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 static int shm_deallocate(struct shmfd *shmfd, off_t *offset,
134 off_t *length, int flags);
135
136 static fo_rdwr_t shm_read;
137 static fo_rdwr_t shm_write;
138 static fo_truncate_t shm_truncate;
139 static fo_ioctl_t shm_ioctl;
140 static fo_stat_t shm_stat;
141 static fo_close_t shm_close;
142 static fo_chmod_t shm_chmod;
143 static fo_chown_t shm_chown;
144 static fo_seek_t shm_seek;
145 static fo_fill_kinfo_t shm_fill_kinfo;
146 static fo_mmap_t shm_mmap;
147 static fo_get_seals_t shm_get_seals;
148 static fo_add_seals_t shm_add_seals;
149 static fo_fallocate_t shm_fallocate;
150 static fo_fspacectl_t shm_fspacectl;
151
152 /* File descriptor operations. */
153 const struct fileops shm_ops = {
154 .fo_read = shm_read,
155 .fo_write = shm_write,
156 .fo_truncate = shm_truncate,
157 .fo_ioctl = shm_ioctl,
158 .fo_poll = invfo_poll,
159 .fo_kqfilter = invfo_kqfilter,
160 .fo_stat = shm_stat,
161 .fo_close = shm_close,
162 .fo_chmod = shm_chmod,
163 .fo_chown = shm_chown,
164 .fo_sendfile = vn_sendfile,
165 .fo_seek = shm_seek,
166 .fo_fill_kinfo = shm_fill_kinfo,
167 .fo_mmap = shm_mmap,
168 .fo_get_seals = shm_get_seals,
169 .fo_add_seals = shm_add_seals,
170 .fo_fallocate = shm_fallocate,
171 .fo_fspacectl = shm_fspacectl,
172 .fo_cmp = file_kcmp_generic,
173 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE,
174 };
175
176 FEATURE(posix_shm, "POSIX shared memory");
177
178 static SYSCTL_NODE(_vm, OID_AUTO, largepages, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
179 "");
180
181 static int largepage_reclaim_tries = 1;
182 SYSCTL_INT(_vm_largepages, OID_AUTO, reclaim_tries,
183 CTLFLAG_RWTUN, &largepage_reclaim_tries, 0,
184 "Number of contig reclaims before giving up for default alloc policy");
185
186 #define shm_rangelock_unlock(shmfd, cookie) \
187 rangelock_unlock(&(shmfd)->shm_rl, (cookie))
188 #define shm_rangelock_rlock(shmfd, start, end) \
189 rangelock_rlock(&(shmfd)->shm_rl, (start), (end))
190 #define shm_rangelock_tryrlock(shmfd, start, end) \
191 rangelock_tryrlock(&(shmfd)->shm_rl, (start), (end))
192 #define shm_rangelock_wlock(shmfd, start, end) \
193 rangelock_wlock(&(shmfd)->shm_rl, (start), (end))
194
195 static int
uiomove_object_page(vm_object_t obj,size_t len,struct uio * uio)196 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio)
197 {
198 vm_page_t m;
199 vm_pindex_t idx;
200 size_t tlen;
201 int error, offset, rv;
202
203 idx = OFF_TO_IDX(uio->uio_offset);
204 offset = uio->uio_offset & PAGE_MASK;
205 tlen = MIN(PAGE_SIZE - offset, len);
206
207 rv = vm_page_grab_valid_unlocked(&m, obj, idx,
208 VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOCREAT);
209 if (rv == VM_PAGER_OK)
210 goto found;
211
212 /*
213 * Read I/O without either a corresponding resident page or swap
214 * page: use zero_region. This is intended to avoid instantiating
215 * pages on read from a sparse region.
216 */
217 VM_OBJECT_WLOCK(obj);
218 m = vm_page_lookup(obj, idx);
219 if (uio->uio_rw == UIO_READ && m == NULL &&
220 !vm_pager_has_page(obj, idx, NULL, NULL)) {
221 VM_OBJECT_WUNLOCK(obj);
222 return (uiomove(__DECONST(void *, zero_region), tlen, uio));
223 }
224
225 /*
226 * Although the tmpfs vnode lock is held here, it is
227 * nonetheless safe to sleep waiting for a free page. The
228 * pageout daemon does not need to acquire the tmpfs vnode
229 * lock to page out tobj's pages because tobj is a OBJT_SWAP
230 * type object.
231 */
232 rv = vm_page_grab_valid(&m, obj, idx,
233 VM_ALLOC_NORMAL | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY);
234 if (rv != VM_PAGER_OK) {
235 VM_OBJECT_WUNLOCK(obj);
236 if (bootverbose) {
237 printf("uiomove_object: vm_obj %p idx %jd "
238 "pager error %d\n", obj, idx, rv);
239 }
240 return (rv == VM_PAGER_AGAIN ? ENOSPC : EIO);
241 }
242 VM_OBJECT_WUNLOCK(obj);
243
244 found:
245 error = uiomove_fromphys(&m, offset, tlen, uio);
246 if (uio->uio_rw == UIO_WRITE && error == 0)
247 vm_page_set_dirty(m);
248 vm_page_activate(m);
249 vm_page_sunbusy(m);
250
251 return (error);
252 }
253
254 int
uiomove_object(vm_object_t obj,off_t obj_size,struct uio * uio)255 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
256 {
257 ssize_t resid;
258 size_t len;
259 int error;
260
261 error = 0;
262 while ((resid = uio->uio_resid) > 0) {
263 if (obj_size <= uio->uio_offset)
264 break;
265 len = MIN(obj_size - uio->uio_offset, resid);
266 if (len == 0)
267 break;
268 error = uiomove_object_page(obj, len, uio);
269 if (error != 0 || resid == uio->uio_resid)
270 break;
271 }
272 return (error);
273 }
274
275 static u_long count_largepages[MAXPAGESIZES];
276
277 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)278 shm_largepage_phys_populate(vm_object_t object, vm_pindex_t pidx,
279 int fault_type, vm_prot_t max_prot, vm_pindex_t *first, vm_pindex_t *last)
280 {
281 vm_page_t m __diagused;
282 int psind;
283
284 psind = object->un_pager.phys.data_val;
285 if (psind == 0 || pidx >= object->size)
286 return (VM_PAGER_FAIL);
287 *first = rounddown2(pidx, pagesizes[psind] / PAGE_SIZE);
288
289 /*
290 * We only busy the first page in the superpage run. It is
291 * useless to busy whole run since we only remove full
292 * superpage, and it takes too long to busy e.g. 512 * 512 ==
293 * 262144 pages constituing 1G amd64 superage.
294 */
295 m = vm_page_grab(object, *first, VM_ALLOC_NORMAL | VM_ALLOC_NOCREAT);
296 MPASS(m != NULL);
297
298 *last = *first + atop(pagesizes[psind]) - 1;
299 return (VM_PAGER_OK);
300 }
301
302 static boolean_t
shm_largepage_phys_haspage(vm_object_t object,vm_pindex_t pindex,int * before,int * after)303 shm_largepage_phys_haspage(vm_object_t object, vm_pindex_t pindex,
304 int *before, int *after)
305 {
306 int psind;
307
308 psind = object->un_pager.phys.data_val;
309 if (psind == 0 || pindex >= object->size)
310 return (FALSE);
311 if (before != NULL) {
312 *before = pindex - rounddown2(pindex, pagesizes[psind] /
313 PAGE_SIZE);
314 }
315 if (after != NULL) {
316 *after = roundup2(pindex, pagesizes[psind] / PAGE_SIZE) -
317 pindex;
318 }
319 return (TRUE);
320 }
321
322 static void
shm_largepage_phys_ctor(vm_object_t object,vm_prot_t prot,vm_ooffset_t foff,struct ucred * cred)323 shm_largepage_phys_ctor(vm_object_t object, vm_prot_t prot,
324 vm_ooffset_t foff, struct ucred *cred)
325 {
326 }
327
328 static void
shm_largepage_phys_dtor(vm_object_t object)329 shm_largepage_phys_dtor(vm_object_t object)
330 {
331 int psind;
332
333 psind = object->un_pager.phys.data_val;
334 if (psind != 0) {
335 atomic_subtract_long(&count_largepages[psind],
336 object->size / (pagesizes[psind] / PAGE_SIZE));
337 vm_wire_sub(object->size);
338 } else {
339 KASSERT(object->size == 0,
340 ("largepage phys obj %p not initialized bit size %#jx > 0",
341 object, (uintmax_t)object->size));
342 }
343 }
344
345 static const struct phys_pager_ops shm_largepage_phys_ops = {
346 .phys_pg_populate = shm_largepage_phys_populate,
347 .phys_pg_haspage = shm_largepage_phys_haspage,
348 .phys_pg_ctor = shm_largepage_phys_ctor,
349 .phys_pg_dtor = shm_largepage_phys_dtor,
350 };
351
352 bool
shm_largepage(struct shmfd * shmfd)353 shm_largepage(struct shmfd *shmfd)
354 {
355 return (shmfd->shm_object->type == OBJT_PHYS);
356 }
357
358 static void
shm_pager_freespace(vm_object_t obj,vm_pindex_t start,vm_size_t size)359 shm_pager_freespace(vm_object_t obj, vm_pindex_t start, vm_size_t size)
360 {
361 struct shmfd *shm;
362 vm_size_t c;
363
364 swap_pager_freespace(obj, start, size, &c);
365 if (c == 0)
366 return;
367
368 shm = obj->un_pager.swp.swp_priv;
369 if (shm == NULL)
370 return;
371 KASSERT(shm->shm_pages >= c,
372 ("shm %p pages %jd free %jd", shm,
373 (uintmax_t)shm->shm_pages, (uintmax_t)c));
374 shm->shm_pages -= c;
375 }
376
377 static void
shm_page_inserted(vm_object_t obj,vm_page_t m)378 shm_page_inserted(vm_object_t obj, vm_page_t m)
379 {
380 struct shmfd *shm;
381
382 shm = obj->un_pager.swp.swp_priv;
383 if (shm == NULL)
384 return;
385 if (!vm_pager_has_page(obj, m->pindex, NULL, NULL))
386 shm->shm_pages += 1;
387 }
388
389 static void
shm_page_removed(vm_object_t obj,vm_page_t m)390 shm_page_removed(vm_object_t obj, vm_page_t m)
391 {
392 struct shmfd *shm;
393
394 shm = obj->un_pager.swp.swp_priv;
395 if (shm == NULL)
396 return;
397 if (!vm_pager_has_page(obj, m->pindex, NULL, NULL)) {
398 KASSERT(shm->shm_pages >= 1,
399 ("shm %p pages %jd free 1", shm,
400 (uintmax_t)shm->shm_pages));
401 shm->shm_pages -= 1;
402 }
403 }
404
405 static struct pagerops shm_swap_pager_ops = {
406 .pgo_kvme_type = KVME_TYPE_SWAP,
407 .pgo_freespace = shm_pager_freespace,
408 .pgo_page_inserted = shm_page_inserted,
409 .pgo_page_removed = shm_page_removed,
410 };
411 static int shmfd_pager_type = -1;
412
413 static int
shm_seek(struct file * fp,off_t offset,int whence,struct thread * td)414 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
415 {
416 struct shmfd *shmfd;
417 off_t foffset;
418 int error;
419
420 shmfd = fp->f_data;
421 foffset = foffset_lock(fp, 0);
422 error = 0;
423 switch (whence) {
424 case L_INCR:
425 if (foffset < 0 ||
426 (offset > 0 && foffset > OFF_MAX - offset)) {
427 error = EOVERFLOW;
428 break;
429 }
430 offset += foffset;
431 break;
432 case L_XTND:
433 if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
434 error = EOVERFLOW;
435 break;
436 }
437 offset += shmfd->shm_size;
438 break;
439 case L_SET:
440 break;
441 default:
442 error = EINVAL;
443 }
444 if (error == 0) {
445 if (offset < 0 || offset > shmfd->shm_size)
446 error = EINVAL;
447 else
448 td->td_uretoff.tdu_off = offset;
449 }
450 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
451 return (error);
452 }
453
454 static int
shm_read(struct file * fp,struct uio * uio,struct ucred * active_cred,int flags,struct thread * td)455 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
456 int flags, struct thread *td)
457 {
458 struct shmfd *shmfd;
459 void *rl_cookie;
460 int error;
461
462 shmfd = fp->f_data;
463 #ifdef MAC
464 error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
465 if (error)
466 return (error);
467 #endif
468 foffset_lock_uio(fp, uio, flags);
469 rl_cookie = shm_rangelock_rlock(shmfd, uio->uio_offset,
470 uio->uio_offset + uio->uio_resid);
471 error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
472 shm_rangelock_unlock(shmfd, rl_cookie);
473 foffset_unlock_uio(fp, uio, flags);
474 return (error);
475 }
476
477 static int
shm_write(struct file * fp,struct uio * uio,struct ucred * active_cred,int flags,struct thread * td)478 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
479 int flags, struct thread *td)
480 {
481 struct shmfd *shmfd;
482 void *rl_cookie;
483 int error;
484 off_t size;
485
486 shmfd = fp->f_data;
487 #ifdef MAC
488 error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
489 if (error)
490 return (error);
491 #endif
492 if (shm_largepage(shmfd) && shmfd->shm_lp_psind == 0)
493 return (EINVAL);
494 foffset_lock_uio(fp, uio, flags);
495 if (uio->uio_resid > OFF_MAX - uio->uio_offset) {
496 /*
497 * Overflow is only an error if we're supposed to expand on
498 * write. Otherwise, we'll just truncate the write to the
499 * size of the file, which can only grow up to OFF_MAX.
500 */
501 if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0) {
502 foffset_unlock_uio(fp, uio, flags);
503 return (EFBIG);
504 }
505
506 size = shmfd->shm_size;
507 } else {
508 size = uio->uio_offset + uio->uio_resid;
509 }
510 if ((flags & FOF_OFFSET) == 0)
511 rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
512 else
513 rl_cookie = shm_rangelock_wlock(shmfd, uio->uio_offset, size);
514 if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) {
515 error = EPERM;
516 } else {
517 error = 0;
518 if ((shmfd->shm_flags & SHM_GROW_ON_WRITE) != 0 &&
519 size > shmfd->shm_size) {
520 error = shm_dotruncate_cookie(shmfd, size, rl_cookie);
521 }
522 if (error == 0)
523 error = uiomove_object(shmfd->shm_object,
524 shmfd->shm_size, uio);
525 }
526 shm_rangelock_unlock(shmfd, rl_cookie);
527 foffset_unlock_uio(fp, uio, flags);
528 return (error);
529 }
530
531 static int
shm_truncate(struct file * fp,off_t length,struct ucred * active_cred,struct thread * td)532 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
533 struct thread *td)
534 {
535 struct shmfd *shmfd;
536 #ifdef MAC
537 int error;
538 #endif
539
540 shmfd = fp->f_data;
541 #ifdef MAC
542 error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
543 if (error)
544 return (error);
545 #endif
546 return (shm_dotruncate(shmfd, length));
547 }
548
549 int
shm_ioctl(struct file * fp,u_long com,void * data,struct ucred * active_cred,struct thread * td)550 shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
551 struct thread *td)
552 {
553 struct shmfd *shmfd;
554 struct shm_largepage_conf *conf;
555 void *rl_cookie;
556
557 shmfd = fp->f_data;
558 switch (com) {
559 case FIONBIO:
560 case FIOASYNC:
561 /*
562 * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work,
563 * just like it would on an unlinked regular file
564 */
565 return (0);
566 case FIOSSHMLPGCNF:
567 if (!shm_largepage(shmfd))
568 return (ENOTTY);
569 conf = data;
570 if (shmfd->shm_lp_psind != 0 &&
571 conf->psind != shmfd->shm_lp_psind)
572 return (EINVAL);
573 if (conf->psind <= 0 || conf->psind >= MAXPAGESIZES ||
574 pagesizes[conf->psind] == 0)
575 return (EINVAL);
576 if (conf->alloc_policy != SHM_LARGEPAGE_ALLOC_DEFAULT &&
577 conf->alloc_policy != SHM_LARGEPAGE_ALLOC_NOWAIT &&
578 conf->alloc_policy != SHM_LARGEPAGE_ALLOC_HARD)
579 return (EINVAL);
580
581 rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
582 shmfd->shm_lp_psind = conf->psind;
583 shmfd->shm_lp_alloc_policy = conf->alloc_policy;
584 shmfd->shm_object->un_pager.phys.data_val = conf->psind;
585 shm_rangelock_unlock(shmfd, rl_cookie);
586 return (0);
587 case FIOGSHMLPGCNF:
588 if (!shm_largepage(shmfd))
589 return (ENOTTY);
590 conf = data;
591 rl_cookie = shm_rangelock_rlock(shmfd, 0, OFF_MAX);
592 conf->psind = shmfd->shm_lp_psind;
593 conf->alloc_policy = shmfd->shm_lp_alloc_policy;
594 shm_rangelock_unlock(shmfd, rl_cookie);
595 return (0);
596 default:
597 return (ENOTTY);
598 }
599 }
600
601 static int
shm_stat(struct file * fp,struct stat * sb,struct ucred * active_cred)602 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred)
603 {
604 struct shmfd *shmfd;
605 #ifdef MAC
606 int error;
607 #endif
608
609 shmfd = fp->f_data;
610
611 #ifdef MAC
612 error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
613 if (error)
614 return (error);
615 #endif
616
617 /*
618 * Attempt to return sanish values for fstat() on a memory file
619 * descriptor.
620 */
621 bzero(sb, sizeof(*sb));
622 sb->st_blksize = PAGE_SIZE;
623 sb->st_size = shmfd->shm_size;
624 mtx_lock(&shm_timestamp_lock);
625 sb->st_atim = shmfd->shm_atime;
626 sb->st_ctim = shmfd->shm_ctime;
627 sb->st_mtim = shmfd->shm_mtime;
628 sb->st_birthtim = shmfd->shm_birthtime;
629 sb->st_mode = S_IFREG | shmfd->shm_mode; /* XXX */
630 sb->st_uid = shmfd->shm_uid;
631 sb->st_gid = shmfd->shm_gid;
632 mtx_unlock(&shm_timestamp_lock);
633 sb->st_dev = shm_dev_ino;
634 sb->st_ino = shmfd->shm_ino;
635 sb->st_nlink = shmfd->shm_object->ref_count;
636 if (shm_largepage(shmfd)) {
637 sb->st_blocks = shmfd->shm_object->size /
638 (pagesizes[shmfd->shm_lp_psind] >> PAGE_SHIFT);
639 } else {
640 sb->st_blocks = shmfd->shm_pages;
641 }
642
643 return (0);
644 }
645
646 static int
shm_close(struct file * fp,struct thread * td)647 shm_close(struct file *fp, struct thread *td)
648 {
649 struct shmfd *shmfd;
650
651 shmfd = fp->f_data;
652 fp->f_data = NULL;
653 shm_drop(shmfd);
654
655 return (0);
656 }
657
658 static int
shm_copyin_path(struct thread * td,const char * userpath_in,char ** path_out)659 shm_copyin_path(struct thread *td, const char *userpath_in, char **path_out) {
660 int error;
661 char *path;
662 const char *pr_path;
663 size_t pr_pathlen;
664
665 path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
666 pr_path = td->td_ucred->cr_prison->pr_path;
667
668 /* Construct a full pathname for jailed callers. */
669 pr_pathlen = strcmp(pr_path, "/") ==
670 0 ? 0 : strlcpy(path, pr_path, MAXPATHLEN);
671 error = copyinstr(userpath_in, path + pr_pathlen,
672 MAXPATHLEN - pr_pathlen, NULL);
673 if (error != 0)
674 goto out;
675
676 #ifdef KTRACE
677 if (KTRPOINT(curthread, KTR_NAMEI))
678 ktrnamei(path);
679 #endif
680
681 /* Require paths to start with a '/' character. */
682 if (path[pr_pathlen] != '/') {
683 error = EINVAL;
684 goto out;
685 }
686
687 *path_out = path;
688
689 out:
690 if (error != 0)
691 free(path, M_SHMFD);
692
693 return (error);
694 }
695
696 static int
shm_partial_page_invalidate(vm_object_t object,vm_pindex_t idx,int base,int end)697 shm_partial_page_invalidate(vm_object_t object, vm_pindex_t idx, int base,
698 int end)
699 {
700 int error;
701
702 error = vm_page_grab_zero_partial(object, idx, base, end);
703 if (error == EIO)
704 VM_OBJECT_WUNLOCK(object);
705 return (error);
706 }
707
708 static int
shm_dotruncate_locked(struct shmfd * shmfd,off_t length,void * rl_cookie)709 shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie)
710 {
711 vm_object_t object;
712 vm_pindex_t nobjsize;
713 vm_ooffset_t delta;
714 int base, error;
715
716 KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
717 object = shmfd->shm_object;
718 VM_OBJECT_ASSERT_WLOCKED(object);
719 rangelock_cookie_assert(rl_cookie, RA_WLOCKED);
720 if (length == shmfd->shm_size)
721 return (0);
722 nobjsize = OFF_TO_IDX(length + PAGE_MASK);
723
724 /* Are we shrinking? If so, trim the end. */
725 if (length < shmfd->shm_size) {
726 if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0)
727 return (EPERM);
728
729 /*
730 * Disallow any requests to shrink the size if this
731 * object is mapped into the kernel.
732 */
733 if (shmfd->shm_kmappings > 0)
734 return (EBUSY);
735
736 /*
737 * Zero the truncated part of the last page.
738 */
739 base = length & PAGE_MASK;
740 if (base != 0) {
741 error = shm_partial_page_invalidate(object,
742 OFF_TO_IDX(length), base, PAGE_SIZE);
743 if (error)
744 return (error);
745 }
746 delta = IDX_TO_OFF(object->size - nobjsize);
747
748 if (nobjsize < object->size)
749 vm_object_page_remove(object, nobjsize, object->size,
750 0);
751
752 /* Free the swap accounted for shm */
753 swap_release_by_cred(delta, object->cred);
754 object->charge -= delta;
755 } else {
756 if ((shmfd->shm_seals & F_SEAL_GROW) != 0)
757 return (EPERM);
758
759 /* Try to reserve additional swap space. */
760 delta = IDX_TO_OFF(nobjsize - object->size);
761 if (!swap_reserve_by_cred(delta, object->cred))
762 return (ENOMEM);
763 object->charge += delta;
764 }
765 shmfd->shm_size = length;
766 mtx_lock(&shm_timestamp_lock);
767 vfs_timestamp(&shmfd->shm_ctime);
768 shmfd->shm_mtime = shmfd->shm_ctime;
769 mtx_unlock(&shm_timestamp_lock);
770 object->size = nobjsize;
771 return (0);
772 }
773
774 static int
shm_dotruncate_largepage(struct shmfd * shmfd,off_t length,void * rl_cookie)775 shm_dotruncate_largepage(struct shmfd *shmfd, off_t length, void *rl_cookie)
776 {
777 vm_object_t object;
778 vm_page_t m;
779 vm_pindex_t newobjsz;
780 vm_pindex_t oldobjsz __unused;
781 int aflags, error, i, psind, try;
782
783 KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
784 object = shmfd->shm_object;
785 VM_OBJECT_ASSERT_WLOCKED(object);
786 rangelock_cookie_assert(rl_cookie, RA_WLOCKED);
787
788 oldobjsz = object->size;
789 newobjsz = OFF_TO_IDX(length);
790 if (length == shmfd->shm_size)
791 return (0);
792 psind = shmfd->shm_lp_psind;
793 if (psind == 0 && length != 0)
794 return (EINVAL);
795 if ((length & (pagesizes[psind] - 1)) != 0)
796 return (EINVAL);
797
798 if (length < shmfd->shm_size) {
799 if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0)
800 return (EPERM);
801 if (shmfd->shm_kmappings > 0)
802 return (EBUSY);
803 return (ENOTSUP); /* Pages are unmanaged. */
804 #if 0
805 vm_object_page_remove(object, newobjsz, oldobjsz, 0);
806 object->size = newobjsz;
807 shmfd->shm_size = length;
808 return (0);
809 #endif
810 }
811
812 if ((shmfd->shm_seals & F_SEAL_GROW) != 0)
813 return (EPERM);
814
815 aflags = VM_ALLOC_NORMAL | VM_ALLOC_ZERO;
816 if (shmfd->shm_lp_alloc_policy == SHM_LARGEPAGE_ALLOC_NOWAIT)
817 aflags |= VM_ALLOC_WAITFAIL;
818 try = 0;
819
820 /*
821 * Extend shmfd and object, keeping all already fully
822 * allocated large pages intact even on error, because dropped
823 * object lock might allowed mapping of them.
824 */
825 while (object->size < newobjsz) {
826 m = vm_page_alloc_contig(object, object->size, aflags,
827 pagesizes[psind] / PAGE_SIZE, 0, ~0,
828 pagesizes[psind], 0,
829 VM_MEMATTR_DEFAULT);
830 if (m == NULL) {
831 VM_OBJECT_WUNLOCK(object);
832 if (shmfd->shm_lp_alloc_policy ==
833 SHM_LARGEPAGE_ALLOC_NOWAIT ||
834 (shmfd->shm_lp_alloc_policy ==
835 SHM_LARGEPAGE_ALLOC_DEFAULT &&
836 try >= largepage_reclaim_tries)) {
837 VM_OBJECT_WLOCK(object);
838 return (ENOMEM);
839 }
840 error = vm_page_reclaim_contig(aflags,
841 pagesizes[psind] / PAGE_SIZE, 0, ~0,
842 pagesizes[psind], 0);
843 if (error == ENOMEM)
844 error = vm_wait_intr(object);
845 if (error != 0) {
846 VM_OBJECT_WLOCK(object);
847 return (error);
848 }
849 try++;
850 VM_OBJECT_WLOCK(object);
851 continue;
852 }
853 try = 0;
854 for (i = 0; i < pagesizes[psind] / PAGE_SIZE; i++) {
855 if ((m[i].flags & PG_ZERO) == 0)
856 pmap_zero_page(&m[i]);
857 vm_page_valid(&m[i]);
858 vm_page_xunbusy(&m[i]);
859 }
860 object->size += OFF_TO_IDX(pagesizes[psind]);
861 shmfd->shm_size += pagesizes[psind];
862 atomic_add_long(&count_largepages[psind], 1);
863 vm_wire_add(atop(pagesizes[psind]));
864 }
865 return (0);
866 }
867
868 static int
shm_dotruncate_cookie(struct shmfd * shmfd,off_t length,void * rl_cookie)869 shm_dotruncate_cookie(struct shmfd *shmfd, off_t length, void *rl_cookie)
870 {
871 int error;
872
873 VM_OBJECT_WLOCK(shmfd->shm_object);
874 error = shm_largepage(shmfd) ? shm_dotruncate_largepage(shmfd,
875 length, rl_cookie) : shm_dotruncate_locked(shmfd, length,
876 rl_cookie);
877 VM_OBJECT_WUNLOCK(shmfd->shm_object);
878 return (error);
879 }
880
881 int
shm_dotruncate(struct shmfd * shmfd,off_t length)882 shm_dotruncate(struct shmfd *shmfd, off_t length)
883 {
884 void *rl_cookie;
885 int error;
886
887 rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
888 error = shm_dotruncate_cookie(shmfd, length, rl_cookie);
889 shm_rangelock_unlock(shmfd, rl_cookie);
890 return (error);
891 }
892
893 /*
894 * shmfd object management including creation and reference counting
895 * routines.
896 */
897 struct shmfd *
shm_alloc(struct ucred * ucred,mode_t mode,bool largepage)898 shm_alloc(struct ucred *ucred, mode_t mode, bool largepage)
899 {
900 struct shmfd *shmfd;
901 vm_object_t obj;
902
903 if (largepage) {
904 obj = phys_pager_allocate(NULL, &shm_largepage_phys_ops,
905 NULL, 0, VM_PROT_DEFAULT, 0, ucred);
906 } else {
907 obj = vm_pager_allocate(shmfd_pager_type, NULL, 0,
908 VM_PROT_DEFAULT, 0, ucred);
909 }
910 if (obj == NULL) {
911 /*
912 * swap reservation limits can cause object allocation
913 * to fail.
914 */
915 return (NULL);
916 }
917
918 shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
919 shmfd->shm_uid = ucred->cr_uid;
920 shmfd->shm_gid = ucred->cr_gid;
921 shmfd->shm_mode = mode;
922 if (largepage) {
923 obj->un_pager.phys.phys_priv = shmfd;
924 shmfd->shm_lp_alloc_policy = SHM_LARGEPAGE_ALLOC_DEFAULT;
925 } else {
926 obj->un_pager.swp.swp_priv = shmfd;
927 }
928
929 VM_OBJECT_WLOCK(obj);
930 vm_object_set_flag(obj, OBJ_POSIXSHM);
931 VM_OBJECT_WUNLOCK(obj);
932 shmfd->shm_object = obj;
933 vfs_timestamp(&shmfd->shm_birthtime);
934 shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
935 shmfd->shm_birthtime;
936 shmfd->shm_ino = alloc_unr64(&shm_ino_unr);
937 refcount_init(&shmfd->shm_refs, 1);
938 mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
939 rangelock_init(&shmfd->shm_rl);
940 #ifdef MAC
941 mac_posixshm_init(shmfd);
942 mac_posixshm_create(ucred, shmfd);
943 #endif
944
945 return (shmfd);
946 }
947
948 struct shmfd *
shm_hold(struct shmfd * shmfd)949 shm_hold(struct shmfd *shmfd)
950 {
951
952 refcount_acquire(&shmfd->shm_refs);
953 return (shmfd);
954 }
955
956 void
shm_drop(struct shmfd * shmfd)957 shm_drop(struct shmfd *shmfd)
958 {
959 vm_object_t obj;
960
961 if (refcount_release(&shmfd->shm_refs)) {
962 #ifdef MAC
963 mac_posixshm_destroy(shmfd);
964 #endif
965 rangelock_destroy(&shmfd->shm_rl);
966 mtx_destroy(&shmfd->shm_mtx);
967 obj = shmfd->shm_object;
968 VM_OBJECT_WLOCK(obj);
969 if (shm_largepage(shmfd))
970 obj->un_pager.phys.phys_priv = NULL;
971 else
972 obj->un_pager.swp.swp_priv = NULL;
973 VM_OBJECT_WUNLOCK(obj);
974 vm_object_deallocate(obj);
975 free(shmfd, M_SHMFD);
976 }
977 }
978
979 /*
980 * Determine if the credentials have sufficient permissions for a
981 * specified combination of FREAD and FWRITE.
982 */
983 int
shm_access(struct shmfd * shmfd,struct ucred * ucred,int flags)984 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
985 {
986 accmode_t accmode;
987 int error;
988
989 accmode = 0;
990 if (flags & FREAD)
991 accmode |= VREAD;
992 if (flags & FWRITE)
993 accmode |= VWRITE;
994 mtx_lock(&shm_timestamp_lock);
995 error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
996 accmode, ucred);
997 mtx_unlock(&shm_timestamp_lock);
998 return (error);
999 }
1000
1001 static void
shm_init(void * arg)1002 shm_init(void *arg)
1003 {
1004 char name[32];
1005 int i;
1006
1007 mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
1008 sx_init(&shm_dict_lock, "shm dictionary");
1009 shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
1010 new_unrhdr64(&shm_ino_unr, 1);
1011 shm_dev_ino = devfs_alloc_cdp_inode();
1012 KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
1013 shmfd_pager_type = vm_pager_alloc_dyn_type(&shm_swap_pager_ops,
1014 OBJT_SWAP);
1015 MPASS(shmfd_pager_type != -1);
1016
1017 for (i = 1; i < MAXPAGESIZES; i++) {
1018 if (pagesizes[i] == 0)
1019 break;
1020 #define M (1024 * 1024)
1021 #define G (1024 * M)
1022 if (pagesizes[i] >= G)
1023 snprintf(name, sizeof(name), "%luG", pagesizes[i] / G);
1024 else if (pagesizes[i] >= M)
1025 snprintf(name, sizeof(name), "%luM", pagesizes[i] / M);
1026 else
1027 snprintf(name, sizeof(name), "%lu", pagesizes[i]);
1028 #undef G
1029 #undef M
1030 SYSCTL_ADD_ULONG(NULL, SYSCTL_STATIC_CHILDREN(_vm_largepages),
1031 OID_AUTO, name, CTLFLAG_RD, &count_largepages[i],
1032 "number of non-transient largepages allocated");
1033 }
1034 }
1035 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
1036
1037 /*
1038 * Remove all shared memory objects that belong to a prison.
1039 */
1040 void
shm_remove_prison(struct prison * pr)1041 shm_remove_prison(struct prison *pr)
1042 {
1043 struct shm_mapping *shmm, *tshmm;
1044 u_long i;
1045
1046 sx_xlock(&shm_dict_lock);
1047 for (i = 0; i < shm_hash + 1; i++) {
1048 LIST_FOREACH_SAFE(shmm, &shm_dictionary[i], sm_link, tshmm) {
1049 if (shmm->sm_shmfd->shm_object->cred &&
1050 shmm->sm_shmfd->shm_object->cred->cr_prison == pr)
1051 shm_doremove(shmm);
1052 }
1053 }
1054 sx_xunlock(&shm_dict_lock);
1055 }
1056
1057 /*
1058 * Dictionary management. We maintain an in-kernel dictionary to map
1059 * paths to shmfd objects. We use the FNV hash on the path to store
1060 * the mappings in a hash table.
1061 */
1062 static struct shmfd *
shm_lookup(char * path,Fnv32_t fnv)1063 shm_lookup(char *path, Fnv32_t fnv)
1064 {
1065 struct shm_mapping *map;
1066
1067 LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
1068 if (map->sm_fnv != fnv)
1069 continue;
1070 if (strcmp(map->sm_path, path) == 0)
1071 return (map->sm_shmfd);
1072 }
1073
1074 return (NULL);
1075 }
1076
1077 static void
shm_insert(char * path,Fnv32_t fnv,struct shmfd * shmfd)1078 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
1079 {
1080 struct shm_mapping *map;
1081
1082 map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
1083 map->sm_path = path;
1084 map->sm_fnv = fnv;
1085 map->sm_shmfd = shm_hold(shmfd);
1086 shmfd->shm_path = path;
1087 LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
1088 }
1089
1090 static int
shm_remove(char * path,Fnv32_t fnv,struct ucred * ucred)1091 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
1092 {
1093 struct shm_mapping *map;
1094 int error;
1095
1096 LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
1097 if (map->sm_fnv != fnv)
1098 continue;
1099 if (strcmp(map->sm_path, path) == 0) {
1100 #ifdef MAC
1101 error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
1102 if (error)
1103 return (error);
1104 #endif
1105 error = shm_access(map->sm_shmfd, ucred,
1106 FREAD | FWRITE);
1107 if (error)
1108 return (error);
1109 shm_doremove(map);
1110 return (0);
1111 }
1112 }
1113
1114 return (ENOENT);
1115 }
1116
1117 static void
shm_doremove(struct shm_mapping * map)1118 shm_doremove(struct shm_mapping *map)
1119 {
1120 map->sm_shmfd->shm_path = NULL;
1121 LIST_REMOVE(map, sm_link);
1122 shm_drop(map->sm_shmfd);
1123 free(map->sm_path, M_SHMFD);
1124 free(map, M_SHMFD);
1125 }
1126
1127 int
kern_shm_open2(struct thread * td,const char * userpath,int flags,mode_t mode,int shmflags,struct filecaps * fcaps,const char * name __unused)1128 kern_shm_open2(struct thread *td, const char *userpath, int flags, mode_t mode,
1129 int shmflags, struct filecaps *fcaps, const char *name __unused)
1130 {
1131 struct pwddesc *pdp;
1132 struct shmfd *shmfd;
1133 struct file *fp;
1134 char *path;
1135 void *rl_cookie;
1136 Fnv32_t fnv;
1137 mode_t cmode;
1138 int error, fd, initial_seals;
1139 bool largepage;
1140
1141 if ((shmflags & ~(SHM_ALLOW_SEALING | SHM_GROW_ON_WRITE |
1142 SHM_LARGEPAGE)) != 0)
1143 return (EINVAL);
1144
1145 initial_seals = F_SEAL_SEAL;
1146 if ((shmflags & SHM_ALLOW_SEALING) != 0)
1147 initial_seals &= ~F_SEAL_SEAL;
1148
1149 AUDIT_ARG_FFLAGS(flags);
1150 AUDIT_ARG_MODE(mode);
1151
1152 if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
1153 return (EINVAL);
1154
1155 if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
1156 return (EINVAL);
1157
1158 largepage = (shmflags & SHM_LARGEPAGE) != 0;
1159 if (largepage && !PMAP_HAS_LARGEPAGES)
1160 return (ENOTTY);
1161
1162 /*
1163 * Currently only F_SEAL_SEAL may be set when creating or opening shmfd.
1164 * If the decision is made later to allow additional seals, care must be
1165 * taken below to ensure that the seals are properly set if the shmfd
1166 * already existed -- this currently assumes that only F_SEAL_SEAL can
1167 * be set and doesn't take further precautions to ensure the validity of
1168 * the seals being added with respect to current mappings.
1169 */
1170 if ((initial_seals & ~F_SEAL_SEAL) != 0)
1171 return (EINVAL);
1172
1173 if (userpath != SHM_ANON) {
1174 error = shm_copyin_path(td, userpath, &path);
1175 if (error != 0)
1176 return (error);
1177
1178 #ifdef CAPABILITY_MODE
1179 /*
1180 * shm_open(2) is only allowed for anonymous objects.
1181 */
1182 if (CAP_TRACING(td))
1183 ktrcapfail(CAPFAIL_NAMEI, path);
1184 if (IN_CAPABILITY_MODE(td)) {
1185 error = ECAPMODE;
1186 goto outnofp;
1187 }
1188 #endif
1189
1190 AUDIT_ARG_UPATH1_CANON(path);
1191 } else {
1192 path = NULL;
1193 }
1194
1195 pdp = td->td_proc->p_pd;
1196 cmode = (mode & ~pdp->pd_cmask) & ACCESSPERMS;
1197
1198 /*
1199 * shm_open(2) created shm should always have O_CLOEXEC set, as mandated
1200 * by POSIX. We allow it to be unset here so that an in-kernel
1201 * interface may be written as a thin layer around shm, optionally not
1202 * setting CLOEXEC. For shm_open(2), O_CLOEXEC is set unconditionally
1203 * in sys_shm_open() to keep this implementation compliant.
1204 */
1205 error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps);
1206 if (error != 0)
1207 goto outnofp;
1208
1209 /* A SHM_ANON path pointer creates an anonymous object. */
1210 if (userpath == SHM_ANON) {
1211 /* A read-only anonymous object is pointless. */
1212 if ((flags & O_ACCMODE) == O_RDONLY) {
1213 error = EINVAL;
1214 goto out;
1215 }
1216 shmfd = shm_alloc(td->td_ucred, cmode, largepage);
1217 if (shmfd == NULL) {
1218 error = ENOMEM;
1219 goto out;
1220 }
1221 shmfd->shm_seals = initial_seals;
1222 shmfd->shm_flags = shmflags;
1223 } else {
1224 fnv = fnv_32_str(path, FNV1_32_INIT);
1225 sx_xlock(&shm_dict_lock);
1226 shmfd = shm_lookup(path, fnv);
1227 if (shmfd == NULL) {
1228 /* Object does not yet exist, create it if requested. */
1229 if (flags & O_CREAT) {
1230 #ifdef MAC
1231 error = mac_posixshm_check_create(td->td_ucred,
1232 path);
1233 if (error == 0) {
1234 #endif
1235 shmfd = shm_alloc(td->td_ucred, cmode,
1236 largepage);
1237 if (shmfd == NULL) {
1238 error = ENOMEM;
1239 } else {
1240 shmfd->shm_seals =
1241 initial_seals;
1242 shmfd->shm_flags = shmflags;
1243 shm_insert(path, fnv, shmfd);
1244 path = NULL;
1245 }
1246 #ifdef MAC
1247 }
1248 #endif
1249 } else {
1250 error = ENOENT;
1251 }
1252 } else {
1253 /*
1254 * Object already exists, obtain a new reference if
1255 * requested and permitted.
1256 */
1257 rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
1258
1259 /*
1260 * kern_shm_open() likely shouldn't ever error out on
1261 * trying to set a seal that already exists, unlike
1262 * F_ADD_SEALS. This would break terribly as
1263 * shm_open(2) actually sets F_SEAL_SEAL to maintain
1264 * historical behavior where the underlying file could
1265 * not be sealed.
1266 */
1267 initial_seals &= ~shmfd->shm_seals;
1268
1269 /*
1270 * initial_seals can't set additional seals if we've
1271 * already been set F_SEAL_SEAL. If F_SEAL_SEAL is set,
1272 * then we've already removed that one from
1273 * initial_seals. This is currently redundant as we
1274 * only allow setting F_SEAL_SEAL at creation time, but
1275 * it's cheap to check and decreases the effort required
1276 * to allow additional seals.
1277 */
1278 if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 &&
1279 initial_seals != 0)
1280 error = EPERM;
1281 else if ((flags & (O_CREAT | O_EXCL)) ==
1282 (O_CREAT | O_EXCL))
1283 error = EEXIST;
1284 else if (shmflags != 0 && shmflags != shmfd->shm_flags)
1285 error = EINVAL;
1286 else {
1287 #ifdef MAC
1288 error = mac_posixshm_check_open(td->td_ucred,
1289 shmfd, FFLAGS(flags & O_ACCMODE));
1290 if (error == 0)
1291 #endif
1292 error = shm_access(shmfd, td->td_ucred,
1293 FFLAGS(flags & O_ACCMODE));
1294 }
1295
1296 /*
1297 * Truncate the file back to zero length if
1298 * O_TRUNC was specified and the object was
1299 * opened with read/write.
1300 */
1301 if (error == 0 &&
1302 (flags & (O_ACCMODE | O_TRUNC)) ==
1303 (O_RDWR | O_TRUNC)) {
1304 VM_OBJECT_WLOCK(shmfd->shm_object);
1305 #ifdef MAC
1306 error = mac_posixshm_check_truncate(
1307 td->td_ucred, fp->f_cred, shmfd);
1308 if (error == 0)
1309 #endif
1310 error = shm_dotruncate_locked(shmfd, 0,
1311 rl_cookie);
1312 VM_OBJECT_WUNLOCK(shmfd->shm_object);
1313 }
1314 if (error == 0) {
1315 /*
1316 * Currently we only allow F_SEAL_SEAL to be
1317 * set initially. As noted above, this would
1318 * need to be reworked should that change.
1319 */
1320 shmfd->shm_seals |= initial_seals;
1321 shm_hold(shmfd);
1322 }
1323 shm_rangelock_unlock(shmfd, rl_cookie);
1324 }
1325 sx_xunlock(&shm_dict_lock);
1326
1327 if (error != 0)
1328 goto out;
1329 }
1330
1331 finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
1332
1333 td->td_retval[0] = fd;
1334 fdrop(fp, td);
1335 free(path, M_SHMFD);
1336
1337 return (0);
1338
1339 out:
1340 fdclose(td, fp, fd);
1341 fdrop(fp, td);
1342 outnofp:
1343 free(path, M_SHMFD);
1344
1345 return (error);
1346 }
1347
1348 /* System calls. */
1349 #ifdef COMPAT_FREEBSD12
1350 int
freebsd12_shm_open(struct thread * td,struct freebsd12_shm_open_args * uap)1351 freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap)
1352 {
1353
1354 return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC,
1355 uap->mode, NULL));
1356 }
1357 #endif
1358
1359 int
sys_shm_unlink(struct thread * td,struct shm_unlink_args * uap)1360 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
1361 {
1362 char *path;
1363 Fnv32_t fnv;
1364 int error;
1365
1366 error = shm_copyin_path(td, uap->path, &path);
1367 if (error != 0)
1368 return (error);
1369
1370 AUDIT_ARG_UPATH1_CANON(path);
1371 fnv = fnv_32_str(path, FNV1_32_INIT);
1372 sx_xlock(&shm_dict_lock);
1373 error = shm_remove(path, fnv, td->td_ucred);
1374 sx_xunlock(&shm_dict_lock);
1375 free(path, M_SHMFD);
1376
1377 return (error);
1378 }
1379
1380 int
sys_shm_rename(struct thread * td,struct shm_rename_args * uap)1381 sys_shm_rename(struct thread *td, struct shm_rename_args *uap)
1382 {
1383 char *path_from = NULL, *path_to = NULL;
1384 Fnv32_t fnv_from, fnv_to;
1385 struct shmfd *fd_from;
1386 struct shmfd *fd_to;
1387 int error;
1388 int flags;
1389
1390 flags = uap->flags;
1391 AUDIT_ARG_FFLAGS(flags);
1392
1393 /*
1394 * Make sure the user passed only valid flags.
1395 * If you add a new flag, please add a new term here.
1396 */
1397 if ((flags & ~(
1398 SHM_RENAME_NOREPLACE |
1399 SHM_RENAME_EXCHANGE
1400 )) != 0) {
1401 error = EINVAL;
1402 goto out;
1403 }
1404
1405 /*
1406 * EXCHANGE and NOREPLACE don't quite make sense together. Let's
1407 * force the user to choose one or the other.
1408 */
1409 if ((flags & SHM_RENAME_NOREPLACE) != 0 &&
1410 (flags & SHM_RENAME_EXCHANGE) != 0) {
1411 error = EINVAL;
1412 goto out;
1413 }
1414
1415 /* Renaming to or from anonymous makes no sense */
1416 if (uap->path_from == SHM_ANON || uap->path_to == SHM_ANON) {
1417 error = EINVAL;
1418 goto out;
1419 }
1420
1421 error = shm_copyin_path(td, uap->path_from, &path_from);
1422 if (error != 0)
1423 goto out;
1424
1425 error = shm_copyin_path(td, uap->path_to, &path_to);
1426 if (error != 0)
1427 goto out;
1428
1429 AUDIT_ARG_UPATH1_CANON(path_from);
1430 AUDIT_ARG_UPATH2_CANON(path_to);
1431
1432 /* Rename with from/to equal is a no-op */
1433 if (strcmp(path_from, path_to) == 0)
1434 goto out;
1435
1436 fnv_from = fnv_32_str(path_from, FNV1_32_INIT);
1437 fnv_to = fnv_32_str(path_to, FNV1_32_INIT);
1438
1439 sx_xlock(&shm_dict_lock);
1440
1441 fd_from = shm_lookup(path_from, fnv_from);
1442 if (fd_from == NULL) {
1443 error = ENOENT;
1444 goto out_locked;
1445 }
1446
1447 fd_to = shm_lookup(path_to, fnv_to);
1448 if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) {
1449 error = EEXIST;
1450 goto out_locked;
1451 }
1452
1453 /*
1454 * Unconditionally prevents shm_remove from invalidating the 'from'
1455 * shm's state.
1456 */
1457 shm_hold(fd_from);
1458 error = shm_remove(path_from, fnv_from, td->td_ucred);
1459
1460 /*
1461 * One of my assumptions failed if ENOENT (e.g. locking didn't
1462 * protect us)
1463 */
1464 KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s",
1465 path_from));
1466 if (error != 0) {
1467 shm_drop(fd_from);
1468 goto out_locked;
1469 }
1470
1471 /*
1472 * If we are exchanging, we need to ensure the shm_remove below
1473 * doesn't invalidate the dest shm's state.
1474 */
1475 if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL)
1476 shm_hold(fd_to);
1477
1478 /*
1479 * NOTE: if path_to is not already in the hash, c'est la vie;
1480 * it simply means we have nothing already at path_to to unlink.
1481 * That is the ENOENT case.
1482 *
1483 * If we somehow don't have access to unlink this guy, but
1484 * did for the shm at path_from, then relink the shm to path_from
1485 * and abort with EACCES.
1486 *
1487 * All other errors: that is weird; let's relink and abort the
1488 * operation.
1489 */
1490 error = shm_remove(path_to, fnv_to, td->td_ucred);
1491 if (error != 0 && error != ENOENT) {
1492 shm_insert(path_from, fnv_from, fd_from);
1493 shm_drop(fd_from);
1494 /* Don't free path_from now, since the hash references it */
1495 path_from = NULL;
1496 goto out_locked;
1497 }
1498
1499 error = 0;
1500
1501 shm_insert(path_to, fnv_to, fd_from);
1502
1503 /* Don't free path_to now, since the hash references it */
1504 path_to = NULL;
1505
1506 /* We kept a ref when we removed, and incremented again in insert */
1507 shm_drop(fd_from);
1508 KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1509 fd_from->shm_refs));
1510
1511 if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) {
1512 shm_insert(path_from, fnv_from, fd_to);
1513 path_from = NULL;
1514 shm_drop(fd_to);
1515 KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1516 fd_to->shm_refs));
1517 }
1518
1519 out_locked:
1520 sx_xunlock(&shm_dict_lock);
1521
1522 out:
1523 free(path_from, M_SHMFD);
1524 free(path_to, M_SHMFD);
1525 return (error);
1526 }
1527
1528 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)1529 shm_mmap_large(struct shmfd *shmfd, vm_map_t map, vm_offset_t *addr,
1530 vm_size_t size, vm_prot_t prot, vm_prot_t max_prot, int flags,
1531 vm_ooffset_t foff, struct thread *td)
1532 {
1533 struct vmspace *vms;
1534 vm_map_entry_t next_entry, prev_entry;
1535 vm_offset_t align, mask, maxaddr;
1536 int docow, error, rv, try;
1537 bool curmap;
1538
1539 if (shmfd->shm_lp_psind == 0)
1540 return (EINVAL);
1541
1542 /* MAP_PRIVATE is disabled */
1543 if ((flags & ~(MAP_SHARED | MAP_FIXED | MAP_EXCL |
1544 MAP_NOCORE | MAP_32BIT | MAP_ALIGNMENT_MASK)) != 0)
1545 return (EINVAL);
1546
1547 vms = td->td_proc->p_vmspace;
1548 curmap = map == &vms->vm_map;
1549 if (curmap) {
1550 error = kern_mmap_racct_check(td, map, size);
1551 if (error != 0)
1552 return (error);
1553 }
1554
1555 docow = shmfd->shm_lp_psind << MAP_SPLIT_BOUNDARY_SHIFT;
1556 docow |= MAP_INHERIT_SHARE;
1557 if ((flags & MAP_NOCORE) != 0)
1558 docow |= MAP_DISABLE_COREDUMP;
1559
1560 mask = pagesizes[shmfd->shm_lp_psind] - 1;
1561 if ((foff & mask) != 0)
1562 return (EINVAL);
1563 maxaddr = vm_map_max(map);
1564 if ((flags & MAP_32BIT) != 0 && maxaddr > MAP_32BIT_MAX_ADDR)
1565 maxaddr = MAP_32BIT_MAX_ADDR;
1566 if (size == 0 || (size & mask) != 0 ||
1567 (*addr != 0 && ((*addr & mask) != 0 ||
1568 *addr + size < *addr || *addr + size > maxaddr)))
1569 return (EINVAL);
1570
1571 align = flags & MAP_ALIGNMENT_MASK;
1572 if (align == 0) {
1573 align = pagesizes[shmfd->shm_lp_psind];
1574 } else if (align == MAP_ALIGNED_SUPER) {
1575 /*
1576 * MAP_ALIGNED_SUPER is only supported on superpage sizes,
1577 * i.e., [1, VM_NRESERVLEVEL]. shmfd->shm_lp_psind < 1 is
1578 * handled above.
1579 */
1580 if (
1581 #if VM_NRESERVLEVEL > 0
1582 shmfd->shm_lp_psind > VM_NRESERVLEVEL
1583 #else
1584 shmfd->shm_lp_psind > 1
1585 #endif
1586 )
1587 return (EINVAL);
1588 align = pagesizes[shmfd->shm_lp_psind];
1589 } else {
1590 align >>= MAP_ALIGNMENT_SHIFT;
1591 align = 1ULL << align;
1592 /* Also handles overflow. */
1593 if (align < pagesizes[shmfd->shm_lp_psind])
1594 return (EINVAL);
1595 }
1596
1597 vm_map_lock(map);
1598 if ((flags & MAP_FIXED) == 0) {
1599 try = 1;
1600 if (curmap && (*addr == 0 ||
1601 (*addr >= round_page((vm_offset_t)vms->vm_taddr) &&
1602 *addr < round_page((vm_offset_t)vms->vm_daddr +
1603 lim_max(td, RLIMIT_DATA))))) {
1604 *addr = roundup2((vm_offset_t)vms->vm_daddr +
1605 lim_max(td, RLIMIT_DATA),
1606 pagesizes[shmfd->shm_lp_psind]);
1607 }
1608 again:
1609 rv = vm_map_find_aligned(map, addr, size, maxaddr, align);
1610 if (rv != KERN_SUCCESS) {
1611 if (try == 1) {
1612 try = 2;
1613 *addr = vm_map_min(map);
1614 if ((*addr & mask) != 0)
1615 *addr = (*addr + mask) & mask;
1616 goto again;
1617 }
1618 goto fail1;
1619 }
1620 } else if ((flags & MAP_EXCL) == 0) {
1621 rv = vm_map_delete(map, *addr, *addr + size);
1622 if (rv != KERN_SUCCESS)
1623 goto fail1;
1624 } else {
1625 error = ENOSPC;
1626 if (vm_map_lookup_entry(map, *addr, &prev_entry))
1627 goto fail;
1628 next_entry = vm_map_entry_succ(prev_entry);
1629 if (next_entry->start < *addr + size)
1630 goto fail;
1631 }
1632
1633 rv = vm_map_insert(map, shmfd->shm_object, foff, *addr, *addr + size,
1634 prot, max_prot, docow);
1635 fail1:
1636 error = vm_mmap_to_errno(rv);
1637 fail:
1638 vm_map_unlock(map);
1639 return (error);
1640 }
1641
1642 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 max_maxprot,int flags,vm_ooffset_t foff,struct thread * td)1643 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
1644 vm_prot_t prot, vm_prot_t max_maxprot, int flags,
1645 vm_ooffset_t foff, struct thread *td)
1646 {
1647 struct shmfd *shmfd;
1648 vm_prot_t maxprot;
1649 int error;
1650 bool writecnt;
1651 void *rl_cookie;
1652
1653 shmfd = fp->f_data;
1654 maxprot = VM_PROT_NONE;
1655
1656 rl_cookie = shm_rangelock_rlock(shmfd, 0, objsize);
1657 /* FREAD should always be set. */
1658 if ((fp->f_flag & FREAD) != 0)
1659 maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
1660
1661 /*
1662 * If FWRITE's set, we can allow VM_PROT_WRITE unless it's a shared
1663 * mapping with a write seal applied. Private mappings are always
1664 * writeable.
1665 */
1666 if ((flags & MAP_SHARED) == 0) {
1667 if ((max_maxprot & VM_PROT_WRITE) != 0)
1668 maxprot |= VM_PROT_WRITE;
1669 writecnt = false;
1670 } else {
1671 if ((fp->f_flag & FWRITE) != 0 &&
1672 (shmfd->shm_seals & F_SEAL_WRITE) == 0)
1673 maxprot |= VM_PROT_WRITE;
1674
1675 /*
1676 * Any mappings from a writable descriptor may be upgraded to
1677 * VM_PROT_WRITE with mprotect(2), unless a write-seal was
1678 * applied between the open and subsequent mmap(2). We want to
1679 * reject application of a write seal as long as any such
1680 * mapping exists so that the seal cannot be trivially bypassed.
1681 */
1682 writecnt = (maxprot & VM_PROT_WRITE) != 0;
1683 if (!writecnt && (prot & VM_PROT_WRITE) != 0) {
1684 error = EACCES;
1685 goto out;
1686 }
1687 }
1688 maxprot &= max_maxprot;
1689
1690 /* See comment in vn_mmap(). */
1691 if (
1692 #ifdef _LP64
1693 objsize > OFF_MAX ||
1694 #endif
1695 foff > OFF_MAX - objsize) {
1696 error = EINVAL;
1697 goto out;
1698 }
1699
1700 #ifdef MAC
1701 error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
1702 if (error != 0)
1703 goto out;
1704 #endif
1705
1706 mtx_lock(&shm_timestamp_lock);
1707 vfs_timestamp(&shmfd->shm_atime);
1708 mtx_unlock(&shm_timestamp_lock);
1709 vm_object_reference(shmfd->shm_object);
1710
1711 if (shm_largepage(shmfd)) {
1712 writecnt = false;
1713 error = shm_mmap_large(shmfd, map, addr, objsize, prot,
1714 maxprot, flags, foff, td);
1715 } else {
1716 if (writecnt) {
1717 vm_pager_update_writecount(shmfd->shm_object, 0,
1718 objsize);
1719 }
1720 error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
1721 shmfd->shm_object, foff, writecnt, td);
1722 }
1723 if (error != 0) {
1724 if (writecnt)
1725 vm_pager_release_writecount(shmfd->shm_object, 0,
1726 objsize);
1727 vm_object_deallocate(shmfd->shm_object);
1728 }
1729 out:
1730 shm_rangelock_unlock(shmfd, rl_cookie);
1731 return (error);
1732 }
1733
1734 static int
shm_chmod(struct file * fp,mode_t mode,struct ucred * active_cred,struct thread * td)1735 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
1736 struct thread *td)
1737 {
1738 struct shmfd *shmfd;
1739 int error;
1740
1741 error = 0;
1742 shmfd = fp->f_data;
1743 mtx_lock(&shm_timestamp_lock);
1744 /*
1745 * SUSv4 says that x bits of permission need not be affected.
1746 * Be consistent with our shm_open there.
1747 */
1748 #ifdef MAC
1749 error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
1750 if (error != 0)
1751 goto out;
1752 #endif
1753 error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
1754 VADMIN, active_cred);
1755 if (error != 0)
1756 goto out;
1757 shmfd->shm_mode = mode & ACCESSPERMS;
1758 out:
1759 mtx_unlock(&shm_timestamp_lock);
1760 return (error);
1761 }
1762
1763 static int
shm_chown(struct file * fp,uid_t uid,gid_t gid,struct ucred * active_cred,struct thread * td)1764 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1765 struct thread *td)
1766 {
1767 struct shmfd *shmfd;
1768 int error;
1769
1770 error = 0;
1771 shmfd = fp->f_data;
1772 mtx_lock(&shm_timestamp_lock);
1773 #ifdef MAC
1774 error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
1775 if (error != 0)
1776 goto out;
1777 #endif
1778 if (uid == (uid_t)-1)
1779 uid = shmfd->shm_uid;
1780 if (gid == (gid_t)-1)
1781 gid = shmfd->shm_gid;
1782 if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
1783 (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
1784 (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN)))
1785 goto out;
1786 shmfd->shm_uid = uid;
1787 shmfd->shm_gid = gid;
1788 out:
1789 mtx_unlock(&shm_timestamp_lock);
1790 return (error);
1791 }
1792
1793 /*
1794 * Helper routines to allow the backing object of a shared memory file
1795 * descriptor to be mapped in the kernel.
1796 */
1797 int
shm_map(struct file * fp,size_t size,off_t offset,void ** memp)1798 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
1799 {
1800 struct shmfd *shmfd;
1801 vm_offset_t kva, ofs;
1802 vm_object_t obj;
1803 int rv;
1804
1805 if (fp->f_type != DTYPE_SHM)
1806 return (EINVAL);
1807 shmfd = fp->f_data;
1808 obj = shmfd->shm_object;
1809 VM_OBJECT_WLOCK(obj);
1810 /*
1811 * XXXRW: This validation is probably insufficient, and subject to
1812 * sign errors. It should be fixed.
1813 */
1814 if (offset >= shmfd->shm_size ||
1815 offset + size > round_page(shmfd->shm_size)) {
1816 VM_OBJECT_WUNLOCK(obj);
1817 return (EINVAL);
1818 }
1819
1820 shmfd->shm_kmappings++;
1821 vm_object_reference_locked(obj);
1822 VM_OBJECT_WUNLOCK(obj);
1823
1824 /* Map the object into the kernel_map and wire it. */
1825 kva = vm_map_min(kernel_map);
1826 ofs = offset & PAGE_MASK;
1827 offset = trunc_page(offset);
1828 size = round_page(size + ofs);
1829 rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1830 VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1831 VM_PROT_READ | VM_PROT_WRITE, 0);
1832 if (rv == KERN_SUCCESS) {
1833 rv = vm_map_wire(kernel_map, kva, kva + size,
1834 VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1835 if (rv == KERN_SUCCESS) {
1836 *memp = (void *)(kva + ofs);
1837 return (0);
1838 }
1839 vm_map_remove(kernel_map, kva, kva + size);
1840 } else
1841 vm_object_deallocate(obj);
1842
1843 /* On failure, drop our mapping reference. */
1844 VM_OBJECT_WLOCK(obj);
1845 shmfd->shm_kmappings--;
1846 VM_OBJECT_WUNLOCK(obj);
1847
1848 return (vm_mmap_to_errno(rv));
1849 }
1850
1851 /*
1852 * We require the caller to unmap the entire entry. This allows us to
1853 * safely decrement shm_kmappings when a mapping is removed.
1854 */
1855 int
shm_unmap(struct file * fp,void * mem,size_t size)1856 shm_unmap(struct file *fp, void *mem, size_t size)
1857 {
1858 struct shmfd *shmfd;
1859 vm_map_entry_t entry;
1860 vm_offset_t kva, ofs;
1861 vm_object_t obj;
1862 vm_pindex_t pindex;
1863 vm_prot_t prot;
1864 boolean_t wired;
1865 vm_map_t map;
1866 int rv;
1867
1868 if (fp->f_type != DTYPE_SHM)
1869 return (EINVAL);
1870 shmfd = fp->f_data;
1871 kva = (vm_offset_t)mem;
1872 ofs = kva & PAGE_MASK;
1873 kva = trunc_page(kva);
1874 size = round_page(size + ofs);
1875 map = kernel_map;
1876 rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1877 &obj, &pindex, &prot, &wired);
1878 if (rv != KERN_SUCCESS)
1879 return (EINVAL);
1880 if (entry->start != kva || entry->end != kva + size) {
1881 vm_map_lookup_done(map, entry);
1882 return (EINVAL);
1883 }
1884 vm_map_lookup_done(map, entry);
1885 if (obj != shmfd->shm_object)
1886 return (EINVAL);
1887 vm_map_remove(map, kva, kva + size);
1888 VM_OBJECT_WLOCK(obj);
1889 KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1890 shmfd->shm_kmappings--;
1891 VM_OBJECT_WUNLOCK(obj);
1892 return (0);
1893 }
1894
1895 static int
shm_fill_kinfo_locked(struct shmfd * shmfd,struct kinfo_file * kif,bool list)1896 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list)
1897 {
1898 const char *path, *pr_path;
1899 size_t pr_pathlen;
1900 bool visible;
1901
1902 sx_assert(&shm_dict_lock, SA_LOCKED);
1903 kif->kf_type = KF_TYPE_SHM;
1904 kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;
1905 kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1906 if (shmfd->shm_path != NULL) {
1907 path = shmfd->shm_path;
1908 pr_path = curthread->td_ucred->cr_prison->pr_path;
1909 if (strcmp(pr_path, "/") != 0) {
1910 /* Return the jail-rooted pathname. */
1911 pr_pathlen = strlen(pr_path);
1912 visible = strncmp(path, pr_path, pr_pathlen) == 0 &&
1913 path[pr_pathlen] == '/';
1914 if (list && !visible)
1915 return (EPERM);
1916 if (visible)
1917 path += pr_pathlen;
1918 }
1919 strlcpy(kif->kf_path, path, sizeof(kif->kf_path));
1920 }
1921 return (0);
1922 }
1923
1924 static int
shm_fill_kinfo(struct file * fp,struct kinfo_file * kif,struct filedesc * fdp __unused)1925 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif,
1926 struct filedesc *fdp __unused)
1927 {
1928 int res;
1929
1930 sx_slock(&shm_dict_lock);
1931 res = shm_fill_kinfo_locked(fp->f_data, kif, false);
1932 sx_sunlock(&shm_dict_lock);
1933 return (res);
1934 }
1935
1936 static int
shm_add_seals(struct file * fp,int seals)1937 shm_add_seals(struct file *fp, int seals)
1938 {
1939 struct shmfd *shmfd;
1940 void *rl_cookie;
1941 vm_ooffset_t writemappings;
1942 int error, nseals;
1943
1944 error = 0;
1945 shmfd = fp->f_data;
1946 rl_cookie = shm_rangelock_wlock(shmfd, 0, OFF_MAX);
1947
1948 /* Even already-set seals should result in EPERM. */
1949 if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) {
1950 error = EPERM;
1951 goto out;
1952 }
1953 nseals = seals & ~shmfd->shm_seals;
1954 if ((nseals & F_SEAL_WRITE) != 0) {
1955 if (shm_largepage(shmfd)) {
1956 error = ENOTSUP;
1957 goto out;
1958 }
1959
1960 /*
1961 * The rangelock above prevents writable mappings from being
1962 * added after we've started applying seals. The RLOCK here
1963 * is to avoid torn reads on ILP32 arches as unmapping/reducing
1964 * writemappings will be done without a rangelock.
1965 */
1966 VM_OBJECT_RLOCK(shmfd->shm_object);
1967 writemappings = shmfd->shm_object->un_pager.swp.writemappings;
1968 VM_OBJECT_RUNLOCK(shmfd->shm_object);
1969 /* kmappings are also writable */
1970 if (writemappings > 0) {
1971 error = EBUSY;
1972 goto out;
1973 }
1974 }
1975 shmfd->shm_seals |= nseals;
1976 out:
1977 shm_rangelock_unlock(shmfd, rl_cookie);
1978 return (error);
1979 }
1980
1981 static int
shm_get_seals(struct file * fp,int * seals)1982 shm_get_seals(struct file *fp, int *seals)
1983 {
1984 struct shmfd *shmfd;
1985
1986 shmfd = fp->f_data;
1987 *seals = shmfd->shm_seals;
1988 return (0);
1989 }
1990
1991 static int
shm_deallocate(struct shmfd * shmfd,off_t * offset,off_t * length,int flags)1992 shm_deallocate(struct shmfd *shmfd, off_t *offset, off_t *length, int flags)
1993 {
1994 vm_object_t object;
1995 vm_pindex_t pistart, pi, piend;
1996 vm_ooffset_t off, len;
1997 int startofs, endofs, end;
1998 int error;
1999
2000 off = *offset;
2001 len = *length;
2002 KASSERT(off + len <= (vm_ooffset_t)OFF_MAX, ("off + len overflows"));
2003 if (off + len > shmfd->shm_size)
2004 len = shmfd->shm_size - off;
2005 object = shmfd->shm_object;
2006 startofs = off & PAGE_MASK;
2007 endofs = (off + len) & PAGE_MASK;
2008 pistart = OFF_TO_IDX(off);
2009 piend = OFF_TO_IDX(off + len);
2010 pi = OFF_TO_IDX(off + PAGE_MASK);
2011 error = 0;
2012
2013 /* Handle the case when offset is on or beyond shm size. */
2014 if ((off_t)len <= 0) {
2015 *length = 0;
2016 return (0);
2017 }
2018
2019 VM_OBJECT_WLOCK(object);
2020
2021 if (startofs != 0) {
2022 end = pistart != piend ? PAGE_SIZE : endofs;
2023 error = shm_partial_page_invalidate(object, pistart, startofs,
2024 end);
2025 if (error)
2026 goto out;
2027 off += end - startofs;
2028 len -= end - startofs;
2029 }
2030
2031 if (pi < piend) {
2032 vm_object_page_remove(object, pi, piend, 0);
2033 off += IDX_TO_OFF(piend - pi);
2034 len -= IDX_TO_OFF(piend - pi);
2035 }
2036
2037 if (endofs != 0 && pistart != piend) {
2038 error = shm_partial_page_invalidate(object, piend, 0, endofs);
2039 if (error)
2040 goto out;
2041 off += endofs;
2042 len -= endofs;
2043 }
2044
2045 out:
2046 VM_OBJECT_WUNLOCK(shmfd->shm_object);
2047 *offset = off;
2048 *length = len;
2049 return (error);
2050 }
2051
2052 static int
shm_fspacectl(struct file * fp,int cmd,off_t * offset,off_t * length,int flags,struct ucred * active_cred,struct thread * td)2053 shm_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags,
2054 struct ucred *active_cred, struct thread *td)
2055 {
2056 void *rl_cookie;
2057 struct shmfd *shmfd;
2058 off_t off, len;
2059 int error;
2060
2061 KASSERT(cmd == SPACECTL_DEALLOC, ("shm_fspacectl: Invalid cmd"));
2062 KASSERT((flags & ~SPACECTL_F_SUPPORTED) == 0,
2063 ("shm_fspacectl: non-zero flags"));
2064 KASSERT(*offset >= 0 && *length > 0 && *length <= OFF_MAX - *offset,
2065 ("shm_fspacectl: offset/length overflow or underflow"));
2066 error = EINVAL;
2067 shmfd = fp->f_data;
2068 off = *offset;
2069 len = *length;
2070
2071 rl_cookie = shm_rangelock_wlock(shmfd, off, off + len);
2072 switch (cmd) {
2073 case SPACECTL_DEALLOC:
2074 if ((shmfd->shm_seals & F_SEAL_WRITE) != 0) {
2075 error = EPERM;
2076 break;
2077 }
2078 error = shm_deallocate(shmfd, &off, &len, flags);
2079 *offset = off;
2080 *length = len;
2081 break;
2082 default:
2083 __assert_unreachable();
2084 }
2085 shm_rangelock_unlock(shmfd, rl_cookie);
2086 return (error);
2087 }
2088
2089
2090 static int
shm_fallocate(struct file * fp,off_t offset,off_t len,struct thread * td)2091 shm_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td)
2092 {
2093 void *rl_cookie;
2094 struct shmfd *shmfd;
2095 size_t size;
2096 int error;
2097
2098 /* This assumes that the caller already checked for overflow. */
2099 error = 0;
2100 shmfd = fp->f_data;
2101 size = offset + len;
2102
2103 /*
2104 * Just grab the rangelock for the range that we may be attempting to
2105 * grow, rather than blocking read/write for regions we won't be
2106 * touching while this (potential) resize is in progress. Other
2107 * attempts to resize the shmfd will have to take a write lock from 0 to
2108 * OFF_MAX, so this being potentially beyond the current usable range of
2109 * the shmfd is not necessarily a concern. If other mechanisms are
2110 * added to grow a shmfd, this may need to be re-evaluated.
2111 */
2112 rl_cookie = shm_rangelock_wlock(shmfd, offset, size);
2113 if (size > shmfd->shm_size)
2114 error = shm_dotruncate_cookie(shmfd, size, rl_cookie);
2115 shm_rangelock_unlock(shmfd, rl_cookie);
2116 /* Translate to posix_fallocate(2) return value as needed. */
2117 if (error == ENOMEM)
2118 error = ENOSPC;
2119 return (error);
2120 }
2121
2122 static int
sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS)2123 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS)
2124 {
2125 struct shm_mapping *shmm;
2126 struct sbuf sb;
2127 struct kinfo_file kif;
2128 u_long i;
2129 int error, error2;
2130
2131 sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req);
2132 sbuf_clear_flags(&sb, SBUF_INCLUDENUL);
2133 error = 0;
2134 sx_slock(&shm_dict_lock);
2135 for (i = 0; i < shm_hash + 1; i++) {
2136 LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) {
2137 error = shm_fill_kinfo_locked(shmm->sm_shmfd,
2138 &kif, true);
2139 if (error == EPERM) {
2140 error = 0;
2141 continue;
2142 }
2143 if (error != 0)
2144 break;
2145 pack_kinfo(&kif);
2146 error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ?
2147 0 : ENOMEM;
2148 if (error != 0)
2149 break;
2150 }
2151 }
2152 sx_sunlock(&shm_dict_lock);
2153 error2 = sbuf_finish(&sb);
2154 sbuf_delete(&sb);
2155 return (error != 0 ? error : error2);
2156 }
2157
2158 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list,
2159 CTLFLAG_RD | CTLFLAG_PRISON | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE,
2160 NULL, 0, sysctl_posix_shm_list, "",
2161 "POSIX SHM list");
2162
2163 int
kern_shm_open(struct thread * td,const char * path,int flags,mode_t mode,struct filecaps * caps)2164 kern_shm_open(struct thread *td, const char *path, int flags, mode_t mode,
2165 struct filecaps *caps)
2166 {
2167
2168 return (kern_shm_open2(td, path, flags, mode, 0, caps, NULL));
2169 }
2170
2171 /*
2172 * This version of the shm_open() interface leaves CLOEXEC behavior up to the
2173 * caller, and libc will enforce it for the traditional shm_open() call. This
2174 * allows other consumers, like memfd_create(), to opt-in for CLOEXEC. This
2175 * interface also includes a 'name' argument that is currently unused, but could
2176 * potentially be exported later via some interface for debugging purposes.
2177 * From the kernel's perspective, it is optional. Individual consumers like
2178 * memfd_create() may require it in order to be compatible with other systems
2179 * implementing the same function.
2180 */
2181 int
sys_shm_open2(struct thread * td,struct shm_open2_args * uap)2182 sys_shm_open2(struct thread *td, struct shm_open2_args *uap)
2183 {
2184
2185 return (kern_shm_open2(td, uap->path, uap->flags, uap->mode,
2186 uap->shmflags, NULL, uap->name));
2187 }
2188
2189 int
shm_get_path(struct vm_object * obj,char * path,size_t sz)2190 shm_get_path(struct vm_object *obj, char *path, size_t sz)
2191 {
2192 struct shmfd *shmfd;
2193 int error;
2194
2195 error = 0;
2196 shmfd = NULL;
2197 sx_slock(&shm_dict_lock);
2198 VM_OBJECT_RLOCK(obj);
2199 if ((obj->flags & OBJ_POSIXSHM) == 0) {
2200 error = EINVAL;
2201 } else {
2202 if (obj->type == shmfd_pager_type)
2203 shmfd = obj->un_pager.swp.swp_priv;
2204 else if (obj->type == OBJT_PHYS)
2205 shmfd = obj->un_pager.phys.phys_priv;
2206 if (shmfd == NULL) {
2207 error = ENXIO;
2208 } else {
2209 strlcpy(path, shmfd->shm_path == NULL ? "anon" :
2210 shmfd->shm_path, sz);
2211 }
2212 }
2213 if (error != 0)
2214 path[0] = '\0';
2215 VM_OBJECT_RUNLOCK(obj);
2216 sx_sunlock(&shm_dict_lock);
2217 return (error);
2218 }
2219