]>
git.ipfire.org Git - thirdparty/systemd.git/blob - src/nspawn/nspawn-patch-uid.c
1 /* SPDX-License-Identifier: LGPL-2.1+ */
4 #include <sys/statvfs.h>
9 #include "dirent-util.h"
12 #include "missing_magic.h"
13 #include "nspawn-def.h"
14 #include "nspawn-patch-uid.h"
15 #include "stat-util.h"
16 #include "stdio-util.h"
17 #include "string-util.h"
19 #include "user-util.h"
23 static int get_acl(int fd
, const char *name
, acl_type_t type
, acl_t
*ret
) {
24 char procfs_path
[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int) + 1];
31 _cleanup_close_
int child_fd
= -1;
33 child_fd
= openat(fd
, name
, O_PATH
|O_CLOEXEC
|O_NOFOLLOW
);
37 xsprintf(procfs_path
, "/proc/self/fd/%i", child_fd
);
38 acl
= acl_get_file(procfs_path
, type
);
39 } else if (type
== ACL_TYPE_ACCESS
)
42 xsprintf(procfs_path
, "/proc/self/fd/%i", fd
);
43 acl
= acl_get_file(procfs_path
, type
);
52 static int set_acl(int fd
, const char *name
, acl_type_t type
, acl_t acl
) {
53 char procfs_path
[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int) + 1];
60 _cleanup_close_
int child_fd
= -1;
62 child_fd
= openat(fd
, name
, O_PATH
|O_CLOEXEC
|O_NOFOLLOW
);
66 xsprintf(procfs_path
, "/proc/self/fd/%i", child_fd
);
67 r
= acl_set_file(procfs_path
, type
, acl
);
68 } else if (type
== ACL_TYPE_ACCESS
)
69 r
= acl_set_fd(fd
, acl
);
71 xsprintf(procfs_path
, "/proc/self/fd/%i", fd
);
72 r
= acl_set_file(procfs_path
, type
, acl
);
80 static int shift_acl(acl_t acl
, uid_t shift
, acl_t
*ret
) {
81 _cleanup_(acl_freep
) acl_t copy
= NULL
;
88 r
= acl_get_entry(acl
, ACL_FIRST_ENTRY
, &i
);
92 uid_t
*old_uid
, new_uid
;
96 if (acl_get_tag_type(i
, &tag
) < 0)
99 if (IN_SET(tag
, ACL_USER
, ACL_GROUP
)) {
101 /* We don't distinguish here between uid_t and gid_t, let's make sure the compiler checks that
102 * this is actually OK */
103 assert_cc(sizeof(uid_t
) == sizeof(gid_t
));
105 old_uid
= acl_get_qualifier(i
);
109 new_uid
= shift
| (*old_uid
& UINT32_C(0xFFFF));
110 if (!uid_is_valid(new_uid
))
113 modify
= new_uid
!= *old_uid
;
114 if (modify
&& !copy
) {
117 /* There's no copy of the ACL yet? if so, let's create one, and start the loop from the
118 * beginning, so that we copy all entries, starting from the first, this time. */
120 n
= acl_entries(acl
);
128 /* Seek back to the beginning */
129 r
= acl_get_entry(acl
, ACL_FIRST_ENTRY
, &i
);
137 acl_entry_t new_entry
;
139 if (acl_create_entry(©
, &new_entry
) < 0)
142 if (acl_copy_entry(new_entry
, i
) < 0)
146 if (acl_set_qualifier(new_entry
, &new_uid
) < 0)
150 r
= acl_get_entry(acl
, ACL_NEXT_ENTRY
, &i
);
155 *ret
= TAKE_PTR(copy
);
160 static int patch_acls(int fd
, const char *name
, const struct stat
*st
, uid_t shift
) {
161 _cleanup_(acl_freep
) acl_t acl
= NULL
, shifted
= NULL
;
162 bool changed
= false;
168 /* ACLs are not supported on symlinks, there's no point in trying */
169 if (S_ISLNK(st
->st_mode
))
172 r
= get_acl(fd
, name
, ACL_TYPE_ACCESS
, &acl
);
173 if (r
== -EOPNOTSUPP
)
178 r
= shift_acl(acl
, shift
, &shifted
);
182 r
= set_acl(fd
, name
, ACL_TYPE_ACCESS
, shifted
);
189 if (S_ISDIR(st
->st_mode
)) {
193 acl
= shifted
= NULL
;
195 r
= get_acl(fd
, name
, ACL_TYPE_DEFAULT
, &acl
);
199 r
= shift_acl(acl
, shift
, &shifted
);
203 r
= set_acl(fd
, name
, ACL_TYPE_DEFAULT
, shifted
);
216 static int patch_acls(int fd
, const char *name
, const struct stat
*st
, uid_t shift
) {
222 static int patch_fd(int fd
, const char *name
, const struct stat
*st
, uid_t shift
) {
225 bool changed
= false;
231 new_uid
= shift
| (st
->st_uid
& UINT32_C(0xFFFF));
232 new_gid
= (gid_t
) shift
| (st
->st_gid
& UINT32_C(0xFFFF));
234 if (!uid_is_valid(new_uid
) || !gid_is_valid(new_gid
))
237 if (st
->st_uid
!= new_uid
|| st
->st_gid
!= new_gid
) {
239 r
= fchownat(fd
, name
, new_uid
, new_gid
, AT_SYMLINK_NOFOLLOW
);
241 r
= fchown(fd
, new_uid
, new_gid
);
245 /* The Linux kernel alters the mode in some cases of chown(). Let's undo this. */
247 if (!S_ISLNK(st
->st_mode
))
248 r
= fchmodat(fd
, name
, st
->st_mode
, 0);
249 else /* AT_SYMLINK_NOFOLLOW is not available for fchmodat() */
252 r
= fchmod(fd
, st
->st_mode
);
259 r
= patch_acls(fd
, name
, st
, shift
);
263 return r
> 0 || changed
;
267 * Check if the filesystem is fully compatible with user namespaces or
268 * UID/GID patching. Some filesystems in this list can be fully mounted inside
269 * user namespaces, however their inodes may relate to host resources or only
270 * valid in the global user namespace, therefore no patching should be applied.
272 static int is_fs_fully_userns_compatible(const struct statfs
*sfs
) {
276 return F_TYPE_EQUAL(sfs
->f_type
, BINFMTFS_MAGIC
) ||
277 F_TYPE_EQUAL(sfs
->f_type
, CGROUP_SUPER_MAGIC
) ||
278 F_TYPE_EQUAL(sfs
->f_type
, CGROUP2_SUPER_MAGIC
) ||
279 F_TYPE_EQUAL(sfs
->f_type
, DEBUGFS_MAGIC
) ||
280 F_TYPE_EQUAL(sfs
->f_type
, DEVPTS_SUPER_MAGIC
) ||
281 F_TYPE_EQUAL(sfs
->f_type
, EFIVARFS_MAGIC
) ||
282 F_TYPE_EQUAL(sfs
->f_type
, HUGETLBFS_MAGIC
) ||
283 F_TYPE_EQUAL(sfs
->f_type
, MQUEUE_MAGIC
) ||
284 F_TYPE_EQUAL(sfs
->f_type
, PROC_SUPER_MAGIC
) ||
285 F_TYPE_EQUAL(sfs
->f_type
, PSTOREFS_MAGIC
) ||
286 F_TYPE_EQUAL(sfs
->f_type
, SELINUX_MAGIC
) ||
287 F_TYPE_EQUAL(sfs
->f_type
, SMACK_MAGIC
) ||
288 F_TYPE_EQUAL(sfs
->f_type
, SECURITYFS_MAGIC
) ||
289 F_TYPE_EQUAL(sfs
->f_type
, BPF_FS_MAGIC
) ||
290 F_TYPE_EQUAL(sfs
->f_type
, TRACEFS_MAGIC
) ||
291 F_TYPE_EQUAL(sfs
->f_type
, SYSFS_MAGIC
);
294 static int recurse_fd(int fd
, bool donate_fd
, const struct stat
*st
, uid_t shift
, bool is_toplevel
) {
295 _cleanup_closedir_
DIR *d
= NULL
;
296 bool changed
= false;
302 if (fstatfs(fd
, &sfs
) < 0)
305 /* We generally want to permit crossing of mount boundaries when patching the UIDs/GIDs. However, we probably
306 * shouldn't do this for /proc and /sys if that is already mounted into place. Hence, let's stop the recursion
307 * when we hit procfs, sysfs or some other special file systems. */
309 r
= is_fs_fully_userns_compatible(&sfs
);
313 r
= 0; /* don't recurse */
317 /* Also, if we hit a read-only file system, then don't bother, skip the whole subtree */
318 if ((sfs
.f_flags
& ST_RDONLY
) ||
319 access_fd(fd
, W_OK
) == -EROFS
)
322 if (S_ISDIR(st
->st_mode
)) {
328 copy
= fcntl(fd
, F_DUPFD_CLOEXEC
, 3);
345 FOREACH_DIRENT_ALL(de
, d
, r
= -errno
; goto finish
) {
348 if (dot_or_dot_dot(de
->d_name
))
351 if (fstatat(dirfd(d
), de
->d_name
, &fst
, AT_SYMLINK_NOFOLLOW
) < 0) {
356 if (S_ISDIR(fst
.st_mode
)) {
359 subdir_fd
= openat(dirfd(d
), de
->d_name
, O_RDONLY
|O_NONBLOCK
|O_DIRECTORY
|O_CLOEXEC
|O_NOFOLLOW
|O_NOATIME
);
366 r
= recurse_fd(subdir_fd
, true, &fst
, shift
, false);
373 r
= patch_fd(dirfd(d
), de
->d_name
, &fst
, shift
);
382 /* After we descended, also patch the directory itself. It's key to do this in this order so that the top-level
383 * directory is patched as very last object in the tree, so that we can use it as quick indicator whether the
384 * tree is properly chown()ed already. */
385 r
= patch_fd(d
? dirfd(d
) : fd
, NULL
, st
, shift
);
396 _cleanup_free_
char *name
= NULL
;
398 /* When we hit a ready-only subtree we simply skip it, but log about it. */
399 (void) fd_get_path(fd
, &name
);
400 log_debug("Skipping read-only file or directory %s.", strna(name
));
411 static int fd_patch_uid_internal(int fd
, bool donate_fd
, uid_t shift
, uid_t range
) {
417 /* Recursively adjusts the UID/GIDs of all files of a directory tree. This is used to automatically fix up an
418 * OS tree to the used user namespace UID range. Note that this automatic adjustment only works for UID ranges
419 * following the concept that the upper 16bit of a UID identify the container, and the lower 16bit are the actual
420 * UID within the container. */
422 if ((shift
& 0xFFFF) != 0) {
423 /* We only support containers where the shift starts at a 2^16 boundary */
428 if (shift
== UID_BUSY_BASE
) {
433 if (range
!= 0x10000) {
434 /* We only support containers with 16bit UID ranges for the patching logic */
439 if (fstat(fd
, &st
) < 0) {
444 if ((uint32_t) st
.st_uid
>> 16 != (uint32_t) st
.st_gid
>> 16) {
445 /* We only support containers where the uid/gid container ID match */
450 /* Try to detect if the range is already right. Of course, this a pretty drastic optimization, as we assume
451 * that if the top-level dir has the right upper 16bit assigned, then everything below will have too... */
452 if (((uint32_t) (st
.st_uid
^ shift
) >> 16) == 0)
455 /* Before we start recursively chowning, mark the top-level dir as "busy" by chowning it to the "busy"
456 * range. Should we be interrupted in the middle of our work, we'll see it owned by this user and will start
457 * chown()ing it again, unconditionally, as the busy UID is not a valid UID we'd everpick for ourselves. */
459 if ((st
.st_uid
& UID_BUSY_MASK
) != UID_BUSY_BASE
) {
461 UID_BUSY_BASE
| (st
.st_uid
& ~UID_BUSY_MASK
),
462 (gid_t
) UID_BUSY_BASE
| (st
.st_gid
& ~(gid_t
) UID_BUSY_MASK
)) < 0) {
468 return recurse_fd(fd
, donate_fd
, &st
, shift
, true);
477 int path_patch_uid(const char *path
, uid_t shift
, uid_t range
) {
480 fd
= open(path
, O_RDONLY
|O_NONBLOCK
|O_DIRECTORY
|O_CLOEXEC
|O_NOFOLLOW
|O_NOATIME
);
484 return fd_patch_uid_internal(fd
, true, shift
, range
);