_vm, OID_AUTO, largepages, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
175     "");
176 
177 static int largepage_reclaim_tries = 1;
178 SYSCTL_INT(_vm_largepages, OID_AUTO, reclaim_tries,
179     CTLFLAG_RWTUN, &largepage_reclaim_tries, 0,
180     "Number of contig reclaims before giving up for default alloc policy");
181 
182 #define	shm_rangelock_unlock(shmfd, cookie)				\
183 	rangelock_unlock(&(shmfd)->shm_rl, (cookie), &(shmfd)->shm_mtx)
184 #define	shm_rangelock_rlock(shmfd, start, end)				\
185 	rangelock_rlock(&(shmfd)->shm_rl, (start), (end), &(shmfd)->shm_mtx)
186 #define	shm_rangelock_tryrlock(shmfd, start, end)			\
187 	rangelock_tryrlock(&(shmfd)->shm_rl, (start), (end), &(shmfd)->shm_mtx)
188 #define	shm_rangelock_wlock(shmfd, start, end)				\
189 	rangelock_wlock(&(shmfd)->shm_rl, (start), (end), &(shmfd)->shm_mtx)
190 
191 static int
uiomove_object_page(vm_object_t obj,size_t len,struct uio * uio)192 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio)
193 {
194 	vm_page_t m;
195 	vm_pindex_t idx;
196 	size_t tlen;
197 	int error, offset, rv;
198 
199 	idx = OFF_TO_IDX(uio->uio_offset);
200 	offset = uio->uio_offset & PAGE_MASK;
201 	tlen = MIN(PAGE_SIZE - offset, len);
202 
203 	rv = vm_page_grab_valid_unlocked(&m, obj, idx,
204 	    VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOCREAT);
205 	if (rv == VM_PAGER_OK)
206 		goto found;
207 
208 	/*
209 	 * Read I/O without either a corresponding resident page or swap
210 	 * page: use zero_region.  This is intended to avoid instantiating
211 	 * pages on read from a sparse region.
212 	 */
213 	VM_OBJECT_WLOCK(obj);
214 	m = vm_page_lookup(obj, idx);
215 	if (uio->uio_rw == UIO_READ && m == NULL &&
216 	    !vm_pager_has_page(obj, idx, NULL, NULL)) {
217 		VM_OBJECT_WUNLOCK(obj);
218 		return (uiomove(__DECONST(void *, zero_region), tlen, uio));
219 	}
220 
221 	/*
222 	 * Although the tmpfs vnode lock is held here, it is
223 	 * nonetheless safe to sleep waiting for a free page.  The
224 	 * pageout daemon does not need to acquire the tmpfs vnode
225 	 * lock to page out tobj's pages because tobj is a OBJT_SWAP
226 	 * type object.
227 	 */
228 	rv = vm_page_grab_valid(&m, obj, idx,
229 	    VM_ALLOC_NORMAL | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY);
230 	if (rv != VM_PAGER_OK) {
231 		VM_OBJECT_WUNLOCK(obj);
232 		if (bootverbose) {
233 			printf("uiomove_object: vm_obj %p idx %jd "
234 			    "pager error %d\n", obj, idx, rv);
235 		}
236 		return (rv == VM_PAGER_AGAIN ? ENOSPC : EIO);
237 	}
238 	VM_OBJECT_WUNLOCK(obj);
239 
240 found:
241 	error = uiomove_fromphys(&m, offset, tlen, uio);
242 	if (uio->uio_rw == UIO_WRITE && error == 0)
243 		vm_page_set_dirty(m);
244 	vm_page_activate(m);
245 	vm_page_sunbusy(m);
246 
247 	return (error);
248 }
249 
250 int
uiomove_object(vm_object_t obj,off_t obj_size,struct uio * uio)251 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
252 {
253 	ssize_t resid;
254 	size_t len;
255 	int error;
256 
257 	error = 0;
258 	while ((resid = uio->uio_resid) > 0) {
259 		if (obj_size <= uio->uio_offset)
260 			break;
261 		len = MIN(obj_size - uio->uio_offset, resid);
262 		if (len == 0)
263 			break;
264 		error = uiomove_object_page(obj, len, uio);
265 		if (error != 0 || resid == uio->uio_resid)
266 			break;
267 	}
268 	return (error);
269 }
270 
271 static u_long count_largepages[MAXPAGESIZES];
272 
273 static int
shm_largepage_phys_populate(vm_object_t object,vm_pindex_t pidx,int fault_type,vm_prot_t max_prot,vm_pindex_t * first,vm_pindex_t * last)274 shm_largepage_phys_populate(vm_object_t object, vm_pindex_t pidx,
275     int fault_type, vm_prot_t max_prot, vm_pindex_t *first, vm_pindex_t *last)
276 {
277 	vm_page_t m __diagused;
278 	int psind;
279 
280 	psind = object->un_pager.phys.data_val;
281 	if (psind == 0 || pidx >= object->size)
282 		return (VM_PAGER_FAIL);
283 	*first = rounddown2(pidx, pagesizes[psind] / PAGE_SIZE);
284 
285 	/*
286 	 * We only busy the first page in the superpage run.  It is
287 	 * useless to busy whole run since we only remove full
288 	 * superpage, and it takes too long to busy e.g. 512 * 512 ==
289 	 * 262144 pages constituing 1G amd64 superage.
290 	 */
291 	m = vm_page_grab(object, *first, VM_ALLOC_NORMAL | VM_ALLOC_NOCREAT);
292 	MPASS(m != NULL);
293 
294 	*last = *first + atop(pagesizes[psind]) - 1;
295 	return (VM_PAGER_OK);
296 }
297 
298 static boolean_t
shm_largepage_phys_haspage(vm_object_t object,vm_pindex_t pindex,int * before,int * after)299 shm_largepage_phys_haspage(vm_object_t object, vm_pindex_t pindex,
300     int *before, int *after)
301 {
302 	int psind;
303 
304 	psind = object->un_pager.phys.data_val;
305 	if (psind == 0 || pindex >= object->size)
306 		return (FALSE);
307 	if (before != NULL) {
308 		*before = pindex - rounddown2(pindex, pagesizes[psind] /
309 		    PAGE_SIZE);
310 	}
311 	if (after != NULL) {
312 		*after = roundup2(pindex, pagesizes[psind] / PAGE_SIZE) -
313 		    pindex;
314 	}
315 	return (TRUE);
316 }
317 
318 static void
shm_largepage_phys_ctor(vm_object_t object,vm_prot_t prot,vm_ooffset_t foff,struct ucred * cred)319 shm_largepage_phys_ctor(vm_object_t object, vm_prot_t prot,
320     vm_ooffset_t foff, struct ucred *cred)
321 {
322 }
323 
324 static void
shm_largepage_phys_dtor(vm_object_t object)325 shm_largepage_phys_dtor(vm_object_t object)
326 {
327 	int psind;
328 
329 	psind = object->un_pager.phys.data_val;
330 	if (psind != 0) {
331 		atomic_subtract_long(&count_largepages[psind],
332 		    object->size / (pagesizes[psind] / PAGE_SIZE));
333 		vm_wire_sub(object->size);
334 	} else {
335 		KASSERT(object->size == 0,
336 		    ("largepage phys obj %p not initialized bit size %#jx > 0",
337 		    object, (uintmax_t)object->size));
338 	}
339 }
340 
341 static const struct phys_pager_ops shm_largepage_phys_ops = {
342 	.phys_pg_populate =	shm_largepage_phys_populate,
343 	.phys_pg_haspage =	shm_largepage_phys_haspage,
344 	.phys_pg_ctor =		shm_largepage_phys_ctor,
345 	.phys_pg_dtor =		shm_largepage_phys_dtor,
346 };
347 
348 bool
shm_largepage(struct shmfd * shmfd)349 shm_largepage(struct shmfd *shmfd)
350 {
351 	return (shmfd->shm_object->type == OBJT_PHYS);
352 }
353 
354 static void
shm_pager_freespace(vm_object_t obj,vm_pindex_t start,vm_size_t size)355 shm_pager_freespace(vm_object_t obj, vm_pindex_t start, vm_size_t size)
356 {
357 	struct shmfd *shm;
358 	vm_size_t c;
359 
360 	swap_pager_freespace(obj, start, size, &c);
361 	if (c == 0)
362 		return;
363 
364 	shm = obj->un_pager.swp.swp_priv;
365 	if (shm == NULL)
366 		return;
367 	KASSERT(shm->shm_pages >= c,
368 	    ("shm %p pages %jd free %jd", shm,
369 	    (uintmax_t)shm->shm_pages, (uintmax_t)c));
370 	shm->shm_pages -= c;
371 }
372 
373 static void
shm_page_inserted(vm_object_t obj,vm_page_t m)374 shm_page_inserted(vm_object_t obj, vm_page_t m)
375 {
376 	struct shmfd *shm;
377 
378 	shm = obj->un_pager.swp.swp_priv;
379 	if (shm == NULL)
380 		return;
381 	if (!vm_pager_has_page(obj, m->pindex, NULL, NULL))
382 		shm->shm_pages += 1;
383 }
384 
385 static void
shm_page_removed(vm_object_t obj,vm_page_t m)386 shm_page_removed(vm_object_t obj, vm_page_t m)
387 {
388 	struct shmfd *shm;
389 
390 	shm = obj->un_pager.swp.swp_priv;
391 	if (shm == NULL)
392 		return;
393 	if (!vm_pager_has_page(obj, m->pindex, NULL, NULL)) {
394 		KASSERT(shm->shm_pages >= 1,
395 		    ("shm %p pages %jd free 1", shm,
396 		    (uintmax_t)shm->shm_pages));
397 		shm->shm_pages -= 1;
398 	}
399 }
400 
401 static struct pagerops shm_swap_pager_ops = {
402 	.pgo_kvme_type = KVME_TYPE_SWAP,
403 	.pgo_freespace = shm_pager_freespace,
404 	.pgo_page_inserted = shm_page_inserted,
405 	.pgo_page_removed = shm_page_removed,
406 };
407 static int shmfd_pager_type = -1;
408 
409 static int
shm_seek(struct file * fp,off_t offset,int whence,struct thread * td)410 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
411 {
412 	struct shmfd *shmfd;
413 	off_t foffset;
414 	int error;
415 
416 	shmfd = fp->f_data;
417 	foffset = foffset_lock(fp, 0);
418 	error = 0;
419 	switch (whence) {
420 	case L_INCR:
421 		if (foffset < 0 ||
422 		    (offset > 0 && foffset > OFF_MAX - offset)) {
423 			error = EOVERFLOW;
424 			break;
425 		}
426 		offset += foffset;
427 		break;
428 	case L_XTND:
429 		if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
430 			error = EOVERFLOW;
431 			break;
432 		}
433 		offset += shmfd->shm_size;
434 		break;
435 	case L_SET:
436 		break;
437 	default:
438 		error = EINVAL;
439 	}
440 	if (error == 0) {
441 		if (offset < 0 || offset > shmfd->shm_size)
442 			error = EINVAL;
443 		else
444 			td->td_uretoff.tdu_off = offset;
445 	}
446 	foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
447 	return (error);
448 }
449 
450 static int
shm_read(struct file * fp,struct uio * uio,struct ucred * active_cred,int flags,struct thread * td)451 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
452     int flags, struct thread *td)
453 {
454 	struct shmfd *shmfd;
455 	void *rl_cookie;
456 	int error;
457 
458 	shmfd = fp->f_data;
459 #ifdef MAC
460 	error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
461 	if (error)
462 		return (error);
463 #endif
464 	foffset_lock_uio(fp, uio, flags);
465 	rl_cookie = shm_rangelock_rlock(shmfd, uio->uio_offset,
466 	    uio->uio_offset + uio->uio_resid);
467 	error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
468 	shm_rangelock_unlock(shmfd, rl_cookie);
469 	foffset_unlock_uio(fp, uio, flags);
470 	return (error);
471 }
472 
473 static int
shm_write(struct file * fp,struct uio * uio,struct ucred * active_cred,int flags,struct thread * td)474 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
475     int flags, struct thread *td)
476 {
477 	struct shmfd *shmfd;
478 	void *rl_cookie;
479 	int error;
480 	off_t size;
481 
482 	shmfd = fp->f_data;
483 #ifdef MAC
484 	error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
485 	if (error)
486 		return (error);
487 #endif
488 	if (shm_largepage(shmfd) && shmfd->shm_lp_psind == 0)
489 		return (EINVAL);
490 	foffset_lock_uio(fp, uio, flags);
491 	if (uio->uio_resid > OFF_MAX - uio->uio_offset) {
492 		/*
493 		 * Overflow is only an error if we're supposed to expand on
494 		 * write.  Otherwise, we'll just truncate the write to the
495 		 * size of the file, which can only grow up to OFF_MAX.
496 		 */
497 		if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0) {
498 			foffset_unlock_uio(fp, uio, flags);
499 			return (EFBIG);
500 		}
501 
502 		size = shmfd->shm_size;
503 	} else {
504 		size = uio->uio_offset + uio->uio_resid;
505 	}
506 	if ((flags & FOF_OFFSET) == 0)
507 		rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
508 	else
509 		rl_cookie = shm_rangelock_wlock(shmfd, uio->uio_offset, size);
510 	if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) {
511 		error = EPERM;
512 	} else {
513 		error = 0;
514 		if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0 &&
515 		    size > shmfd->shm_size) {
516 			error = shm_dotruncate_cookie(shmfd, size, rl_cookie);
517 		}
518 		if (error == 0)
519 			error = uiomove_object(shmfd->shm_object,
520 			    shmfd->shm_size, uio);
521 	}
522 	shm_rangelock_unlock(shmfd, rl_cookie);
523 	foffset_unlock_uio(fp, uio, flags);
524 	return (error);
525 }
526 
527 static int
shm_truncate(struct file * fp,off_t length,struct ucred * active_cred,struct thread * td)528 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
529     struct thread *td)
530 {
531 	struct shmfd *shmfd;
532 #ifdef MAC
533 	int error;
534 #endif
535 
536 	shmfd = fp->f_data;
537 #ifdef MAC
538 	error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
539 	if (error)
540 		return (error);
541 #endif
542 	return (shm_dotruncate(shmfd, length));
543 }
544 
545 int
shm_ioctl(struct file * fp,u_long com,void * data,struct ucred * active_cred,struct thread * td)546 shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
547     struct thread *td)
548 {
549 	struct shmfd *shmfd;
550 	struct shm_largepage_conf *conf;
551 	void *rl_cookie;
552 
553 	shmfd = fp->f_data;
554 	switch (com) {
555 	case FIONBIO:
556 	case FIOASYNC:
557 		/*
558 		 * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work,
559 		 * just like it would on an unlinked regular file
560 		 */
561 		return (0);
562 	case FIOSSHMLPGCNF:
563 		if (!shm_largepage(shmfd))
564 			return (ENOTTY);
565 		conf = data;
566 		if (shmfd->shm_lp_psind != 0 &&
567 		    conf->psind != shmfd->shm_lp_psind)
568 			return (EINVAL);
569 		if (conf->psind <= 0 || conf->psind >= MAXPAGESIZES ||
570 		    pagesizes[conf->psind] == 0)
571 			return (EINVAL);
572 		if (conf->alloc_policy != SHM_LARGEPAGE_ALLOC_DEFAULT &&
573 		    conf->alloc_policy != SHM_LARGEPAGE_ALLOC_NOWAIT &&
574 		    conf->alloc_policy != SHM_LARGEPAGE_ALLOC_HARD)
575 			return (EINVAL);
576 
577 		rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
578 		shmfd->shm_lp_psind = conf->psind;
579 		shmfd->shm_lp_alloc_policy = conf->alloc_policy;
580 		shmfd->shm_object->un_pager.phys.data_val = conf->psind;
581 		shm_rangelock_unlock(shmfd, rl_cookie);
582 		return (0);
583 	case FIOGSHMLPGCNF:
584 		if (!shm_largepage(shmfd))
585 			return (ENOTTY);
586 		conf = data;
587 		rl_cookie = shm_rangelock_rlock(shmfd, 0, OFF_MAX);
588 		conf->psind = shmfd->shm_lp_psind;
589 		conf->alloc_policy = shmfd->shm_lp_alloc_policy;
590 		shm_rangelock_unlock(shmfd, rl_cookie);
591 		return (0);
592 	default:
593 		return (ENOTTY);
594 	}
595 }
596 
597 static int
shm_stat(struct file * fp,struct stat * sb,struct ucred * active_cred,struct thread * td)598 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
599     struct thread *td)
600 {
601 	struct shmfd *shmfd;
602 #ifdef MAC
603 	int error;
604 #endif
605 
606 	shmfd = fp->f_data;
607 
608 #ifdef MAC
609 	error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
610 	if (error)
611 		return (error);
612 #endif
613 
614 	/*
615 	 * Attempt to return sanish values for fstat() on a memory file
616 	 * descriptor.
617 	 */
618 	bzero(sb, sizeof(*sb));
619 	sb->st_blksize = PAGE_SIZE;
620 	sb->st_size = shmfd->shm_size;
621 	mtx_lock(&shm_timestamp_lock);
622 	sb->st_atim = shmfd->shm_atime;
623 	sb->st_ctim = shmfd->shm_ctime;
624 	sb->st_mtim = shmfd->shm_mtime;
625 	sb->st_birthtim = shmfd->shm_birthtime;
626 	sb->st_mode = S_IFREG | shmfd->shm_mode;		/* XXX */
627 	sb->st_uid = shmfd->shm_uid;
628 	sb->st_gid = shmfd->shm_gid;
629 	mtx_unlock(&shm_timestamp_lock);
630 	sb->st_dev = shm_dev_ino;
631 	sb->st_ino = shmfd->shm_ino;
632 	sb->st_nlink = shmfd->shm_object->ref_count;
633 	if (shm_largepage(shmfd)) {
634 		sb->st_blocks = shmfd->shm_object->size /
635 		    (pagesizes[shmfd->shm_lp_psind] >> PAGE_SHIFT);
636 	} else {
637 		sb->st_blocks = shmfd->shm_pages;
638 	}
639 
640 	return (0);
641 }
642 
643 static int
shm_close(struct file * fp,struct thread * td)644 shm_close(struct file *fp, struct thread *td)
645 {
646 	struct shmfd *shmfd;
647 
648 	shmfd = fp->f_data;
649 	fp->f_data = NULL;
650 	shm_drop(shmfd);
651 
652 	return (0);
653 }
654 
655 static int
shm_copyin_path(struct thread * td,const char * userpath_in,char ** path_out)656 shm_copyin_path(struct thread *td, const char *userpath_in, char **path_out) {
657 	int error;
658 	char *path;
659 	const char *pr_path;
660 	size_t pr_pathlen;
661 
662 	path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
663 	pr_path = td->td_ucred->cr_prison->pr_path;
664 
665 	/* Construct a full pathname for jailed callers. */
666 	pr_pathlen = strcmp(pr_path, "/") ==
667 	    0 ? 0 : strlcpy(path, pr_path, MAXPATHLEN);
668 	error = copyinstr(userpath_in, path + pr_pathlen,
669 	    MAXPATHLEN - pr_pathlen, NULL);
670 	if (error != 0)
671 		goto out;
672 
673 #ifdef KTRACE
674 	if (KTRPOINT(curthread, KTR_NAMEI))
675 		ktrnamei(path);
676 #endif
677 
678 	/* Require paths to start with a '/' character. */
679 	if (path[pr_pathlen] != '/') {
680 		error = EINVAL;
681 		goto out;
682 	}
683 
684 	*path_out = path;
685 
686 out:
687 	if (error != 0)
688 		free(path, M_SHMFD);
689 
690 	return (error);
691 }
692 
693 static int
shm_dotruncate_locked(struct shmfd * shmfd,off_t length,void * rl_cookie)694 shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie)
695 {
696 	vm_object_t object;
697 	vm_page_t m;
698 	vm_pindex_t idx, nobjsize;
699 	vm_ooffset_t delta;
700 	int base, rv;
701 
702 	KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
703 	object = shmfd->shm_object;
704 	VM_OBJECT_ASSERT_WLOCKED(object);
705 	rangelock_cookie_assert(rl_cookie, RA_WLOCKED);
706 	if (length == shmfd->shm_size)
707 		return (0);
708 	nobjsize = OFF_TO_IDX(length + PAGE_MASK);
709 
710 	/* Are we shrinking?  If so, trim the end. */
711 	if (length < shmfd->shm_size) {
712 		if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0)
713 			return (EPERM);
714 
715 		/*
716 		 * Disallow any requests to shrink the size if this
717 		 * object is mapped into the kernel.
718 		 */
719 		if (shmfd->shm_kmappings > 0)
720 			return (EBUSY);
721 
722 		/*
723 		 * Zero the truncated part of the last page.
724 		 */
725 		base = length & PAGE_MASK;
726 		if (base != 0) {
727 			idx = OFF_TO_IDX(length);
728 retry:
729 			m = vm_page_grab(object, idx, VM_ALLOC_NOCREAT);
730 			if (m != NULL) {
731 				MPASS(vm_page_all_valid(m));
732 			} else if (vm_pager_has_page(object, idx, NULL, NULL)) {
733 				m = vm_page_alloc(object, idx,
734 				    VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL);
735 				if (m == NULL)
736 					goto retry;
737 				vm_object_pip_add(object, 1);
738 				VM_OBJECT_WUNLOCK(object);
739 				rv = vm_pager_get_pages(object, &m, 1, NULL,
740 				    NULL);
741 				VM_OBJECT_WLOCK(object);
742 				vm_object_pip_wakeup(object);
743 				if (rv == VM_PAGER_OK) {
744 					/*
745 					 * Since the page was not resident,
746 					 * and therefore not recently
747 					 * accessed, immediately enqueue it
748 					 * for asynchronous laundering.  The
749 					 * current operation is not regarded
750 					 * as an access.
751 					 */
752 					vm_page_launder(m);
753 				} else {
754 					vm_page_free(m);
755 					VM_OBJECT_WUNLOCK(object);
756 					return (EIO);
757 				}
758 			}
759 			if (m != NULL) {
760 				pmap_zero_page_area(m, base, PAGE_SIZE - base);
761 				KASSERT(vm_page_all_valid(m),
762 				    ("shm_dotruncate: page %p is invalid", m));
763 				vm_page_set_dirty(m);
764 				vm_page_xunbusy(m);
765 			}
766 		}
767 		delta = IDX_TO_OFF(object->size - nobjsize);
768 
769 		if (nobjsize < object->size)
770 			vm_object_page_remove(object, nobjsize, object->size,
771 			    0);
772 
773 		/* Free the swap accounted for shm */
774 		swap_release_by_cred(delta, object->cred);
775 		object->charge -= delta;
776 	} else {
777 		if ((shmfd->shm_seals & F_SEAL_GROW) != 0)
778 			return (EPERM);
779 
780 		/* Try to reserve additional swap space. */
781 		delta = IDX_TO_OFF(nobjsize - object->size);
782 		if (!swap_reserve_by_cred(delta, object->cred))
783 			return (ENOMEM);
784 		object->charge += delta;
785 	}
786 	shmfd->shm_size = length;
787 	mtx_lock(&shm_timestamp_lock);
788 	vfs_timestamp(&shmfd->shm_ctime);
789 	shmfd->shm_mtime = shmfd->shm_ctime;
790 	mtx_unlock(&shm_timestamp_lock);
791 	object->size = nobjsize;
792 	return (0);
793 }
794 
795 static int
shm_dotruncate_largepage(struct shmfd * shmfd,off_t length,void * rl_cookie)796 shm_dotruncate_largepage(struct shmfd *shmfd, off_t length, void *rl_cookie)
797 {
798 	vm_object_t object;
799 	vm_page_t m;
800 	vm_pindex_t newobjsz;
801 	vm_pindex_t oldobjsz __unused;
802 	int aflags, error, i, psind, try;
803 
804 	KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
805 	object = shmfd->shm_object;
806 	VM_OBJECT_ASSERT_WLOCKED(object);
807 	rangelock_cookie_assert(rl_cookie, RA_WLOCKED);
808 
809 	oldobjsz = object->size;
810 	newobjsz = OFF_TO_IDX(length);
811 	if (length == shmfd->shm_size)
812 		return (0);
813 	psind = shmfd->shm_lp_psind;
814 	if (psind == 0 && length != 0)
815 		return (EINVAL);
816 	if ((length & (pagesizes[psind] - 1)) != 0)
817 		return (EINVAL);
818 
819 	if (length < shmfd->shm_size) {
820 		if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0)
821 			return (EPERM);
822 		if (shmfd->shm_kmappings > 0)
823 			return (EBUSY);
824 		return (ENOTSUP);	/* Pages are unmanaged. */
825 #if 0
826 		vm_object_page_remove(object, newobjsz, oldobjsz, 0);
827 		object->size = newobjsz;
828 		shmfd->shm_size = length;
829 		return (0);
830 #endif
831 	}
832 
833 	if ((shmfd->shm_seals & F_SEAL_GROW) != 0)
834 		return (EPERM);
835 
836 	aflags = VM_ALLOC_NORMAL | VM_ALLOC_ZERO;
837 	if (shmfd->shm_lp_alloc_policy == SHM_LARGEPAGE_ALLOC_NOWAIT)
838 		aflags |= VM_ALLOC_WAITFAIL;
839 	try = 0;
840 
841 	/*
842 	 * Extend shmfd and object, keeping all already fully
843 	 * allocated large pages intact even on error, because dropped
844 	 * object lock might allowed mapping of them.
845 	 */
846 	while (object->size < newobjsz) {
847 		m = vm_page_alloc_contig(object, object->size, aflags,
848 		    pagesizes[psind] / PAGE_SIZE, 0, ~0,
849 		    pagesizes[psind], 0,
850 		    VM_MEMATTR_DEFAULT);
851 		if (m == NULL) {
852 			VM_OBJECT_WUNLOCK(object);
853 			if (shmfd->shm_lp_alloc_policy ==
854 			    SHM_LARGEPAGE_ALLOC_NOWAIT ||
855 			    (shmfd->shm_lp_alloc_policy ==
856 			    SHM_LARGEPAGE_ALLOC_DEFAULT &&
857 			    try >= largepage_reclaim_tries)) {
858 				VM_OBJECT_WLOCK(object);
859 				return (ENOMEM);
860 			}
861 			error = vm_page_reclaim_contig(aflags,
862 			    pagesizes[psind] / PAGE_SIZE, 0, ~0,
863 			    pagesizes[psind], 0) ? 0 :
864 			    vm_wait_intr(object);
865 			if (error != 0) {
866 				VM_OBJECT_WLOCK(object);
867 				return (error);
868 			}
869 			try++;
870 			VM_OBJECT_WLOCK(object);
871 			continue;
872 		}
873 		try = 0;
874 		for (i = 0; i < pagesizes[psind] / PAGE_SIZE; i++) {
875 			if ((m[i].flags & PG_ZERO) == 0)
876 				pmap_zero_page(&m[i]);
877 			vm_page_valid(&m[i]);
878 			vm_page_xunbusy(&m[i]);
879 		}
880 		object->size += OFF_TO_IDX(pagesizes[psind]);
881 		shmfd->shm_size += pagesizes[psind];
882 		atomic_add_long(&count_largepages[psind], 1);
883 		vm_wire_add(atop(pagesizes[psind]));
884 	}
885 	return (0);
886 }
887 
888 static int
shm_dotruncate_cookie(struct shmfd * shmfd,off_t length,void * rl_cookie)889 shm_dotruncate_cookie(struct shmfd *shmfd, off_t length, void *rl_cookie)
890 {
891 	int error;
892 
893 	VM_OBJECT_WLOCK(shmfd->shm_object);
894 	error = shm_largepage(shmfd) ? shm_dotruncate_largepage(shmfd,
895 	    length, rl_cookie) : shm_dotruncate_locked(shmfd, length,
896 	    rl_cookie);
897 	VM_OBJECT_WUNLOCK(shmfd->shm_object);
898 	return (error);
899 }
900 
901 int
shm_dotruncate(struct shmfd * shmfd,off_t length)902 shm_dotruncate(struct shmfd *shmfd, off_t length)
903 {
904 	void *rl_cookie;
905 	int error;
906 
907 	rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
908 	error = shm_dotruncate_cookie(shmfd, length, rl_cookie);
909 	shm_rangelock_unlock(shmfd, rl_cookie);
910 	return (error);
911 }
912 
913 /*
914  * shmfd object management including creation and reference counting
915  * routines.
916  */
917 struct shmfd *
shm_alloc(struct ucred * ucred,mode_t mode,bool largepage)918 shm_alloc(struct ucred *ucred, mode_t mode, bool largepage)
919 {
920 	struct shmfd *shmfd;
921 	vm_object_t obj;
922 
923 	shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
924 	shmfd->shm_size = 0;
925 	shmfd->shm_uid = ucred->cr_uid;
926 	shmfd->shm_gid = ucred->cr_gid;
927 	shmfd->shm_mode = mode;
928 	if (largepage) {
929 		shmfd->shm_object = phys_pager_allocate(NULL,
930 		    &shm_largepage_phys_ops, NULL, shmfd->shm_size,
931 		    VM_PROT_DEFAULT, 0, ucred);
932 		shmfd->shm_lp_alloc_policy = SHM_LARGEPAGE_ALLOC_DEFAULT;
933 	} else {
934 		obj = vm_pager_allocate(shmfd_pager_type, NULL,
935 		    shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
936 		VM_OBJECT_WLOCK(obj);
937 		obj->un_pager.swp.swp_priv = shmfd;
938 		VM_OBJECT_WUNLOCK(obj);
939 		shmfd->shm_object = obj;
940 	}
941 	KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
942 	vfs_timestamp(&shmfd->shm_birthtime);
943 	shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
944 	    shmfd->shm_birthtime;
945 	shmfd->shm_ino = alloc_unr64(&shm_ino_unr);
946 	refcount_init(&shmfd->shm_refs, 1);
947 	mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
948 	rangelock_init(&shmfd->shm_rl);
949 #ifdef MAC
950 	mac_posixshm_init(shmfd);
951 	mac_posixshm_create(ucred, shmfd);
952 #endif
953 
954 	return (shmfd);
955 }
956 
957 struct shmfd *
shm_hold(struct shmfd * shmfd)958 shm_hold(struct shmfd *shmfd)
959 {
960 
961 	refcount_acquire(&shmfd->shm_refs);
962 	return (shmfd);
963 }
964 
965 void
shm_drop(struct shmfd * shmfd)966 shm_drop(struct shmfd *shmfd)
967 {
968 	vm_object_t obj;
969 
970 	if (refcount_release(&shmfd->shm_refs)) {
971 #ifdef MAC
972 		mac_posixshm_destroy(shmfd);
973 #endif
974 		rangelock_destroy(&shmfd->shm_rl);
975 		mtx_destroy(&shmfd->shm_mtx);
976 		obj = shmfd->shm_object;
977 		if (!shm_largepage(shmfd)) {
978 			VM_OBJECT_WLOCK(obj);
979 			obj->un_pager.swp.swp_priv = NULL;
980 			VM_OBJECT_WUNLOCK(obj);
981 		}
982 		vm_object_deallocate(obj);
983 		free(shmfd, M_SHMFD);
984 	}
985 }
986 
987 /*
988  * Determine if the credentials have sufficient permissions for a
989  * specified combination of FREAD and FWRITE.
990  */
991 int
shm_access(struct shmfd * shmfd,struct ucred * ucred,int flags)992 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
993 {
994 	accmode_t accmode;
995 	int error;
996 
997 	accmode = 0;
998 	if (flags & FREAD)
999 		accmode |= VREAD;
1000 	if (flags & FWRITE)
1001 		accmode |= VWRITE;
1002 	mtx_lock(&shm_timestamp_lock);
1003 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
1004 	    accmode, ucred);
1005 	mtx_unlock(&shm_timestamp_lock);
1006 	return (error);
1007 }
1008 
1009 static void
shm_init(void * arg)1010 shm_init(void *arg)
1011 {
1012 	char name[32];
1013 	int i;
1014 
1015 	mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
1016 	sx_init(&shm_dict_lock, "shm dictionary");
1017 	shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
1018 	new_unrhdr64(&shm_ino_unr, 1);
1019 	shm_dev_ino = devfs_alloc_cdp_inode();
1020 	KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
1021 	shmfd_pager_type = vm_pager_alloc_dyn_type(&shm_swap_pager_ops,
1022 	    OBJT_SWAP);
1023 	MPASS(shmfd_pager_type != -1);
1024 
1025 	for (i = 1; i < MAXPAGESIZES; i++) {
1026 		if (pagesizes[i] == 0)
1027 			break;
1028 #define	M	(1024 * 1024)
1029 #define	G	(1024 * M)
1030 		if (pagesizes[i] >= G)
1031 			snprintf(name, sizeof(name), "%luG", pagesizes[i] / G);
1032 		else if (pagesizes[i] >= M)
1033 			snprintf(name, sizeof(name), "%luM", pagesizes[i] / M);
1034 		else
1035 			snprintf(name, sizeof(name), "%lu", pagesizes[i]);
1036 #undef G
1037 #undef M
1038 		SYSCTL_ADD_ULONG(NULL, SYSCTL_STATIC_CHILDREN(_vm_largepages),
1039 		    OID_AUTO, name, CTLFLAG_RD, &count_largepages[i],
1040 		    "number of non-transient largepages allocated");
1041 	}
1042 }
1043 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
1044 
1045 /*
1046  * Remove all shared memory objects that belong to a prison.
1047  */
1048 void
shm_remove_prison(struct prison * pr)1049 shm_remove_prison(struct prison *pr)
1050 {
1051 	struct shm_mapping *shmm, *tshmm;
1052 	u_long i;
1053 
1054 	sx_xlock(&shm_dict_lock);
1055 	for (i = 0; i < shm_hash + 1; i++) {
1056 		LIST_FOREACH_SAFE(shmm, &shm_dictionary[i], sm_link, tshmm) {
1057 			if (shmm->sm_shmfd->shm_object->cred &&
1058 			    shmm->sm_shmfd->shm_object->cred->cr_prison == pr)
1059 				shm_doremove(shmm);
1060 		}
1061 	}
1062 	sx_xunlock(&shm_dict_lock);
1063 }
1064 
1065 /*
1066  * Dictionary management.  We maintain an in-kernel dictionary to map
1067  * paths to shmfd objects.  We use the FNV hash on the path to store
1068  * the mappings in a hash table.
1069  */
1070 static struct shmfd *
shm_lookup(char * path,Fnv32_t fnv)1071 shm_lookup(char *path, Fnv32_t fnv)
1072 {
1073 	struct shm_mapping *map;
1074 
1075 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
1076 		if (map->sm_fnv != fnv)
1077 			continue;
1078 		if (strcmp(map->sm_path, path) == 0)
1079 			return (map->sm_shmfd);
1080 	}
1081 
1082 	return (NULL);
1083 }
1084 
1085 static void
shm_insert(char * path,Fnv32_t fnv,struct shmfd * shmfd)1086 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
1087 {
1088 	struct shm_mapping *map;
1089 
1090 	map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
1091 	map->sm_path = path;
1092 	map->sm_fnv = fnv;
1093 	map->sm_shmfd = shm_hold(shmfd);
1094 	shmfd->shm_path = path;
1095 	LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
1096 }
1097 
1098 static int
shm_remove(char * path,Fnv32_t fnv,struct ucred * ucred)1099 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
1100 {
1101 	struct shm_mapping *map;
1102 	int error;
1103 
1104 	LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
1105 		if (map->sm_fnv != fnv)
1106 			continue;
1107 		if (strcmp(map->sm_path, path) == 0) {
1108 #ifdef MAC
1109 			error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
1110 			if (error)
1111 				return (error);
1112 #endif
1113 			error = shm_access(map->sm_shmfd, ucred,
1114 			    FREAD | FWRITE);
1115 			if (error)
1116 				return (error);
1117 			shm_doremove(map);
1118 			return (0);
1119 		}
1120 	}
1121 
1122 	return (ENOENT);
1123 }
1124 
1125 static void
shm_doremove(struct shm_mapping * map)1126 shm_doremove(struct shm_mapping *map)
1127 {
1128 	map->sm_shmfd->shm_path = NULL;
1129 	LIST_REMOVE(map, sm_link);
1130 	shm_drop(map->sm_shmfd);
1131 	free(map->sm_path, M_SHMFD);
1132 	free(map, M_SHMFD);
1133 }
1134 
1135 int
kern_shm_open2(struct thread * td,const char * userpath,int flags,mode_t mode,int shmflags,struct filecaps * fcaps,const char * name __unused)1136 kern_shm_open2(struct thread *td, const char *userpath, int flags, mode_t mode,
1137     int shmflags, struct filecaps *fcaps, const char *name __unused)
1138 {
1139 	struct pwddesc *pdp;
1140 	struct shmfd *shmfd;
1141 	struct file *fp;
1142 	char *path;
1143 	void *rl_cookie;
1144 	Fnv32_t fnv;
1145 	mode_t cmode;
1146 	int error, fd, initial_seals;
1147 	bool largepage;
1148 
1149 	if ((shmflags & ~(SHM_ALLOW_SEALING | SHM_GROW_ON_WRITE |
1150 	    SHM_LARGEPAGE)) != 0)
1151 		return (EINVAL);
1152 
1153 	initial_seals = F_SEAL_SEAL;
1154 	if ((shmflags & SHM_ALLOW_SEALING) != 0)
1155 		initial_seals &= ~F_SEAL_SEAL;
1156 
1157 #ifdef CAPABILITY_MODE
1158 	/*
1159 	 * shm_open(2) is only allowed for anonymous objects.
1160 	 */
1161 	if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON))
1162 		return (ECAPMODE);
1163 #endif
1164 
1165 	AUDIT_ARG_FFLAGS(flags);
1166 	AUDIT_ARG_MODE(mode);
1167 
1168 	if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
1169 		return (EINVAL);
1170 
1171 	if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
1172 		return (EINVAL);
1173 
1174 	largepage = (shmflags & SHM_LARGEPAGE) != 0;
1175 	if (largepage && !PMAP_HAS_LARGEPAGES)
1176 		return (ENOTTY);
1177 
1178 	/*
1179 	 * Currently only F_SEAL_SEAL may be set when creating or opening shmfd.
1180 	 * If the decision is made later to allow additional seals, care must be
1181 	 * taken below to ensure that the seals are properly set if the shmfd
1182 	 * already existed -- this currently assumes that only F_SEAL_SEAL can
1183 	 * be set and doesn't take further precautions to ensure the validity of
1184 	 * the seals being added with respect to current mappings.
1185 	 */
1186 	if ((initial_seals & ~F_SEAL_SEAL) != 0)
1187 		return (EINVAL);
1188 
1189 	pdp = td->td_proc->p_pd;
1190 	cmode = (mode & ~pdp->pd_cmask) & ACCESSPERMS;
1191 
1192 	/*
1193 	 * shm_open(2) created shm should always have O_CLOEXEC set, as mandated
1194 	 * by POSIX.  We allow it to be unset here so that an in-kernel
1195 	 * interface may be written as a thin layer around shm, optionally not
1196 	 * setting CLOEXEC.  For shm_open(2), O_CLOEXEC is set unconditionally
1197 	 * in sys_shm_open() to keep this implementation compliant.
1198 	 */
1199 	error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps);
1200 	if (error)
1201 		return (error);
1202 
1203 	/* A SHM_ANON path pointer creates an anonymous object. */
1204 	if (userpath == SHM_ANON) {
1205 		/* A read-only anonymous object is pointless. */
1206 		if ((flags & O_ACCMODE) == O_RDONLY) {
1207 			fdclose(td, fp, fd);
1208 			fdrop(fp, td);
1209 			return (EINVAL);
1210 		}
1211 		shmfd = shm_alloc(td->td_ucred, cmode, largepage);
1212 		shmfd->shm_seals = initial_seals;
1213 		shmfd->shm_flags = shmflags;
1214 	} else {
1215 		error = shm_copyin_path(td, userpath, &path);
1216 		if (error != 0) {
1217 			fdclose(td, fp, fd);
1218 			fdrop(fp, td);
1219 			return (error);
1220 		}
1221 
1222 		AUDIT_ARG_UPATH1_CANON(path);
1223 		fnv = fnv_32_str(path, FNV1_32_INIT);
1224 		sx_xlock(&shm_dict_lock);
1225 		shmfd = shm_lookup(path, fnv);
1226 		if (shmfd == NULL) {
1227 			/* Object does not yet exist, create it if requested. */
1228 			if (flags & O_CREAT) {
1229 #ifdef MAC
1230 				error = mac_posixshm_check_create(td->td_ucred,
1231 				    path);
1232 				if (error == 0) {
1233 #endif
1234 					shmfd = shm_alloc(td->td_ucred, cmode,
1235 					    largepage);
1236 					shmfd->shm_seals = initial_seals;
1237 					shmfd->shm_flags = shmflags;
1238 					shm_insert(path, fnv, shmfd);
1239 #ifdef MAC
1240 				}
1241 #endif
1242 			} else {
1243 				free(path, M_SHMFD);
1244 				error = ENOENT;
1245 			}
1246 		} else {
1247 			rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
1248 
1249 			/*
1250 			 * kern_shm_open() likely shouldn't ever error out on
1251 			 * trying to set a seal that already exists, unlike
1252 			 * F_ADD_SEALS.  This would break terribly as
1253 			 * shm_open(2) actually sets F_SEAL_SEAL to maintain
1254 			 * historical behavior where the underlying file could
1255 			 * not be sealed.
1256 			 */
1257 			initial_seals &= ~shmfd->shm_seals;
1258 
1259 			/*
1260 			 * Object already exists, obtain a new
1261 			 * reference if requested and permitted.
1262 			 */
1263 			free(path, M_SHMFD);
1264 
1265 			/*
1266 			 * initial_seals can't set additional seals if we've
1267 			 * already been set F_SEAL_SEAL.  If F_SEAL_SEAL is set,
1268 			 * then we've already removed that one from
1269 			 * initial_seals.  This is currently redundant as we
1270 			 * only allow setting F_SEAL_SEAL at creation time, but
1271 			 * it's cheap to check and decreases the effort required
1272 			 * to allow additional seals.
1273 			 */
1274 			if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 &&
1275 			    initial_seals != 0)
1276 				error = EPERM;
1277 			else if ((flags & (O_CREAT | O_EXCL)) ==
1278 			    (O_CREAT | O_EXCL))
1279 				error = EEXIST;
1280 			else if (shmflags != 0 && shmflags != shmfd->shm_flags)
1281 				error = EINVAL;
1282 			else {
1283 #ifdef MAC
1284 				error = mac_posixshm_check_open(td->td_ucred,
1285 				    shmfd, FFLAGS(flags & O_ACCMODE));
1286 				if (error == 0)
1287 #endif
1288 				error = shm_access(shmfd, td->td_ucred,
1289 				    FFLAGS(flags & O_ACCMODE));
1290 			}
1291 
1292 			/*
1293 			 * Truncate the file back to zero length if
1294 			 * O_TRUNC was specified and the object was
1295 			 * opened with read/write.
1296 			 */
1297 			if (error == 0 &&
1298 			    (flags & (O_ACCMODE | O_TRUNC)) ==
1299 			    (O_RDWR | O_TRUNC)) {
1300 				VM_OBJECT_WLOCK(shmfd->shm_object);
1301 #ifdef MAC
1302 				error = mac_posixshm_check_truncate(
1303 					td->td_ucred, fp->f_cred, shmfd);
1304 				if (error == 0)
1305 #endif
1306 					error = shm_dotruncate_locked(shmfd, 0,
1307 					    rl_cookie);
1308 				VM_OBJECT_WUNLOCK(shmfd->shm_object);
1309 			}
1310 			if (error == 0) {
1311 				/*
1312 				 * Currently we only allow F_SEAL_SEAL to be
1313 				 * set initially.  As noted above, this would
1314 				 * need to be reworked should that change.
1315 				 */
1316 				shmfd->shm_seals |= initial_seals;
1317 				shm_hold(shmfd);
1318 			}
1319 			shm_rangelock_unlock(shmfd, rl_cookie);
1320 		}
1321 		sx_xunlock(&shm_dict_lock);
1322 
1323 		if (error) {
1324 			fdclose(td, fp, fd);
1325 			fdrop(fp, td);
1326 			return (error);
1327 		}
1328 	}
1329 
1330 	finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
1331 
1332 	td->td_retval[0] = fd;
1333 	fdrop(fp, td);
1334 
1335 	return (0);
1336 }
1337 
1338 /* System calls. */
1339 #ifdef COMPAT_FREEBSD12
1340 int
freebsd12_shm_open(struct thread * td,struct freebsd12_shm_open_args * uap)1341 freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap)
1342 {
1343 
1344 	return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC,
1345 	    uap->mode, NULL));
1346 }
1347 #endif
1348 
1349 int
sys_shm_unlink(struct thread * td,struct shm_unlink_args * uap)1350 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
1351 {
1352 	char *path;
1353 	Fnv32_t fnv;
1354 	int error;
1355 
1356 	error = shm_copyin_path(td, uap->path, &path);
1357 	if (error != 0)
1358 		return (error);
1359 
1360 	AUDIT_ARG_UPATH1_CANON(path);
1361 	fnv = fnv_32_str(path, FNV1_32_INIT);
1362 	sx_xlock(&shm_dict_lock);
1363 	error = shm_remove(path, fnv, td->td_ucred);
1364 	sx_xunlock(&shm_dict_lock);
1365 	free(path, M_SHMFD);
1366 
1367 	return (error);
1368 }
1369 
1370 int
sys_shm_rename(struct thread * td,struct shm_rename_args * uap)1371 sys_shm_rename(struct thread *td, struct shm_rename_args *uap)
1372 {
1373 	char *path_from = NULL, *path_to = NULL;
1374 	Fnv32_t fnv_from, fnv_to;
1375 	struct shmfd *fd_from;
1376 	struct shmfd *fd_to;
1377 	int error;
1378 	int flags;
1379 
1380 	flags = uap->flags;
1381 	AUDIT_ARG_FFLAGS(flags);
1382 
1383 	/*
1384 	 * Make sure the user passed only valid flags.
1385 	 * If you add a new flag, please add a new term here.
1386 	 */
1387 	if ((flags & ~(
1388 	    SHM_RENAME_NOREPLACE |
1389 	    SHM_RENAME_EXCHANGE
1390 	    )) != 0) {
1391 		error = EINVAL;
1392 		goto out;
1393 	}
1394 
1395 	/*
1396 	 * EXCHANGE and NOREPLACE don't quite make sense together. Let's
1397 	 * force the user to choose one or the other.
1398 	 */
1399 	if ((flags & SHM_RENAME_NOREPLACE) != 0 &&
1400 	    (flags & SHM_RENAME_EXCHANGE) != 0) {
1401 		error = EINVAL;
1402 		goto out;
1403 	}
1404 
1405 	/* Renaming to or from anonymous makes no sense */
1406 	if (uap->path_from == SHM_ANON || uap->path_to == SHM_ANON) {
1407 		error = EINVAL;
1408 		goto out;
1409 	}
1410 
1411 	error = shm_copyin_path(td, uap->path_from, &path_from);
1412 	if (error != 0)
1413 		goto out;
1414 
1415 	error = shm_copyin_path(td, uap->path_to, &path_to);
1416 	if (error != 0)
1417 		goto out;
1418 
1419 	AUDIT_ARG_UPATH1_CANON(path_from);
1420 	AUDIT_ARG_UPATH2_CANON(path_to);
1421 
1422 	/* Rename with from/to equal is a no-op */
1423 	if (strcmp(path_from, path_to) == 0)
1424 		goto out;
1425 
1426 	fnv_from = fnv_32_str(path_from, FNV1_32_INIT);
1427 	fnv_to = fnv_32_str(path_to, FNV1_32_INIT);
1428 
1429 	sx_xlock(&shm_dict_lock);
1430 
1431 	fd_from = shm_lookup(path_from, fnv_from);
1432 	if (fd_from == NULL) {
1433 		error = ENOENT;
1434 		goto out_locked;
1435 	}
1436 
1437 	fd_to = shm_lookup(path_to, fnv_to);
1438 	if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) {
1439 		error = EEXIST;
1440 		goto out_locked;
1441 	}
1442 
1443 	/*
1444 	 * Unconditionally prevents shm_remove from invalidating the 'from'
1445 	 * shm's state.
1446 	 */
1447 	shm_hold(fd_from);
1448 	error = shm_remove(path_from, fnv_from, td->td_ucred);
1449 
1450 	/*
1451 	 * One of my assumptions failed if ENOENT (e.g. locking didn't
1452 	 * protect us)
1453 	 */
1454 	KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s",
1455 	    path_from));
1456 	if (error != 0) {
1457 		shm_drop(fd_from);
1458 		goto out_locked;
1459 	}
1460 
1461 	/*
1462 	 * If we are exchanging, we need to ensure the shm_remove below
1463 	 * doesn't invalidate the dest shm's state.
1464 	 */
1465 	if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL)
1466 		shm_hold(fd_to);
1467 
1468 	/*
1469 	 * NOTE: if path_to is not already in the hash, c'est la vie;
1470 	 * it simply means we have nothing already at path_to to unlink.
1471 	 * That is the ENOENT case.
1472 	 *
1473 	 * If we somehow don't have access to unlink this guy, but
1474 	 * did for the shm at path_from, then relink the shm to path_from
1475 	 * and abort with EACCES.
1476 	 *
1477 	 * All other errors: that is weird; let's relink and abort the
1478 	 * operation.
1479 	 */
1480 	error = shm_remove(path_to, fnv_to, td->td_ucred);
1481 	if (error != 0 && error != ENOENT) {
1482 		shm_insert(path_from, fnv_from, fd_from);
1483 		shm_drop(fd_from);
1484 		/* Don't free path_from now, since the hash references it */
1485 		path_from = NULL;
1486 		goto out_locked;
1487 	}
1488 
1489 	error = 0;
1490 
1491 	shm_insert(path_to, fnv_to, fd_from);
1492 
1493 	/* Don't free path_to now, since the hash references it */
1494 	path_to = NULL;
1495 
1496 	/* We kept a ref when we removed, and incremented again in insert */
1497 	shm_drop(fd_from);
1498 	KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1499 	    fd_from->shm_refs));
1500 
1501 	if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) {
1502 		shm_insert(path_from, fnv_from, fd_to);
1503 		path_from = NULL;
1504 		shm_drop(fd_to);
1505 		KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1506 		    fd_to->shm_refs));
1507 	}
1508 
1509 out_locked:
1510 	sx_xunlock(&shm_dict_lock);
1511 
1512 out:
1513 	free(path_from, M_SHMFD);
1514 	free(path_to, M_SHMFD);
1515 	return (error);
1516 }
1517 
1518 static int
shm_mmap_large(struct shmfd * shmfd,vm_map_t map,vm_offset_t * addr,vm_size_t size,vm_prot_t prot,vm_prot_t max_prot,int flags,vm_ooffset_t foff,struct thread * td)1519 shm_mmap_large(struct shmfd *shmfd, vm_map_t map, vm_offset_t *addr,
1520     vm_size_t size, vm_prot_t prot, vm_prot_t max_prot, int flags,
1521     vm_ooffset_t foff, struct thread *td)
1522 {
1523 	struct vmspace *vms;
1524 	vm_map_entry_t next_entry, prev_entry;
1525 	vm_offset_t align, mask, maxaddr;
1526 	int docow, error, rv, try;
1527 	bool curmap;
1528 
1529 	if (shmfd->shm_lp_psind == 0)
1530 		return (EINVAL);
1531 
1532 	/* MAP_PRIVATE is disabled */
1533 	if ((flags & ~(MAP_SHARED | MAP_FIXED | MAP_EXCL |
1534 	    MAP_NOCORE |
1535 #ifdef MAP_32BIT
1536 	    MAP_32BIT |
1537 #endif
1538 	    MAP_ALIGNMENT_MASK)) != 0)
1539 		return (EINVAL);
1540 
1541 	vms = td->td_proc->p_vmspace;
1542 	curmap = map == &vms->vm_map;
1543 	if (curmap) {
1544 		error = kern_mmap_racct_check(td, map, size);
1545 		if (error != 0)
1546 			return (error);
1547 	}
1548 
1549 	docow = shmfd->shm_lp_psind << MAP_SPLIT_BOUNDARY_SHIFT;
1550 	docow |= MAP_INHERIT_SHARE;
1551 	if ((flags & MAP_NOCORE) != 0)
1552 		docow |= MAP_DISABLE_COREDUMP;
1553 
1554 	mask = pagesizes[shmfd->shm_lp_psind] - 1;
1555 	if ((foff & mask) != 0)
1556 		return (EINVAL);
1557 	maxaddr = vm_map_max(map);
1558 #ifdef MAP_32BIT
1559 	if ((flags & MAP_32BIT) != 0 && maxaddr > MAP_32BIT_MAX_ADDR)
1560 		maxaddr = MAP_32BIT_MAX_ADDR;
1561 #endif
1562 	if (size == 0 || (size & mask) != 0 ||
1563 	    (*addr != 0 && ((*addr & mask) != 0 ||
1564 	    *addr + size < *addr || *addr + size > maxaddr)))
1565 		return (EINVAL);
1566 
1567 	align = flags & MAP_ALIGNMENT_MASK;
1568 	if (align == 0) {
1569 		align = pagesizes[shmfd->shm_lp_psind];
1570 	} else if (align == MAP_ALIGNED_SUPER) {
1571 		if (shmfd->shm_lp_psind != 1)
1572 			return (EINVAL);
1573 		align = pagesizes[1];
1574 	} else {
1575 		align >>= MAP_ALIGNMENT_SHIFT;
1576 		align = 1ULL << align;
1577 		/* Also handles overflow. */
1578 		if (align < pagesizes[shmfd->shm_lp_psind])
1579 			return (EINVAL);
1580 	}
1581 
1582 	vm_map_lock(map);
1583 	if ((flags & MAP_FIXED) == 0) {
1584 		try = 1;
1585 		if (curmap && (*addr == 0 ||
1586 		    (*addr >= round_page((vm_offset_t)vms->vm_taddr) &&
1587 		    *addr < round_page((vm_offset_t)vms->vm_daddr +
1588 		    lim_max(td, RLIMIT_DATA))))) {
1589 			*addr = roundup2((vm_offset_t)vms->vm_daddr +
1590 			    lim_max(td, RLIMIT_DATA),
1591 			    pagesizes[shmfd->shm_lp_psind]);
1592 		}
1593 again:
1594 		rv = vm_map_find_aligned(map, addr, size, maxaddr, align);
1595 		if (rv != KERN_SUCCESS) {
1596 			if (try == 1) {
1597 				try = 2;
1598 				*addr = vm_map_min(map);
1599 				if ((*addr & mask) != 0)
1600 					*addr = (*addr + mask) & mask;
1601 				goto again;
1602 			}
1603 			goto fail1;
1604 		}
1605 	} else if ((flags & MAP_EXCL) == 0) {
1606 		rv = vm_map_delete(map, *addr, *addr + size);
1607 		if (rv != KERN_SUCCESS)
1608 			goto fail1;
1609 	} else {
1610 		error = ENOSPC;
1611 		if (vm_map_lookup_entry(map, *addr, &prev_entry))
1612 			goto fail;
1613 		next_entry = vm_map_entry_succ(prev_entry);
1614 		if (next_entry->start < *addr + size)
1615 			goto fail;
1616 	}
1617 
1618 	rv = vm_map_insert(map, shmfd->shm_object, foff, *addr, *addr + size,
1619 	    prot, max_prot, docow);
1620 fail1:
1621 	error = vm_mmap_to_errno(rv);
1622 fail:
1623 	vm_map_unlock(map);
1624 	return (error);
1625 }
1626 
1627 static int
shm_mmap(struct file * fp,vm_map_t map,vm_offset_t * addr,vm_size_t objsize,vm_prot_t prot,vm_prot_t cap_maxprot,int flags,vm_ooffset_t foff,struct thread * td)1628 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
1629     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
1630     vm_ooffset_t foff, struct thread *td)
1631 {
1632 	struct shmfd *shmfd;
1633 	vm_prot_t maxprot;
1634 	int error;
1635 	bool writecnt;
1636 	void *rl_cookie;
1637 
1638 	shmfd = fp->f_data;
1639 	maxprot = VM_PROT_NONE;
1640 
1641 	rl_cookie = shm_rangelock_rlock(shmfd, 0, objsize);
1642 	/* FREAD should always be set. */
1643 	if ((fp->f_flag & FREAD) != 0)
1644 		maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
1645 
1646 	/*
1647 	 * If FWRITE's set, we can allow VM_PROT_WRITE unless it's a shared
1648 	 * mapping with a write seal applied.  Private mappings are always
1649 	 * writeable.
1650 	 */
1651 	if ((flags & MAP_SHARED) == 0) {
1652 		cap_maxprot |= VM_PROT_WRITE;
1653 		maxprot |= VM_PROT_WRITE;
1654 		writecnt = false;
1655 	} else {
1656 		if ((fp->f_flag & FWRITE) != 0 &&
1657 		    (shmfd->shm_seals & F_SEAL_WRITE) == 0)
1658 			maxprot |= VM_PROT_WRITE;
1659 
1660 		/*
1661 		 * Any mappings from a writable descriptor may be upgraded to
1662 		 * VM_PROT_WRITE with mprotect(2), unless a write-seal was
1663 		 * applied between the open and subsequent mmap(2).  We want to
1664 		 * reject application of a write seal as long as any such
1665 		 * mapping exists so that the seal cannot be trivially bypassed.
1666 		 */
1667 		writecnt = (maxprot & VM_PROT_WRITE) != 0;
1668 		if (!writecnt && (prot & VM_PROT_WRITE) != 0) {
1669 			error = EACCES;
1670 			goto out;
1671 		}
1672 	}
1673 	maxprot &= cap_maxprot;
1674 
1675 	/* See comment in vn_mmap(). */
1676 	if (
1677 #ifdef _LP64
1678 	    objsize > OFF_MAX ||
1679 #endif
1680 	    foff > OFF_MAX - objsize) {
1681 		error = EINVAL;
1682 		goto out;
1683 	}
1684 
1685 #ifdef MAC
1686 	error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
1687 	if (error != 0)
1688 		goto out;
1689 #endif
1690 
1691 	mtx_lock(&shm_timestamp_lock);
1692 	vfs_timestamp(&shmfd->shm_atime);
1693 	mtx_unlock(&shm_timestamp_lock);
1694 	vm_object_reference(shmfd->shm_object);
1695 
1696 	if (shm_largepage(shmfd)) {
1697 		writecnt = false;
1698 		error = shm_mmap_large(shmfd, map, addr, objsize, prot,
1699 		    maxprot, flags, foff, td);
1700 	} else {
1701 		if (writecnt) {
1702 			vm_pager_update_writecount(shmfd->shm_object, 0,
1703 			    objsize);
1704 		}
1705 		error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
1706 		    shmfd->shm_object, foff, writecnt, td);
1707 	}
1708 	if (error != 0) {
1709 		if (writecnt)
1710 			vm_pager_release_writecount(shmfd->shm_object, 0,
1711 			    objsize);
1712 		vm_object_deallocate(shmfd->shm_object);
1713 	}
1714 out:
1715 	shm_rangelock_unlock(shmfd, rl_cookie);
1716 	return (error);
1717 }
1718 
1719 static int
shm_chmod(struct file * fp,mode_t mode,struct ucred * active_cred,struct thread * td)1720 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
1721     struct thread *td)
1722 {
1723 	struct shmfd *shmfd;
1724 	int error;
1725 
1726 	error = 0;
1727 	shmfd = fp->f_data;
1728 	mtx_lock(&shm_timestamp_lock);
1729 	/*
1730 	 * SUSv4 says that x bits of permission need not be affected.
1731 	 * Be consistent with our shm_open there.
1732 	 */
1733 #ifdef MAC
1734 	error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
1735 	if (error != 0)
1736 		goto out;
1737 #endif
1738 	error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
1739 	    VADMIN, active_cred);
1740 	if (error != 0)
1741 		goto out;
1742 	shmfd->shm_mode = mode & ACCESSPERMS;
1743 out:
1744 	mtx_unlock(&shm_timestamp_lock);
1745 	return (error);
1746 }
1747 
1748 static int
shm_chown(struct file * fp,uid_t uid,gid_t gid,struct ucred * active_cred,struct thread * td)1749 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1750     struct thread *td)
1751 {
1752 	struct shmfd *shmfd;
1753 	int error;
1754 
1755 	error = 0;
1756 	shmfd = fp->f_data;
1757 	mtx_lock(&shm_timestamp_lock);
1758 #ifdef MAC
1759 	error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
1760 	if (error != 0)
1761 		goto out;
1762 #endif
1763 	if (uid == (uid_t)-1)
1764 		uid = shmfd->shm_uid;
1765 	if (gid == (gid_t)-1)
1766                  gid = shmfd->shm_gid;
1767 	if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
1768 	    (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
1769 	    (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN)))
1770 		goto out;
1771 	shmfd->shm_uid = uid;
1772 	shmfd->shm_gid = gid;
1773 out:
1774 	mtx_unlock(&shm_timestamp_lock);
1775 	return (error);
1776 }
1777 
1778 /*
1779  * Helper routines to allow the backing object of a shared memory file
1780  * descriptor to be mapped in the kernel.
1781  */
1782 int
shm_map(struct file * fp,size_t size,off_t offset,void ** memp)1783 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
1784 {
1785 	struct shmfd *shmfd;
1786 	vm_offset_t kva, ofs;
1787 	vm_object_t obj;
1788 	int rv;
1789 
1790 	if (fp->f_type != DTYPE_SHM)
1791 		return (EINVAL);
1792 	shmfd = fp->f_data;
1793 	obj = shmfd->shm_object;
1794 	VM_OBJECT_WLOCK(obj);
1795 	/*
1796 	 * XXXRW: This validation is probably insufficient, and subject to
1797 	 * sign errors.  It should be fixed.
1798 	 */
1799 	if (offset >= shmfd->shm_size ||
1800 	    offset + size > round_page(shmfd->shm_size)) {
1801 		VM_OBJECT_WUNLOCK(obj);
1802 		return (EINVAL);
1803 	}
1804 
1805 	shmfd->shm_kmappings++;
1806 	vm_object_reference_locked(obj);
1807 	VM_OBJECT_WUNLOCK(obj);
1808 
1809 	/* Map the object into the kernel_map and wire it. */
1810 	kva = vm_map_min(kernel_map);
1811 	ofs = offset & PAGE_MASK;
1812 	offset = trunc_page(offset);
1813 	size = round_page(size + ofs);
1814 	rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1815 	    VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1816 	    VM_PROT_READ | VM_PROT_WRITE, 0);
1817 	if (rv == KERN_SUCCESS) {
1818 		rv = vm_map_wire(kernel_map, kva, kva + size,
1819 		    VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1820 		if (rv == KERN_SUCCESS) {
1821 			*memp = (void *)(kva + ofs);
1822 			return (0);
1823 		}
1824 		vm_map_remove(kernel_map, kva, kva + size);
1825 	} else
1826 		vm_object_deallocate(obj);
1827 
1828 	/* On failure, drop our mapping reference. */
1829 	VM_OBJECT_WLOCK(obj);
1830 	shmfd->shm_kmappings--;
1831 	VM_OBJECT_WUNLOCK(obj);
1832 
1833 	return (vm_mmap_to_errno(rv));
1834 }
1835 
1836 /*
1837  * We require the caller to unmap the entire entry.  This allows us to
1838  * safely decrement shm_kmappings when a mapping is removed.
1839  */
1840 int
shm_unmap(struct file * fp,void * mem,size_t size)1841 shm_unmap(struct file *fp, void *mem, size_t size)
1842 {
1843 	struct shmfd *shmfd;
1844 	vm_map_entry_t entry;
1845 	vm_offset_t kva, ofs;
1846 	vm_object_t obj;
1847 	vm_pindex_t pindex;
1848 	vm_prot_t prot;
1849 	boolean_t wired;
1850 	vm_map_t map;
1851 	int rv;
1852 
1853 	if (fp->f_type != DTYPE_SHM)
1854 		return (EINVAL);
1855 	shmfd = fp->f_data;
1856 	kva = (vm_offset_t)mem;
1857 	ofs = kva & PAGE_MASK;
1858 	kva = trunc_page(kva);
1859 	size = round_page(size + ofs);
1860 	map = kernel_map;
1861 	rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1862 	    &obj, &pindex, &prot, &wired);
1863 	if (rv != KERN_SUCCESS)
1864 		return (EINVAL);
1865 	if (entry->start != kva || entry->end != kva + size) {
1866 		vm_map_lookup_done(map, entry);
1867 		return (EINVAL);
1868 	}
1869 	vm_map_lookup_done(map, entry);
1870 	if (obj != shmfd->shm_object)
1871 		return (EINVAL);
1872 	vm_map_remove(map, kva, kva + size);
1873 	VM_OBJECT_WLOCK(obj);
1874 	KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1875 	shmfd->shm_kmappings--;
1876 	VM_OBJECT_WUNLOCK(obj);
1877 	return (0);
1878 }
1879 
1880 static int
shm_fill_kinfo_locked(struct shmfd * shmfd,struct kinfo_file * kif,bool list)1881 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list)
1882 {
1883 	const char *path, *pr_path;
1884 	size_t pr_pathlen;
1885 	bool visible;
1886 
1887 	sx_assert(&shm_dict_lock, SA_LOCKED);
1888 	kif->kf_type = KF_TYPE_SHM;
1889 	kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;
1890 	kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1891 	if (shmfd->shm_path != NULL) {
1892 		if (shmfd->shm_path != NULL) {
1893 			path = shmfd->shm_path;
1894 			pr_path = curthread->td_ucred->cr_prison->pr_path;
1895 			if (strcmp(pr_path, "/") != 0) {
1896 				/* Return the jail-rooted pathname. */
1897 				pr_pathlen = strlen(pr_path);
1898 				visible = strncmp(path, pr_path, pr_pathlen)
1899 				    == 0 && path[pr_pathlen] == '/';
1900 				if (list && !visible)
1901 					return (EPERM);
1902 				if (visible)
1903 					path += pr_pathlen;
1904 			}
1905 			strlcpy(kif->kf_path, path, sizeof(kif->kf_path));
1906 		}
1907 	}
1908 	return (0);
1909 }
1910 
1911 static int
shm_fill_kinfo(struct file * fp,struct kinfo_file * kif,struct filedesc * fdp __unused)1912 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif,
1913     struct filedesc *fdp __unused)
1914 {
1915 	int res;
1916 
1917 	sx_slock(&shm_dict_lock);
1918 	res = shm_fill_kinfo_locked(fp->f_data, kif, false);
1919 	sx_sunlock(&shm_dict_lock);
1920 	return (res);
1921 }
1922 
1923 static int
shm_add_seals(struct file * fp,int seals)1924 shm_add_seals(struct file *fp, int seals)
1925 {
1926 	struct shmfd *shmfd;
1927 	void *rl_cookie;
1928 	vm_ooffset_t writemappings;
1929 	int error, nseals;
1930 
1931 	error = 0;
1932 	shmfd = fp->f_data;
1933 	rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
1934 
1935 	/* Even already-set seals should result in EPERM. */
1936 	if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) {
1937 		error = EPERM;
1938 		goto out;
1939 	}
1940 	nseals = seals & ~shmfd->shm_seals;
1941 	if ((nseals & F_SEAL_WRITE) != 0) {
1942 		if (shm_largepage(shmfd)) {
1943 			error = ENOTSUP;
1944 			goto out;
1945 		}
1946 
1947 		/*
1948 		 * The rangelock above prevents writable mappings from being
1949 		 * added after we've started applying seals.  The RLOCK here
1950 		 * is to avoid torn reads on ILP32 arches as unmapping/reducing
1951 		 * writemappings will be done without a rangelock.
1952 		 */
1953 		VM_OBJECT_RLOCK(shmfd->shm_object);
1954 		writemappings = shmfd->shm_object->un_pager.swp.writemappings;
1955 		VM_OBJECT_RUNLOCK(shmfd->shm_object);
1956 		/* kmappings are also writable */
1957 		if (writemappings > 0) {
1958 			error = EBUSY;
1959 			goto out;
1960 		}
1961 	}
1962 	shmfd->shm_seals |= nseals;
1963 out:
1964 	shm_rangelock_unlock(shmfd, rl_cookie);
1965 	return (error);
1966 }
1967 
1968 static int
shm_get_seals(struct file * fp,int * seals)1969 shm_get_seals(struct file *fp, int *seals)
1970 {
1971 	struct shmfd *shmfd;
1972 
1973 	shmfd = fp->f_data;
1974 	*seals = shmfd->shm_seals;
1975 	return (0);
1976 }
1977 
1978 static int
shm_fallocate(struct file * fp,off_t offset,off_t len,struct thread * td)1979 shm_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
1980 {
1981 	void *rl_cookie;
1982 	struct shmfd *shmfd;
1983 	size_t size;
1984 	int error;
1985 
1986 	/* This assumes that the caller already checked for overflow. */
1987 	error = 0;
1988 	shmfd = fp->f_data;
1989 	size = offset + len;
1990 
1991 	/*
1992 	 * Just grab the rangelock for the range that we may be attempting to
1993 	 * grow, rather than blocking read/write for regions we won't be
1994 	 * touching while this (potential) resize is in progress.  Other
1995 	 * attempts to resize the shmfd will have to take a write lock from 0 to
1996 	 * OFF_MAX, so this being potentially beyond the current usable range of
1997 	 * the shmfd is not necessarily a concern.  If other mechanisms are
1998 	 * added to grow a shmfd, this may need to be re-evaluated.
1999 	 */
2000 	rl_cookie = shm_rangelock_wlock(shmfd, offset, size);
2001 	if (size > shmfd->shm_size)
2002 		error = shm_dotruncate_cookie(shmfd, size, rl_cookie);
2003 	shm_rangelock_unlock(shmfd, rl_cookie);
2004 	/* Translate to posix_fallocate(2) return value as needed. */
2005 	if (error == ENOMEM)
2006 		error = ENOSPC;
2007 	return (error);
2008 }
2009 
2010 static int
sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS)2011 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS)
2012 {
2013 	struct shm_mapping *shmm;
2014 	struct sbuf sb;
2015 	struct kinfo_file kif;
2016 	u_long i;
2017 	ssize_t curlen;
2018 	int error, error2;
2019 
2020 	sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req);
2021 	sbuf_clear_flags(&sb, SBUF_INCLUDENUL);
2022 	curlen = 0;
2023 	error = 0;
2024 	sx_slock(&shm_dict_lock);
2025 	for (i = 0; i < shm_hash + 1; i++) {
2026 		LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) {
2027 			error = shm_fill_kinfo_locked(shmm->sm_shmfd,
2028 			    &kif, true);
2029 			if (error == EPERM) {
2030 				error = 0;
2031 				continue;
2032 			}
2033 			if (error != 0)
2034 				break;
2035 			pack_kinfo(&kif);
2036 			error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ?
2037 			    0 : ENOMEM;
2038 			if (error != 0)
2039 				break;
2040 			curlen += kif.kf_structsize;
2041 		}
2042 	}
2043 	sx_sunlock(&shm_dict_lock);
2044 	error2 = sbuf_finish(&sb);
2045 	sbuf_delete(&sb);
2046 	return (error != 0 ? error : error2);
2047 }
2048 
2049 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list,
2050     CTLFLAG_RD | CTLFLAG_PRISON | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE,
2051     NULL, 0, sysctl_posix_shm_list, "",
2052     "POSIX SHM list");
2053 
2054 int
kern_shm_open(struct thread * td,const char * path,int flags,mode_t mode,struct filecaps * caps)2055 kern_shm_open(struct thread *td, const char *path, int flags, mode_t mode,
2056     struct filecaps *caps)
2057 {
2058 
2059 	return (kern_shm_open2(td, path, flags, mode, 0, caps, NULL));
2060 }
2061 
2062 /*
2063  * This version of the shm_open() interface leaves CLOEXEC behavior up to the
2064  * caller, and libc will enforce it for the traditional shm_open() call.  This
2065  * allows other consumers, like memfd_create(), to opt-in for CLOEXEC.  This
2066  * interface also includes a 'name' argument that is currently unused, but could
2067  * potentially be exported later via some interface for debugging purposes.
2068  * From the kernel's perspective, it is optional.  Individual consumers like
2069  * memfd_create() may require it in order to be compatible with other systems
2070  * implementing the same function.
2071  */
2072 int
sys_shm_open2(struct thread * td,struct shm_open2_args * uap)2073 sys_shm_open2(struct thread *td, struct shm_open2_args *uap)
2074 {
2075 
2076 	return (kern_shm_open2(td, uap->path, uap->flags, uap->mode,
2077 	    uap->shmflags, NULL, uap->name));
2078 }
2079