2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2006, 2011, 2016-2017 Robert N. M. Watson
7 * Portions of this software were developed by BAE Systems, the University of
8 * Cambridge Computer Laboratory, and Memorial University under DARPA/AFRL
9 * contract FA8650-15-C-7558 ("CADETS"), as part of the DARPA Transparent
10 * Computing (TC) research program.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
21 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * Support for shared swap-backed anonymous memory objects via
36 * shm_open(2), shm_rename(2), and shm_unlink(2).
37 * While most of the implementation is here, vm_mmap.c contains
38 * mapping logic changes.
40 * posixshmcontrol(1) allows users to inspect the state of the memory
41 * objects. Per-uid swap resource limit controls total amount of
42 * memory that user can consume for anonymous objects, including
46 #include <sys/cdefs.h>
47 __FBSDID("$FreeBSD$");
49 #include "opt_capsicum.h"
50 #include "opt_ktrace.h"
52 #include <sys/param.h>
53 #include <sys/capsicum.h>
55 #include <sys/fcntl.h>
57 #include <sys/filedesc.h>
58 #include <sys/filio.h>
59 #include <sys/fnv_hash.h>
60 #include <sys/kernel.h>
61 #include <sys/limits.h>
63 #include <sys/signal.h>
65 #include <sys/ktrace.h>
67 #include <sys/malloc.h>
69 #include <sys/mutex.h>
72 #include <sys/refcount.h>
73 #include <sys/resourcevar.h>
74 #include <sys/rwlock.h>
77 #include <sys/syscallsubr.h>
78 #include <sys/sysctl.h>
79 #include <sys/sysproto.h>
80 #include <sys/systm.h>
83 #include <sys/vnode.h>
84 #include <sys/unistd.h>
87 #include <security/audit/audit.h>
88 #include <security/mac/mac_framework.h>
91 #include <vm/vm_param.h>
93 #include <vm/vm_extern.h>
94 #include <vm/vm_map.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_pageout.h>
99 #include <vm/vm_pager.h>
100 #include <vm/swap_pager.h>
105 struct shmfd *sm_shmfd;
106 LIST_ENTRY(shm_mapping) sm_link;
109 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor");
110 static LIST_HEAD(, shm_mapping) *shm_dictionary;
111 static struct sx shm_dict_lock;
112 static struct mtx shm_timestamp_lock;
113 static u_long shm_hash;
114 static struct unrhdr64 shm_ino_unr;
115 static dev_t shm_dev_ino;
117 #define SHM_HASH(fnv) (&shm_dictionary[(fnv) & shm_hash])
119 static void shm_init(void *arg);
120 static void shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd);
121 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv);
122 static int shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred);
123 static int shm_dotruncate_locked(struct shmfd *shmfd, off_t length,
126 static fo_rdwr_t shm_read;
127 static fo_rdwr_t shm_write;
128 static fo_truncate_t shm_truncate;
129 static fo_ioctl_t shm_ioctl;
130 static fo_stat_t shm_stat;
131 static fo_close_t shm_close;
132 static fo_chmod_t shm_chmod;
133 static fo_chown_t shm_chown;
134 static fo_seek_t shm_seek;
135 static fo_fill_kinfo_t shm_fill_kinfo;
136 static fo_mmap_t shm_mmap;
137 static fo_get_seals_t shm_get_seals;
138 static fo_add_seals_t shm_add_seals;
140 /* File descriptor operations. */
141 struct fileops shm_ops = {
143 .fo_write = shm_write,
144 .fo_truncate = shm_truncate,
145 .fo_ioctl = shm_ioctl,
146 .fo_poll = invfo_poll,
147 .fo_kqfilter = invfo_kqfilter,
149 .fo_close = shm_close,
150 .fo_chmod = shm_chmod,
151 .fo_chown = shm_chown,
152 .fo_sendfile = vn_sendfile,
154 .fo_fill_kinfo = shm_fill_kinfo,
156 .fo_get_seals = shm_get_seals,
157 .fo_add_seals = shm_add_seals,
158 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
161 FEATURE(posix_shm, "POSIX shared memory");
164 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio)
169 int error, offset, rv;
171 idx = OFF_TO_IDX(uio->uio_offset);
172 offset = uio->uio_offset & PAGE_MASK;
173 tlen = MIN(PAGE_SIZE - offset, len);
175 VM_OBJECT_WLOCK(obj);
178 * Read I/O without either a corresponding resident page or swap
179 * page: use zero_region. This is intended to avoid instantiating
180 * pages on read from a sparse region.
182 if (uio->uio_rw == UIO_READ && vm_page_lookup(obj, idx) == NULL &&
183 !vm_pager_has_page(obj, idx, NULL, NULL)) {
184 VM_OBJECT_WUNLOCK(obj);
185 return (uiomove(__DECONST(void *, zero_region), tlen, uio));
189 * Parallel reads of the page content from disk are prevented
192 * Although the tmpfs vnode lock is held here, it is
193 * nonetheless safe to sleep waiting for a free page. The
194 * pageout daemon does not need to acquire the tmpfs vnode
195 * lock to page out tobj's pages because tobj is a OBJT_SWAP
198 rv = vm_page_grab_valid(&m, obj, idx,
199 VM_ALLOC_NORMAL | VM_ALLOC_WIRED | VM_ALLOC_NOBUSY);
200 if (rv != VM_PAGER_OK) {
201 VM_OBJECT_WUNLOCK(obj);
202 printf("uiomove_object: vm_obj %p idx %jd pager error %d\n",
206 VM_OBJECT_WUNLOCK(obj);
207 error = uiomove_fromphys(&m, offset, tlen, uio);
208 if (uio->uio_rw == UIO_WRITE && error == 0) {
209 VM_OBJECT_WLOCK(obj);
211 vm_pager_page_unswapped(m);
212 VM_OBJECT_WUNLOCK(obj);
214 vm_page_unwire(m, PQ_ACTIVE);
220 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
227 while ((resid = uio->uio_resid) > 0) {
228 if (obj_size <= uio->uio_offset)
230 len = MIN(obj_size - uio->uio_offset, resid);
233 error = uiomove_object_page(obj, len, uio);
234 if (error != 0 || resid == uio->uio_resid)
241 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
248 foffset = foffset_lock(fp, 0);
253 (offset > 0 && foffset > OFF_MAX - offset)) {
260 if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
264 offset += shmfd->shm_size;
272 if (offset < 0 || offset > shmfd->shm_size)
275 td->td_uretoff.tdu_off = offset;
277 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
282 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
283 int flags, struct thread *td)
291 error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
295 foffset_lock_uio(fp, uio, flags);
296 rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset,
297 uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
298 error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
299 rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
300 foffset_unlock_uio(fp, uio, flags);
305 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
306 int flags, struct thread *td)
314 error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
318 foffset_lock_uio(fp, uio, flags);
319 if ((flags & FOF_OFFSET) == 0) {
320 rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
323 rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset,
324 uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
326 if ((shmfd->shm_seals & F_SEAL_WRITE) != 0)
329 error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
330 rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
331 foffset_unlock_uio(fp, uio, flags);
336 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
346 error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
350 return (shm_dotruncate(shmfd, length));
354 shm_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred,
362 * Allow fcntl(fd, F_SETFL, O_NONBLOCK) to work,
363 * just like it would on an unlinked regular file
372 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
383 error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
389 * Attempt to return sanish values for fstat() on a memory file
392 bzero(sb, sizeof(*sb));
393 sb->st_blksize = PAGE_SIZE;
394 sb->st_size = shmfd->shm_size;
395 sb->st_blocks = howmany(sb->st_size, sb->st_blksize);
396 mtx_lock(&shm_timestamp_lock);
397 sb->st_atim = shmfd->shm_atime;
398 sb->st_ctim = shmfd->shm_ctime;
399 sb->st_mtim = shmfd->shm_mtime;
400 sb->st_birthtim = shmfd->shm_birthtime;
401 sb->st_mode = S_IFREG | shmfd->shm_mode; /* XXX */
402 sb->st_uid = shmfd->shm_uid;
403 sb->st_gid = shmfd->shm_gid;
404 mtx_unlock(&shm_timestamp_lock);
405 sb->st_dev = shm_dev_ino;
406 sb->st_ino = shmfd->shm_ino;
407 sb->st_nlink = shmfd->shm_object->ref_count;
413 shm_close(struct file *fp, struct thread *td)
425 shm_dotruncate_locked(struct shmfd *shmfd, off_t length, void *rl_cookie)
429 vm_pindex_t idx, nobjsize;
433 KASSERT(length >= 0, ("shm_dotruncate: length < 0"));
434 object = shmfd->shm_object;
435 VM_OBJECT_ASSERT_WLOCKED(object);
436 rangelock_cookie_assert(rl_cookie, RA_WLOCKED);
437 if (length == shmfd->shm_size)
439 nobjsize = OFF_TO_IDX(length + PAGE_MASK);
441 /* Are we shrinking? If so, trim the end. */
442 if (length < shmfd->shm_size) {
443 if ((shmfd->shm_seals & F_SEAL_SHRINK) != 0)
447 * Disallow any requests to shrink the size if this
448 * object is mapped into the kernel.
450 if (shmfd->shm_kmappings > 0)
454 * Zero the truncated part of the last page.
456 base = length & PAGE_MASK;
458 idx = OFF_TO_IDX(length);
460 m = vm_page_lookup(object, idx);
462 if (vm_page_sleep_if_busy(m, "shmtrc"))
464 } else if (vm_pager_has_page(object, idx, NULL, NULL)) {
465 m = vm_page_alloc(object, idx,
466 VM_ALLOC_NORMAL | VM_ALLOC_WAITFAIL);
469 rv = vm_pager_get_pages(object, &m, 1, NULL,
471 if (rv == VM_PAGER_OK) {
473 * Since the page was not resident,
474 * and therefore not recently
475 * accessed, immediately enqueue it
476 * for asynchronous laundering. The
477 * current operation is not regarded
484 VM_OBJECT_WUNLOCK(object);
489 pmap_zero_page_area(m, base, PAGE_SIZE - base);
490 KASSERT(m->valid == VM_PAGE_BITS_ALL,
491 ("shm_dotruncate: page %p is invalid", m));
493 vm_pager_page_unswapped(m);
496 delta = IDX_TO_OFF(object->size - nobjsize);
498 /* Toss in memory pages. */
499 if (nobjsize < object->size)
500 vm_object_page_remove(object, nobjsize, object->size,
503 /* Toss pages from swap. */
504 if (object->type == OBJT_SWAP)
505 swap_pager_freespace(object, nobjsize, delta);
507 /* Free the swap accounted for shm */
508 swap_release_by_cred(delta, object->cred);
509 object->charge -= delta;
511 if ((shmfd->shm_seals & F_SEAL_GROW) != 0)
514 /* Try to reserve additional swap space. */
515 delta = IDX_TO_OFF(nobjsize - object->size);
516 if (!swap_reserve_by_cred(delta, object->cred))
518 object->charge += delta;
520 shmfd->shm_size = length;
521 mtx_lock(&shm_timestamp_lock);
522 vfs_timestamp(&shmfd->shm_ctime);
523 shmfd->shm_mtime = shmfd->shm_ctime;
524 mtx_unlock(&shm_timestamp_lock);
525 object->size = nobjsize;
530 shm_dotruncate(struct shmfd *shmfd, off_t length)
535 rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
537 VM_OBJECT_WLOCK(shmfd->shm_object);
538 error = shm_dotruncate_locked(shmfd, length, rl_cookie);
539 VM_OBJECT_WUNLOCK(shmfd->shm_object);
540 rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
545 * shmfd object management including creation and reference counting
549 shm_alloc(struct ucred *ucred, mode_t mode)
553 shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
555 shmfd->shm_uid = ucred->cr_uid;
556 shmfd->shm_gid = ucred->cr_gid;
557 shmfd->shm_mode = mode;
558 shmfd->shm_object = vm_pager_allocate(OBJT_SWAP, NULL,
559 shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
560 KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
561 shmfd->shm_object->pg_color = 0;
562 VM_OBJECT_WLOCK(shmfd->shm_object);
563 vm_object_clear_flag(shmfd->shm_object, OBJ_ONEMAPPING);
564 vm_object_set_flag(shmfd->shm_object, OBJ_COLORED | OBJ_NOSPLIT);
565 VM_OBJECT_WUNLOCK(shmfd->shm_object);
566 vfs_timestamp(&shmfd->shm_birthtime);
567 shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
568 shmfd->shm_birthtime;
569 shmfd->shm_ino = alloc_unr64(&shm_ino_unr);
570 refcount_init(&shmfd->shm_refs, 1);
571 mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
572 rangelock_init(&shmfd->shm_rl);
574 mac_posixshm_init(shmfd);
575 mac_posixshm_create(ucred, shmfd);
582 shm_hold(struct shmfd *shmfd)
585 refcount_acquire(&shmfd->shm_refs);
590 shm_drop(struct shmfd *shmfd)
593 if (refcount_release(&shmfd->shm_refs)) {
595 mac_posixshm_destroy(shmfd);
597 rangelock_destroy(&shmfd->shm_rl);
598 mtx_destroy(&shmfd->shm_mtx);
599 vm_object_deallocate(shmfd->shm_object);
600 free(shmfd, M_SHMFD);
605 * Determine if the credentials have sufficient permissions for a
606 * specified combination of FREAD and FWRITE.
609 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
619 mtx_lock(&shm_timestamp_lock);
620 error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
621 accmode, ucred, NULL);
622 mtx_unlock(&shm_timestamp_lock);
627 * Dictionary management. We maintain an in-kernel dictionary to map
628 * paths to shmfd objects. We use the FNV hash on the path to store
629 * the mappings in a hash table.
635 mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
636 sx_init(&shm_dict_lock, "shm dictionary");
637 shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
638 new_unrhdr64(&shm_ino_unr, 1);
639 shm_dev_ino = devfs_alloc_cdp_inode();
640 KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
642 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
644 static struct shmfd *
645 shm_lookup(char *path, Fnv32_t fnv)
647 struct shm_mapping *map;
649 LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
650 if (map->sm_fnv != fnv)
652 if (strcmp(map->sm_path, path) == 0)
653 return (map->sm_shmfd);
660 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
662 struct shm_mapping *map;
664 map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
667 map->sm_shmfd = shm_hold(shmfd);
668 shmfd->shm_path = path;
669 LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
673 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
675 struct shm_mapping *map;
678 LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
679 if (map->sm_fnv != fnv)
681 if (strcmp(map->sm_path, path) == 0) {
683 error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
687 error = shm_access(map->sm_shmfd, ucred,
691 map->sm_shmfd->shm_path = NULL;
692 LIST_REMOVE(map, sm_link);
693 shm_drop(map->sm_shmfd);
694 free(map->sm_path, M_SHMFD);
704 kern_shm_open(struct thread *td, const char *userpath, int flags, mode_t mode,
705 struct filecaps *fcaps, int initial_seals)
707 struct filedesc *fdp;
718 #ifdef CAPABILITY_MODE
720 * shm_open(2) is only allowed for anonymous objects.
722 if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON))
726 AUDIT_ARG_FFLAGS(flags);
727 AUDIT_ARG_MODE(mode);
729 if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
732 if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
736 * Currently only F_SEAL_SEAL may be set when creating or opening shmfd.
737 * If the decision is made later to allow additional seals, care must be
738 * taken below to ensure that the seals are properly set if the shmfd
739 * already existed -- this currently assumes that only F_SEAL_SEAL can
740 * be set and doesn't take further precautions to ensure the validity of
741 * the seals being added with respect to current mappings.
743 if ((initial_seals & ~F_SEAL_SEAL) != 0)
746 fdp = td->td_proc->p_fd;
747 cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS;
750 * shm_open(2) created shm should always have O_CLOEXEC set, as mandated
751 * by POSIX. We allow it to be unset here so that an in-kernel
752 * interface may be written as a thin layer around shm, optionally not
753 * setting CLOEXEC. For shm_open(2), O_CLOEXEC is set unconditionally
754 * in sys_shm_open() to keep this implementation compliant.
756 error = falloc_caps(td, &fp, &fd, flags & O_CLOEXEC, fcaps);
760 /* A SHM_ANON path pointer creates an anonymous object. */
761 if (userpath == SHM_ANON) {
762 /* A read-only anonymous object is pointless. */
763 if ((flags & O_ACCMODE) == O_RDONLY) {
768 shmfd = shm_alloc(td->td_ucred, cmode);
769 shmfd->shm_seals = initial_seals;
771 path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
772 pr_path = td->td_ucred->cr_prison->pr_path;
774 /* Construct a full pathname for jailed callers. */
775 pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
776 : strlcpy(path, pr_path, MAXPATHLEN);
777 error = copyinstr(userpath, path + pr_pathlen,
778 MAXPATHLEN - pr_pathlen, NULL);
780 if (error == 0 && KTRPOINT(curthread, KTR_NAMEI))
783 /* Require paths to start with a '/' character. */
784 if (error == 0 && path[pr_pathlen] != '/')
793 AUDIT_ARG_UPATH1_CANON(path);
794 fnv = fnv_32_str(path, FNV1_32_INIT);
795 sx_xlock(&shm_dict_lock);
796 shmfd = shm_lookup(path, fnv);
798 /* Object does not yet exist, create it if requested. */
799 if (flags & O_CREAT) {
801 error = mac_posixshm_check_create(td->td_ucred,
805 shmfd = shm_alloc(td->td_ucred, cmode);
806 shmfd->shm_seals = initial_seals;
807 shm_insert(path, fnv, shmfd);
816 rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
820 * kern_shm_open() likely shouldn't ever error out on
821 * trying to set a seal that already exists, unlike
822 * F_ADD_SEALS. This would break terribly as
823 * shm_open(2) actually sets F_SEAL_SEAL to maintain
824 * historical behavior where the underlying file could
827 initial_seals &= ~shmfd->shm_seals;
830 * Object already exists, obtain a new
831 * reference if requested and permitted.
836 * initial_seals can't set additional seals if we've
837 * already been set F_SEAL_SEAL. If F_SEAL_SEAL is set,
838 * then we've already removed that one from
839 * initial_seals. This is currently redundant as we
840 * only allow setting F_SEAL_SEAL at creation time, but
841 * it's cheap to check and decreases the effort required
842 * to allow additional seals.
844 if ((shmfd->shm_seals & F_SEAL_SEAL) != 0 &&
847 else if ((flags & (O_CREAT | O_EXCL)) ==
852 error = mac_posixshm_check_open(td->td_ucred,
853 shmfd, FFLAGS(flags & O_ACCMODE));
856 error = shm_access(shmfd, td->td_ucred,
857 FFLAGS(flags & O_ACCMODE));
861 * Truncate the file back to zero length if
862 * O_TRUNC was specified and the object was
863 * opened with read/write.
866 (flags & (O_ACCMODE | O_TRUNC)) ==
867 (O_RDWR | O_TRUNC)) {
868 VM_OBJECT_WLOCK(shmfd->shm_object);
870 error = mac_posixshm_check_truncate(
871 td->td_ucred, fp->f_cred, shmfd);
874 error = shm_dotruncate_locked(shmfd, 0,
876 VM_OBJECT_WUNLOCK(shmfd->shm_object);
880 * Currently we only allow F_SEAL_SEAL to be
881 * set initially. As noted above, this would
882 * need to be reworked should that change.
884 shmfd->shm_seals |= initial_seals;
887 rangelock_unlock(&shmfd->shm_rl, rl_cookie,
890 sx_xunlock(&shm_dict_lock);
899 finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
901 td->td_retval[0] = fd;
908 #ifdef COMPAT_FREEBSD12
910 freebsd12_shm_open(struct thread *td, struct freebsd12_shm_open_args *uap)
913 return (kern_shm_open(td, uap->path, uap->flags | O_CLOEXEC, uap->mode,
919 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
927 path = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
928 pr_path = td->td_ucred->cr_prison->pr_path;
929 pr_pathlen = strcmp(pr_path, "/") == 0 ? 0
930 : strlcpy(path, pr_path, MAXPATHLEN);
931 error = copyinstr(uap->path, path + pr_pathlen, MAXPATHLEN - pr_pathlen,
938 if (KTRPOINT(curthread, KTR_NAMEI))
941 AUDIT_ARG_UPATH1_CANON(path);
942 fnv = fnv_32_str(path, FNV1_32_INIT);
943 sx_xlock(&shm_dict_lock);
944 error = shm_remove(path, fnv, td->td_ucred);
945 sx_xunlock(&shm_dict_lock);
952 sys_shm_rename(struct thread *td, struct shm_rename_args *uap)
954 char *path_from = NULL, *path_to = NULL;
955 Fnv32_t fnv_from, fnv_to;
956 struct shmfd *fd_from;
964 * Make sure the user passed only valid flags.
965 * If you add a new flag, please add a new term here.
968 SHM_RENAME_NOREPLACE |
976 * EXCHANGE and NOREPLACE don't quite make sense together. Let's
977 * force the user to choose one or the other.
979 if ((flags & SHM_RENAME_NOREPLACE) != 0 &&
980 (flags & SHM_RENAME_EXCHANGE) != 0) {
986 * Malloc zone M_SHMFD, since this path may end up freed later from
987 * M_SHMFD if we end up doing an insert.
989 path_from = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
990 error = copyinstr(uap->path_from, path_from, MAXPATHLEN, NULL);
994 path_to = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
995 error = copyinstr(uap->path_to, path_to, MAXPATHLEN, NULL);
999 /* Rename with from/to equal is a no-op */
1000 if (strncmp(path_from, path_to, MAXPATHLEN) == 0)
1003 fnv_from = fnv_32_str(path_from, FNV1_32_INIT);
1004 fnv_to = fnv_32_str(path_to, FNV1_32_INIT);
1006 sx_xlock(&shm_dict_lock);
1008 fd_from = shm_lookup(path_from, fnv_from);
1009 if (fd_from == NULL) {
1010 sx_xunlock(&shm_dict_lock);
1015 fd_to = shm_lookup(path_to, fnv_to);
1016 if ((flags & SHM_RENAME_NOREPLACE) != 0 && fd_to != NULL) {
1017 sx_xunlock(&shm_dict_lock);
1023 * Unconditionally prevents shm_remove from invalidating the 'from'
1027 error = shm_remove(path_from, fnv_from, td->td_ucred);
1030 * One of my assumptions failed if ENOENT (e.g. locking didn't
1033 KASSERT(error != ENOENT, ("Our shm disappeared during shm_rename: %s",
1037 sx_xunlock(&shm_dict_lock);
1042 * If we are exchanging, we need to ensure the shm_remove below
1043 * doesn't invalidate the dest shm's state.
1045 if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL)
1049 * NOTE: if path_to is not already in the hash, c'est la vie;
1050 * it simply means we have nothing already at path_to to unlink.
1051 * That is the ENOENT case.
1053 * If we somehow don't have access to unlink this guy, but
1054 * did for the shm at path_from, then relink the shm to path_from
1055 * and abort with EACCES.
1057 * All other errors: that is weird; let's relink and abort the
1060 error = shm_remove(path_to, fnv_to, td->td_ucred);
1061 if (error && error != ENOENT) {
1062 shm_insert(path_from, fnv_from, fd_from);
1064 /* Don't free path_from now, since the hash references it */
1066 sx_xunlock(&shm_dict_lock);
1070 shm_insert(path_to, fnv_to, fd_from);
1072 /* Don't free path_to now, since the hash references it */
1075 /* We kept a ref when we removed, and incremented again in insert */
1078 KASSERT(fd_from->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1079 fd_from->shm_refs));
1082 if ((flags & SHM_RENAME_EXCHANGE) != 0 && fd_to != NULL) {
1083 shm_insert(path_from, fnv_from, fd_to);
1087 KASSERT(fd_to->shm_refs > 0, ("Expected >0 refs; got: %d\n",
1093 sx_xunlock(&shm_dict_lock);
1096 if (path_from != NULL)
1097 free(path_from, M_SHMFD);
1098 if (path_to != NULL)
1099 free(path_to, M_SHMFD);
1104 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
1105 vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
1106 vm_ooffset_t foff, struct thread *td)
1108 struct shmfd *shmfd;
1115 maxprot = VM_PROT_NONE;
1117 rl_cookie = rangelock_rlock(&shmfd->shm_rl, 0, objsize,
1119 /* FREAD should always be set. */
1120 if ((fp->f_flag & FREAD) != 0)
1121 maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
1122 if ((fp->f_flag & FWRITE) != 0)
1123 maxprot |= VM_PROT_WRITE;
1125 writecnt = (flags & MAP_SHARED) != 0 && (prot & VM_PROT_WRITE) != 0;
1127 if (writecnt && (shmfd->shm_seals & F_SEAL_WRITE) != 0) {
1132 /* Don't permit shared writable mappings on read-only descriptors. */
1133 if (writecnt && (maxprot & VM_PROT_WRITE) == 0) {
1137 maxprot &= cap_maxprot;
1139 /* See comment in vn_mmap(). */
1142 objsize > OFF_MAX ||
1144 foff < 0 || foff > OFF_MAX - objsize) {
1150 error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
1155 mtx_lock(&shm_timestamp_lock);
1156 vfs_timestamp(&shmfd->shm_atime);
1157 mtx_unlock(&shm_timestamp_lock);
1158 vm_object_reference(shmfd->shm_object);
1161 vm_pager_update_writecount(shmfd->shm_object, 0, objsize);
1162 error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
1163 shmfd->shm_object, foff, writecnt, td);
1166 vm_pager_release_writecount(shmfd->shm_object, 0,
1168 vm_object_deallocate(shmfd->shm_object);
1171 rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
1176 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
1179 struct shmfd *shmfd;
1184 mtx_lock(&shm_timestamp_lock);
1186 * SUSv4 says that x bits of permission need not be affected.
1187 * Be consistent with our shm_open there.
1190 error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
1194 error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid,
1195 shmfd->shm_gid, VADMIN, active_cred, NULL);
1198 shmfd->shm_mode = mode & ACCESSPERMS;
1200 mtx_unlock(&shm_timestamp_lock);
1205 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
1208 struct shmfd *shmfd;
1213 mtx_lock(&shm_timestamp_lock);
1215 error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
1219 if (uid == (uid_t)-1)
1220 uid = shmfd->shm_uid;
1221 if (gid == (gid_t)-1)
1222 gid = shmfd->shm_gid;
1223 if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
1224 (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
1225 (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN)))
1227 shmfd->shm_uid = uid;
1228 shmfd->shm_gid = gid;
1230 mtx_unlock(&shm_timestamp_lock);
1235 * Helper routines to allow the backing object of a shared memory file
1236 * descriptor to be mapped in the kernel.
1239 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
1241 struct shmfd *shmfd;
1242 vm_offset_t kva, ofs;
1246 if (fp->f_type != DTYPE_SHM)
1249 obj = shmfd->shm_object;
1250 VM_OBJECT_WLOCK(obj);
1252 * XXXRW: This validation is probably insufficient, and subject to
1253 * sign errors. It should be fixed.
1255 if (offset >= shmfd->shm_size ||
1256 offset + size > round_page(shmfd->shm_size)) {
1257 VM_OBJECT_WUNLOCK(obj);
1261 shmfd->shm_kmappings++;
1262 vm_object_reference_locked(obj);
1263 VM_OBJECT_WUNLOCK(obj);
1265 /* Map the object into the kernel_map and wire it. */
1266 kva = vm_map_min(kernel_map);
1267 ofs = offset & PAGE_MASK;
1268 offset = trunc_page(offset);
1269 size = round_page(size + ofs);
1270 rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
1271 VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
1272 VM_PROT_READ | VM_PROT_WRITE, 0);
1273 if (rv == KERN_SUCCESS) {
1274 rv = vm_map_wire(kernel_map, kva, kva + size,
1275 VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1276 if (rv == KERN_SUCCESS) {
1277 *memp = (void *)(kva + ofs);
1280 vm_map_remove(kernel_map, kva, kva + size);
1282 vm_object_deallocate(obj);
1284 /* On failure, drop our mapping reference. */
1285 VM_OBJECT_WLOCK(obj);
1286 shmfd->shm_kmappings--;
1287 VM_OBJECT_WUNLOCK(obj);
1289 return (vm_mmap_to_errno(rv));
1293 * We require the caller to unmap the entire entry. This allows us to
1294 * safely decrement shm_kmappings when a mapping is removed.
1297 shm_unmap(struct file *fp, void *mem, size_t size)
1299 struct shmfd *shmfd;
1300 vm_map_entry_t entry;
1301 vm_offset_t kva, ofs;
1309 if (fp->f_type != DTYPE_SHM)
1312 kva = (vm_offset_t)mem;
1313 ofs = kva & PAGE_MASK;
1314 kva = trunc_page(kva);
1315 size = round_page(size + ofs);
1317 rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1318 &obj, &pindex, &prot, &wired);
1319 if (rv != KERN_SUCCESS)
1321 if (entry->start != kva || entry->end != kva + size) {
1322 vm_map_lookup_done(map, entry);
1325 vm_map_lookup_done(map, entry);
1326 if (obj != shmfd->shm_object)
1328 vm_map_remove(map, kva, kva + size);
1329 VM_OBJECT_WLOCK(obj);
1330 KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1331 shmfd->shm_kmappings--;
1332 VM_OBJECT_WUNLOCK(obj);
1337 shm_fill_kinfo_locked(struct shmfd *shmfd, struct kinfo_file *kif, bool list)
1339 const char *path, *pr_path;
1343 sx_assert(&shm_dict_lock, SA_LOCKED);
1344 kif->kf_type = KF_TYPE_SHM;
1345 kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;
1346 kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1347 if (shmfd->shm_path != NULL) {
1348 if (shmfd->shm_path != NULL) {
1349 path = shmfd->shm_path;
1350 pr_path = curthread->td_ucred->cr_prison->pr_path;
1351 if (strcmp(pr_path, "/") != 0) {
1352 /* Return the jail-rooted pathname. */
1353 pr_pathlen = strlen(pr_path);
1354 visible = strncmp(path, pr_path, pr_pathlen)
1355 == 0 && path[pr_pathlen] == '/';
1356 if (list && !visible)
1361 strlcpy(kif->kf_path, path, sizeof(kif->kf_path));
1368 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif,
1369 struct filedesc *fdp __unused)
1373 sx_slock(&shm_dict_lock);
1374 res = shm_fill_kinfo_locked(fp->f_data, kif, false);
1375 sx_sunlock(&shm_dict_lock);
1380 shm_add_seals(struct file *fp, int seals)
1382 struct shmfd *shmfd;
1384 vm_ooffset_t writemappings;
1389 rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
1392 /* Even already-set seals should result in EPERM. */
1393 if ((shmfd->shm_seals & F_SEAL_SEAL) != 0) {
1397 nseals = seals & ~shmfd->shm_seals;
1398 if ((nseals & F_SEAL_WRITE) != 0) {
1400 * The rangelock above prevents writable mappings from being
1401 * added after we've started applying seals. The RLOCK here
1402 * is to avoid torn reads on ILP32 arches as unmapping/reducing
1403 * writemappings will be done without a rangelock.
1405 VM_OBJECT_RLOCK(shmfd->shm_object);
1406 writemappings = shmfd->shm_object->un_pager.swp.writemappings;
1407 VM_OBJECT_RUNLOCK(shmfd->shm_object);
1408 /* kmappings are also writable */
1409 if (writemappings > 0) {
1414 shmfd->shm_seals |= nseals;
1416 rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
1421 shm_get_seals(struct file *fp, int *seals)
1423 struct shmfd *shmfd;
1426 *seals = shmfd->shm_seals;
1431 sysctl_posix_shm_list(SYSCTL_HANDLER_ARGS)
1433 struct shm_mapping *shmm;
1435 struct kinfo_file kif;
1440 sbuf_new_for_sysctl(&sb, NULL, sizeof(struct kinfo_file) * 5, req);
1441 sbuf_clear_flags(&sb, SBUF_INCLUDENUL);
1444 sx_slock(&shm_dict_lock);
1445 for (i = 0; i < shm_hash + 1; i++) {
1446 LIST_FOREACH(shmm, &shm_dictionary[i], sm_link) {
1447 error = shm_fill_kinfo_locked(shmm->sm_shmfd,
1454 if (req->oldptr != NULL &&
1455 kif.kf_structsize + curlen > req->oldlen)
1457 error = sbuf_bcat(&sb, &kif, kif.kf_structsize) == 0 ?
1461 curlen += kif.kf_structsize;
1464 sx_sunlock(&shm_dict_lock);
1465 error2 = sbuf_finish(&sb);
1467 return (error != 0 ? error : error2);
1470 SYSCTL_PROC(_kern_ipc, OID_AUTO, posix_shm_list,
1471 CTLFLAG_RD | CTLFLAG_MPSAFE | CTLTYPE_OPAQUE,
1472 NULL, 0, sysctl_posix_shm_list, "",
1476 kern_shm_open2(struct thread *td, const char *path, int flags, mode_t mode,
1477 int shmflags, const char *name __unused)
1481 if ((shmflags & ~SHM_ALLOW_SEALING) != 0)
1484 initial_seals = F_SEAL_SEAL;
1485 if ((shmflags & SHM_ALLOW_SEALING) != 0)
1486 initial_seals &= ~F_SEAL_SEAL;
1487 return (kern_shm_open(td, path, flags, mode, NULL, initial_seals));
1491 * This version of the shm_open() interface leaves CLOEXEC behavior up to the
1492 * caller, and libc will enforce it for the traditional shm_open() call. This
1493 * allows other consumers, like memfd_create(), to opt-in for CLOEXEC. This
1494 * interface also includes a 'name' argument that is currently unused, but could
1495 * potentially be exported later via some interface for debugging purposes.
1496 * From the kernel's perspective, it is optional. Individual consumers like
1497 * memfd_create() may require it in order to be compatible with other systems
1498 * implementing the same function.
1501 sys_shm_open2(struct thread *td, struct shm_open2_args *uap)
1504 return (kern_shm_open2(td, uap->path, uap->flags, uap->mode,
1505 uap->shmflags, uap->name));