1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
7 #include <sys/statvfs.h>
9 #include <linux/loop.h>
14 #include "alloc-util.h"
16 #include "dissect-image.h"
17 #include "exec-util.h"
18 #include "extract-word.h"
22 #include "glyph-util.h"
24 #include "initrd-util.h"
25 #include "label-util.h"
26 #include "libmount-util.h"
27 #include "missing_mount.h"
28 #include "missing_syscall.h"
29 #include "mkdir-label.h"
30 #include "mount-util.h"
31 #include "mountpoint-util.h"
32 #include "namespace-util.h"
33 #include "parse-util.h"
34 #include "path-util.h"
35 #include "process-util.h"
37 #include "sort-util.h"
38 #include "stat-util.h"
39 #include "stdio-util.h"
40 #include "string-table.h"
41 #include "string-util.h"
43 #include "tmpfile-util.h"
44 #include "user-util.h"
46 int umount_recursive_full(const char *prefix
, int flags
, char **keep
) {
47 _cleanup_fclose_
FILE *f
= NULL
;
50 /* Try to umount everything recursively below a directory. Also, take care of stacked mounts, and
51 * keep unmounting them until they are gone. */
53 f
= fopen("/proc/self/mountinfo", "re"); /* Pin the file, in case we unmount /proc/ as part of the logic here */
55 return log_debug_errno(errno
, "Failed to open /proc/self/mountinfo: %m");
58 _cleanup_(mnt_free_tablep
) struct libmnt_table
*table
= NULL
;
59 _cleanup_(mnt_free_iterp
) struct libmnt_iter
*iter
= NULL
;
62 r
= libmount_parse("/proc/self/mountinfo", f
, &table
, &iter
);
64 return log_debug_errno(r
, "Failed to parse /proc/self/mountinfo: %m");
67 bool shall_keep
= false;
71 r
= mnt_table_next_fs(table
, iter
, &fs
);
75 return log_debug_errno(r
, "Failed to get next entry from /proc/self/mountinfo: %m");
77 path
= mnt_fs_get_target(fs
);
81 if (prefix
&& !path_startswith(path
, prefix
)) {
82 log_trace("Not unmounting %s, outside of prefix: %s", path
, prefix
);
87 /* Match against anything in the path to the dirs to keep, or below the dirs to keep */
88 if (path_startswith(path
, *k
) || path_startswith(*k
, path
)) {
93 log_debug("Not unmounting %s, referenced by keep list.", path
);
97 if (umount2(path
, flags
| UMOUNT_NOFOLLOW
) < 0) {
98 log_debug_errno(errno
, "Failed to umount %s, ignoring: %m", path
);
102 log_trace("Successfully unmounted %s", path
);
119 #define MS_CONVERTIBLE_FLAGS (MS_RDONLY|MS_NOSUID|MS_NODEV|MS_NOEXEC|MS_NOSYMFOLLOW)
121 static uint64_t ms_flags_to_mount_attr(unsigned long a
) {
124 if (FLAGS_SET(a
, MS_RDONLY
))
125 f
|= MOUNT_ATTR_RDONLY
;
127 if (FLAGS_SET(a
, MS_NOSUID
))
128 f
|= MOUNT_ATTR_NOSUID
;
130 if (FLAGS_SET(a
, MS_NODEV
))
131 f
|= MOUNT_ATTR_NODEV
;
133 if (FLAGS_SET(a
, MS_NOEXEC
))
134 f
|= MOUNT_ATTR_NOEXEC
;
136 if (FLAGS_SET(a
, MS_NOSYMFOLLOW
))
137 f
|= MOUNT_ATTR_NOSYMFOLLOW
;
142 static bool skip_mount_set_attr
= false;
144 /* Use this function only if you do not have direct access to /proc/self/mountinfo but the caller can open it
145 * for you. This is the case when /proc is masked or not mounted. Otherwise, use bind_remount_recursive. */
146 int bind_remount_recursive_with_mountinfo(
148 unsigned long new_flags
,
149 unsigned long flags_mask
,
151 FILE *proc_self_mountinfo
) {
153 _cleanup_fclose_
FILE *proc_self_mountinfo_opened
= NULL
;
154 _cleanup_set_free_ Set
*done
= NULL
;
155 unsigned n_tries
= 0;
160 if ((flags_mask
& ~MS_CONVERTIBLE_FLAGS
) == 0 && strv_isempty(deny_list
) && !skip_mount_set_attr
) {
161 /* Let's take a shortcut for all the flags we know how to convert into mount_setattr() flags */
163 if (mount_setattr(AT_FDCWD
, prefix
, AT_SYMLINK_NOFOLLOW
|AT_RECURSIVE
,
164 &(struct mount_attr
) {
165 .attr_set
= ms_flags_to_mount_attr(new_flags
& flags_mask
),
166 .attr_clr
= ms_flags_to_mount_attr(~new_flags
& flags_mask
),
167 }, MOUNT_ATTR_SIZE_VER0
) < 0) {
169 log_debug_errno(errno
, "mount_setattr() failed, falling back to classic remounting: %m");
171 /* We fall through to classic behaviour if not supported (i.e. kernel < 5.12). We
172 * also do this for all other kinds of errors since they are so many different, and
173 * mount_setattr() has no graceful mode where it continues despite seeing errors one
174 * some mounts, but we want that. Moreover mount_setattr() only works on the mount
175 * point inode itself, not a non-mount point inode, and we want to support arbitrary
178 if (ERRNO_IS_NOT_SUPPORTED(errno
)) /* if not supported, then don't bother at all anymore */
179 skip_mount_set_attr
= true;
181 return 0; /* Nice, this worked! */
184 if (!proc_self_mountinfo
) {
185 r
= fopen_unlocked("/proc/self/mountinfo", "re", &proc_self_mountinfo_opened
);
189 proc_self_mountinfo
= proc_self_mountinfo_opened
;
192 /* Recursively remount a directory (and all its submounts) with desired flags (MS_READONLY,
193 * MS_NOSUID, MS_NOEXEC). If the directory is already mounted, we reuse the mount and simply mark it
194 * MS_BIND|MS_RDONLY (or remove the MS_RDONLY for read-write operation), ditto for other flags. If it
195 * isn't we first make it one. Afterwards we apply (or remove) the flags to all submounts we can
196 * access, too. When mounts are stacked on the same mount point we only care for each individual
197 * "top-level" mount on each point, as we cannot influence/access the underlying mounts anyway. We do
198 * not have any effect on future submounts that might get propagated, they might be writable
199 * etc. This includes future submounts that have been triggered via autofs. Also note that we can't
200 * operate atomically here. Mounts established while we process the tree might or might not get
201 * noticed and thus might or might not be covered.
203 * If the "deny_list" parameter is specified it may contain a list of subtrees to exclude from the
204 * remount operation. Note that we'll ignore the deny list for the top-level path. */
207 _cleanup_(mnt_free_tablep
) struct libmnt_table
*table
= NULL
;
208 _cleanup_(mnt_free_iterp
) struct libmnt_iter
*iter
= NULL
;
209 _cleanup_hashmap_free_ Hashmap
*todo
= NULL
;
210 bool top_autofs
= false;
212 if (n_tries
++ >= 32) /* Let's not retry this loop forever */
215 rewind(proc_self_mountinfo
);
217 r
= libmount_parse("/proc/self/mountinfo", proc_self_mountinfo
, &table
, &iter
);
219 return log_debug_errno(r
, "Failed to parse /proc/self/mountinfo: %m");
222 _cleanup_free_
char *d
= NULL
;
223 const char *path
, *type
, *opts
;
224 unsigned long flags
= 0;
225 struct libmnt_fs
*fs
;
227 r
= mnt_table_next_fs(table
, iter
, &fs
);
228 if (r
== 1) /* EOF */
231 return log_debug_errno(r
, "Failed to get next entry from /proc/self/mountinfo: %m");
233 path
= mnt_fs_get_target(fs
);
237 if (!path_startswith(path
, prefix
))
240 type
= mnt_fs_get_fstype(fs
);
244 /* Let's ignore autofs mounts. If they aren't triggered yet, we want to avoid
245 * triggering them, as we don't make any guarantees for future submounts anyway. If
246 * they are already triggered, then we will find another entry for this. */
247 if (streq(type
, "autofs")) {
248 top_autofs
= top_autofs
|| path_equal(path
, prefix
);
252 if (set_contains(done
, path
))
255 /* Ignore this mount if it is deny-listed, but only if it isn't the top-level mount
256 * we shall operate on. */
257 if (!path_equal(path
, prefix
)) {
258 bool deny_listed
= false;
260 STRV_FOREACH(i
, deny_list
) {
261 if (path_equal(*i
, prefix
))
264 if (!path_startswith(*i
, prefix
))
267 if (path_startswith(path
, *i
)) {
269 log_trace("Not remounting %s deny-listed by %s, called for %s", path
, *i
, prefix
);
278 opts
= mnt_fs_get_vfs_options(fs
);
280 r
= mnt_optstr_get_flags(opts
, &flags
, mnt_get_builtin_optmap(MNT_LINUX_MAP
));
282 log_debug_errno(r
, "Could not get flags for '%s', ignoring: %m", path
);
289 r
= hashmap_ensure_put(&todo
, &path_hash_ops_free
, d
, ULONG_TO_PTR(flags
));
291 /* If the same path was recorded, but with different mount flags, update it:
292 * it means a mount point is overmounted, and libmount returns the "bottom" (or
293 * older one) first, but we want to reapply the flags from the "top" (or newer
294 * one). See: https://github.com/systemd/systemd/issues/20032
295 * Note that this shouldn't really fail, as we were just told that the key
296 * exists, and it's an update so we want 'd' to be freed immediately. */
297 r
= hashmap_update(todo
, d
, ULONG_TO_PTR(flags
));
304 /* Check if the top-level directory was among what we have seen so far. For that check both
305 * 'done' and 'todo'. Also check 'top_autofs' because if the top-level dir is an autofs we'll
306 * not include it in either set but will set this bool. */
307 if (!set_contains(done
, prefix
) &&
308 !(top_autofs
|| hashmap_contains(todo
, prefix
))) {
310 /* The prefix directory itself is not yet a mount, make it one. */
311 r
= mount_nofollow(prefix
, prefix
, NULL
, MS_BIND
|MS_REC
, NULL
);
315 /* Immediately rescan, so that we pick up the new mount's flags */
319 /* If we have no submounts to process anymore, we are done */
320 if (hashmap_isempty(todo
))
327 /* Take the first mount from our list of mounts to still process */
328 flags
= PTR_TO_ULONG(hashmap_steal_first_key_and_value(todo
, (void**) &x
));
332 r
= set_ensure_consume(&done
, &path_hash_ops_free
, x
);
333 if (IN_SET(r
, 0, -EEXIST
))
334 continue; /* Already done */
338 /* Now, remount this with the new flags set, but exclude MS_RELATIME from it. (It's
339 * the default anyway, thus redundant, and in userns we'll get an error if we try to
340 * explicitly enable it) */
341 r
= mount_nofollow(NULL
, x
, NULL
, ((flags
& ~flags_mask
)|MS_BIND
|MS_REMOUNT
|new_flags
) & ~MS_RELATIME
, NULL
);
345 /* OK, so the remount of this entry failed. We'll ultimately ignore this in
346 * almost all cases (there are simply so many reasons why this can fail,
347 * think autofs, NFS, FUSE, …), but let's generate useful debug messages at
350 q
= path_is_mount_point(x
, NULL
, 0);
351 if (IN_SET(q
, 0, -ENOENT
)) {
352 /* Hmm, whaaaa? The mount point is not actually a mount point? Then
353 * it is either obstructed by a later mount or somebody has been
354 * racing against us and removed it. Either way the mount point
355 * doesn't matter to us, let's ignore it hence. */
356 log_debug_errno(r
, "Mount point '%s' to remount is not a mount point anymore, ignoring remount failure: %m", x
);
359 if (q
< 0) /* Any other error on this? Just log and continue */
360 log_debug_errno(q
, "Failed to determine whether '%s' is a mount point or not, ignoring: %m", x
);
362 if (((flags
^ new_flags
) & flags_mask
& ~MS_RELATIME
) == 0) { /* ignore MS_RELATIME while comparing */
363 log_debug_errno(r
, "Couldn't remount '%s', but the flags already match what we want, hence ignoring: %m", x
);
367 /* Make this fatal if this is the top-level mount */
368 if (path_equal(x
, prefix
))
371 /* If this is not the top-level mount, then handle this gracefully: log but
372 * otherwise ignore. With NFS, FUSE, autofs there are just too many reasons
373 * this might fail without a chance for us to do anything about it, let's
374 * hence be strict on the top-level mount and lenient on the inner ones. */
375 log_debug_errno(r
, "Couldn't remount submount '%s' for unexpected reason, ignoring: %m", x
);
379 log_trace("Remounted %s.", x
);
384 int bind_remount_one_with_mountinfo(
386 unsigned long new_flags
,
387 unsigned long flags_mask
,
388 FILE *proc_self_mountinfo
) {
390 _cleanup_(mnt_free_tablep
) struct libmnt_table
*table
= NULL
;
391 unsigned long flags
= 0;
392 struct libmnt_fs
*fs
;
397 assert(proc_self_mountinfo
);
399 if ((flags_mask
& ~MS_CONVERTIBLE_FLAGS
) == 0 && !skip_mount_set_attr
) {
400 /* Let's take a shortcut for all the flags we know how to convert into mount_setattr() flags */
402 if (mount_setattr(AT_FDCWD
, path
, AT_SYMLINK_NOFOLLOW
,
403 &(struct mount_attr
) {
404 .attr_set
= ms_flags_to_mount_attr(new_flags
& flags_mask
),
405 .attr_clr
= ms_flags_to_mount_attr(~new_flags
& flags_mask
),
406 }, MOUNT_ATTR_SIZE_VER0
) < 0) {
408 log_debug_errno(errno
, "mount_setattr() didn't work, falling back to classic remounting: %m");
410 if (ERRNO_IS_NOT_SUPPORTED(errno
)) /* if not supported, then don't bother at all anymore */
411 skip_mount_set_attr
= true;
413 return 0; /* Nice, this worked! */
416 rewind(proc_self_mountinfo
);
418 table
= mnt_new_table();
422 r
= mnt_table_parse_stream(table
, proc_self_mountinfo
, "/proc/self/mountinfo");
426 fs
= mnt_table_find_target(table
, path
, MNT_ITER_FORWARD
);
428 if (laccess(path
, F_OK
) < 0) /* Hmm, it's not in the mount table, but does it exist at all? */
431 return -EINVAL
; /* Not a mount point we recognize */
434 opts
= mnt_fs_get_vfs_options(fs
);
436 r
= mnt_optstr_get_flags(opts
, &flags
, mnt_get_builtin_optmap(MNT_LINUX_MAP
));
438 log_debug_errno(r
, "Could not get flags for '%s', ignoring: %m", path
);
441 r
= mount_nofollow(NULL
, path
, NULL
, ((flags
& ~flags_mask
)|MS_BIND
|MS_REMOUNT
|new_flags
) & ~MS_RELATIME
, NULL
);
443 if (((flags
^ new_flags
) & flags_mask
& ~MS_RELATIME
) != 0) /* Ignore MS_RELATIME again,
444 * since kernel adds it in
445 * everywhere, because it's the
449 /* Let's handle redundant remounts gracefully */
450 log_debug_errno(r
, "Failed to remount '%s' but flags already match what we want, ignoring: %m", path
);
456 static int mount_switch_root_pivot(int fd_newroot
, const char *path
) {
457 assert(fd_newroot
>= 0);
460 /* Change into the new rootfs. */
461 if (fchdir(fd_newroot
) < 0)
462 return log_debug_errno(errno
, "Failed to change into new rootfs '%s': %m", path
);
464 /* Let the kernel tuck the new root under the old one. */
465 if (pivot_root(".", ".") < 0)
466 return log_debug_errno(errno
, "Failed to pivot root to new rootfs '%s': %m", path
);
468 /* Get rid of the old root and reveal our brand new root. (This will always operate on the top-most
469 * mount on our cwd, regardless what our current directory actually points to.) */
470 if (umount2(".", MNT_DETACH
) < 0)
471 return log_debug_errno(errno
, "Failed to unmount old rootfs: %m");
476 static int mount_switch_root_move(int fd_newroot
, const char *path
) {
477 assert(fd_newroot
>= 0);
480 /* Change into the new rootfs. */
481 if (fchdir(fd_newroot
) < 0)
482 return log_debug_errno(errno
, "Failed to change into new rootfs '%s': %m", path
);
484 /* Move the new root fs */
485 if (mount(".", "/", NULL
, MS_MOVE
, NULL
) < 0)
486 return log_debug_errno(errno
, "Failed to move new rootfs '%s': %m", path
);
488 /* Also change chroot dir */
490 return log_debug_errno(errno
, "Failed to chroot to new rootfs '%s': %m", path
);
495 int mount_switch_root_full(const char *path
, unsigned long mount_propagation_flag
, bool force_ms_move
) {
496 _cleanup_close_
int fd_newroot
= -EBADF
;
500 assert(mount_propagation_flag_is_valid(mount_propagation_flag
));
502 fd_newroot
= open(path
, O_PATH
|O_DIRECTORY
|O_CLOEXEC
|O_NOFOLLOW
);
504 return log_debug_errno(errno
, "Failed to open new rootfs '%s': %m", path
);
506 if (!force_ms_move
) {
507 r
= mount_switch_root_pivot(fd_newroot
, path
);
509 log_debug_errno(r
, "Failed to pivot into new rootfs '%s', will try to use MS_MOVE instead: %m", path
);
510 force_ms_move
= true;
514 /* Failed to pivot_root() fallback to MS_MOVE. For example, this may happen if the rootfs is
515 * an initramfs in which case pivot_root() isn't supported. */
516 r
= mount_switch_root_move(fd_newroot
, path
);
518 return log_debug_errno(r
, "Failed to switch to new rootfs '%s' with MS_MOVE: %m", path
);
521 /* Finally, let's establish the requested propagation flags. */
522 if (mount_propagation_flag
== 0)
525 if (mount(NULL
, ".", NULL
, mount_propagation_flag
| MS_REC
, 0) < 0)
526 return log_debug_errno(errno
, "Failed to turn new rootfs '%s' into %s mount: %m",
527 mount_propagation_flag_to_string(mount_propagation_flag
), path
);
532 int repeat_unmount(const char *path
, int flags
) {
537 /* If there are multiple mounts on a mount point, this
538 * removes them all */
541 if (umount2(path
, flags
) < 0) {
553 int mode_to_inaccessible_node(
554 const char *runtime_dir
,
558 /* This function maps a node type to a corresponding inaccessible file node. These nodes are created
559 * during early boot by PID 1. In some cases we lacked the privs to create the character and block
560 * devices (maybe because we run in an userns environment, or miss CAP_SYS_MKNOD, or run with a
561 * devices policy that excludes device nodes with major and minor of 0), but that's fine, in that
562 * case we use an AF_UNIX file node instead, which is not the same, but close enough for most
563 * uses. And most importantly, the kernel allows bind mounts from socket nodes to any non-directory
564 * file nodes, and that's the most important thing that matters.
566 * Note that the runtime directory argument shall be the top-level runtime directory, i.e. /run/ if
567 * we operate in system context and $XDG_RUNTIME_DIR if we operate in user context. */
569 _cleanup_free_
char *d
= NULL
;
570 const char *node
= NULL
;
575 runtime_dir
= "/run";
577 switch (mode
& S_IFMT
) {
579 node
= "/systemd/inaccessible/reg";
583 node
= "/systemd/inaccessible/dir";
587 node
= "/systemd/inaccessible/chr";
591 node
= "/systemd/inaccessible/blk";
595 node
= "/systemd/inaccessible/fifo";
599 node
= "/systemd/inaccessible/sock";
605 d
= path_join(runtime_dir
, node
);
609 /* On new kernels unprivileged users are permitted to create 0:0 char device nodes (because they also
610 * act as whiteout inode for overlayfs), but no other char or block device nodes. On old kernels no
611 * device node whatsoever may be created by unprivileged processes. Hence, if the caller asks for the
612 * inaccessible block device node let's see if the block device node actually exists, and if not,
613 * fall back to the character device node. From there fall back to the socket device node. This means
614 * in the best case we'll get the right device node type — but if not we'll hopefully at least get a
615 * device node at all. */
618 access(d
, F_OK
) < 0 && errno
== ENOENT
) {
620 d
= path_join(runtime_dir
, "/systemd/inaccessible/chr");
625 if (IN_SET(mode
& S_IFMT
, S_IFBLK
, S_IFCHR
) &&
626 access(d
, F_OK
) < 0 && errno
== ENOENT
) {
628 d
= path_join(runtime_dir
, "/systemd/inaccessible/sock");
637 int mount_flags_to_string(unsigned long flags
, char **ret
) {
638 static const struct {
642 { .flag
= MS_RDONLY
, .name
= "MS_RDONLY", },
643 { .flag
= MS_NOSUID
, .name
= "MS_NOSUID", },
644 { .flag
= MS_NODEV
, .name
= "MS_NODEV", },
645 { .flag
= MS_NOEXEC
, .name
= "MS_NOEXEC", },
646 { .flag
= MS_SYNCHRONOUS
, .name
= "MS_SYNCHRONOUS", },
647 { .flag
= MS_REMOUNT
, .name
= "MS_REMOUNT", },
648 { .flag
= MS_MANDLOCK
, .name
= "MS_MANDLOCK", },
649 { .flag
= MS_DIRSYNC
, .name
= "MS_DIRSYNC", },
650 { .flag
= MS_NOSYMFOLLOW
, .name
= "MS_NOSYMFOLLOW", },
651 { .flag
= MS_NOATIME
, .name
= "MS_NOATIME", },
652 { .flag
= MS_NODIRATIME
, .name
= "MS_NODIRATIME", },
653 { .flag
= MS_BIND
, .name
= "MS_BIND", },
654 { .flag
= MS_MOVE
, .name
= "MS_MOVE", },
655 { .flag
= MS_REC
, .name
= "MS_REC", },
656 { .flag
= MS_SILENT
, .name
= "MS_SILENT", },
657 { .flag
= MS_POSIXACL
, .name
= "MS_POSIXACL", },
658 { .flag
= MS_UNBINDABLE
, .name
= "MS_UNBINDABLE", },
659 { .flag
= MS_PRIVATE
, .name
= "MS_PRIVATE", },
660 { .flag
= MS_SLAVE
, .name
= "MS_SLAVE", },
661 { .flag
= MS_SHARED
, .name
= "MS_SHARED", },
662 { .flag
= MS_RELATIME
, .name
= "MS_RELATIME", },
663 { .flag
= MS_KERNMOUNT
, .name
= "MS_KERNMOUNT", },
664 { .flag
= MS_I_VERSION
, .name
= "MS_I_VERSION", },
665 { .flag
= MS_STRICTATIME
, .name
= "MS_STRICTATIME", },
666 { .flag
= MS_LAZYTIME
, .name
= "MS_LAZYTIME", },
668 _cleanup_free_
char *str
= NULL
;
672 for (size_t i
= 0; i
< ELEMENTSOF(map
); i
++)
673 if (flags
& map
[i
].flag
) {
674 if (!strextend_with_separator(&str
, "|", map
[i
].name
))
676 flags
&= ~map
[i
].flag
;
679 if (!str
|| flags
!= 0)
680 if (strextendf_with_separator(&str
, "|", "%lx", flags
) < 0)
683 *ret
= TAKE_PTR(str
);
687 int mount_verbose_full(
694 bool follow_symlink
) {
696 _cleanup_free_
char *fl
= NULL
, *o
= NULL
;
700 r
= mount_option_mangle(options
, flags
, &f
, &o
);
702 return log_full_errno(error_log_level
, r
,
703 "Failed to mangle mount options %s: %m",
706 (void) mount_flags_to_string(f
, &fl
);
708 if ((f
& MS_REMOUNT
) && !what
&& !type
)
709 log_debug("Remounting %s (%s \"%s\")...",
710 where
, strnull(fl
), strempty(o
));
711 else if (!what
&& !type
)
712 log_debug("Mounting %s (%s \"%s\")...",
713 where
, strnull(fl
), strempty(o
));
714 else if ((f
& MS_BIND
) && !type
)
715 log_debug("Bind-mounting %s on %s (%s \"%s\")...",
716 what
, where
, strnull(fl
), strempty(o
));
717 else if (f
& MS_MOVE
)
718 log_debug("Moving mount %s %s %s (%s \"%s\")...",
719 what
, special_glyph(SPECIAL_GLYPH_ARROW_RIGHT
), where
, strnull(fl
), strempty(o
));
721 log_debug("Mounting %s (%s) on %s (%s \"%s\")...",
722 strna(what
), strna(type
), where
, strnull(fl
), strempty(o
));
725 r
= RET_NERRNO(mount(what
, where
, type
, f
, o
));
727 r
= mount_nofollow(what
, where
, type
, f
, o
);
729 return log_full_errno(error_log_level
, r
,
730 "Failed to mount %s (type %s) on %s (%s \"%s\"): %m",
731 strna(what
), strna(type
), where
, strnull(fl
), strempty(o
));
742 log_debug("Umounting %s...", what
);
744 if (umount2(what
, flags
) < 0)
745 return log_full_errno(error_log_level
, errno
,
746 "Failed to unmount %s: %m", what
);
751 int mount_option_mangle(
753 unsigned long mount_flags
,
754 unsigned long *ret_mount_flags
,
755 char **ret_remaining_options
) {
757 const struct libmnt_optmap
*map
;
758 _cleanup_free_
char *ret
= NULL
;
761 /* This extracts mount flags from the mount options, and stores
762 * non-mount-flag options to '*ret_remaining_options'.
764 * "rw,nosuid,nodev,relatime,size=1630748k,mode=0700,uid=1000,gid=1000"
765 * is split to MS_NOSUID|MS_NODEV|MS_RELATIME and
766 * "size=1630748k,mode=0700,uid=1000,gid=1000".
767 * See more examples in test-mount-util.c.
769 * If 'options' does not contain any non-mount-flag options,
770 * then '*ret_remaining_options' is set to NULL instead of empty string.
771 * The validity of options stored in '*ret_remaining_options' is not checked.
772 * If 'options' is NULL, this just copies 'mount_flags' to *ret_mount_flags. */
774 assert(ret_mount_flags
);
775 assert(ret_remaining_options
);
777 map
= mnt_get_builtin_optmap(MNT_LINUX_MAP
);
781 for (const char *p
= options
;;) {
782 _cleanup_free_
char *word
= NULL
;
783 const struct libmnt_optmap
*ent
;
785 r
= extract_first_word(&p
, &word
, ",", EXTRACT_KEEP_QUOTE
);
791 for (ent
= map
; ent
->name
; ent
++) {
792 /* All entries in MNT_LINUX_MAP do not take any argument.
793 * Thus, ent->name does not contain "=" or "[=]". */
794 if (!streq(word
, ent
->name
))
797 if (!(ent
->mask
& MNT_INVERT
))
798 mount_flags
|= ent
->id
;
799 else if (mount_flags
& ent
->id
)
800 mount_flags
^= ent
->id
;
805 /* If 'word' is not a mount flag, then store it in '*ret_remaining_options'. */
807 !startswith_no_case(word
, "x-") &&
808 !strextend_with_separator(&ret
, ",", word
))
812 *ret_mount_flags
= mount_flags
;
813 *ret_remaining_options
= TAKE_PTR(ret
);
818 static int mount_in_namespace(
820 const char *propagate_path
,
821 const char *incoming_path
,
825 bool make_file_or_directory
,
826 const MountOptions
*options
,
827 const ImagePolicy
*image_policy
,
830 _cleanup_close_pair_
int errno_pipe_fd
[2] = PIPE_EBADF
;
831 _cleanup_close_
int mntns_fd
= -EBADF
, root_fd
= -EBADF
, pidns_fd
= -EBADF
, chased_src_fd
= -EBADF
;
832 char mount_slave
[] = "/tmp/propagate.XXXXXX", *mount_tmp
, *mount_outside
, *p
;
833 bool mount_slave_created
= false, mount_slave_mounted
= false,
834 mount_tmp_created
= false, mount_tmp_mounted
= false,
835 mount_outside_created
= false, mount_outside_mounted
= false;
836 _cleanup_free_
char *chased_src_path
= NULL
;
842 assert(propagate_path
);
843 assert(incoming_path
);
846 assert(!options
|| is_image
);
848 r
= namespace_open(target
, &pidns_fd
, &mntns_fd
, NULL
, NULL
, &root_fd
);
850 return log_debug_errno(r
, "Failed to retrieve FDs of the target process' namespace: %m");
852 r
= in_same_namespace(target
, 0, NAMESPACE_MOUNT
);
854 return log_debug_errno(r
, "Failed to determine if mount namespaces are equal: %m");
855 /* We can't add new mounts at runtime if the process wasn't started in a namespace */
857 return log_debug_errno(SYNTHETIC_ERRNO(EINVAL
), "Failed to activate bind mount in target, not running in a mount namespace");
859 /* One day, when bind mounting /proc/self/fd/n works across namespace boundaries we should rework
860 * this logic to make use of it... */
862 p
= strjoina(propagate_path
, "/");
863 r
= laccess(p
, F_OK
);
865 return log_debug_errno(r
== -ENOENT
? SYNTHETIC_ERRNO(EOPNOTSUPP
) : r
, "Target does not allow propagation of mount points");
867 r
= chase(src
, NULL
, 0, &chased_src_path
, &chased_src_fd
);
869 return log_debug_errno(r
, "Failed to resolve source path of %s: %m", src
);
870 log_debug("Chased source path of %s to %s", src
, chased_src_path
);
872 if (fstat(chased_src_fd
, &st
) < 0)
873 return log_debug_errno(errno
, "Failed to stat() resolved source path %s: %m", src
);
874 if (S_ISLNK(st
.st_mode
)) /* This shouldn't really happen, given that we just chased the symlinks above, but let's better be safe… */
875 return log_debug_errno(SYNTHETIC_ERRNO(EOPNOTSUPP
), "Source directory %s can't be a symbolic link", src
);
877 /* Our goal is to install a new bind mount into the container,
878 possibly read-only. This is irritatingly complex
879 unfortunately, currently.
881 First, we start by creating a private playground in /tmp,
882 that we can mount MS_SLAVE. (Which is necessary, since
883 MS_MOVE cannot be applied to mounts with MS_SHARED parent
886 if (!mkdtemp(mount_slave
))
887 return log_debug_errno(errno
, "Failed to create playground %s: %m", mount_slave
);
889 mount_slave_created
= true;
891 r
= mount_nofollow_verbose(LOG_DEBUG
, mount_slave
, mount_slave
, NULL
, MS_BIND
, NULL
);
895 mount_slave_mounted
= true;
897 r
= mount_nofollow_verbose(LOG_DEBUG
, NULL
, mount_slave
, NULL
, MS_SLAVE
, NULL
);
901 /* Second, we mount the source file or directory to a directory inside of our MS_SLAVE playground. */
902 mount_tmp
= strjoina(mount_slave
, "/mount");
904 r
= mkdir_p(mount_tmp
, 0700);
906 r
= make_mount_point_inode_from_stat(&st
, mount_tmp
, 0700);
908 log_debug_errno(r
, "Failed to create temporary mount point %s: %m", mount_tmp
);
912 mount_tmp_created
= true;
915 r
= verity_dissect_and_mount(chased_src_fd
, chased_src_path
, mount_tmp
, options
, image_policy
, NULL
, NULL
, NULL
, NULL
);
917 r
= mount_follow_verbose(LOG_DEBUG
, FORMAT_PROC_FD_PATH(chased_src_fd
), mount_tmp
, NULL
, MS_BIND
, NULL
);
921 mount_tmp_mounted
= true;
923 /* Third, we remount the new bind mount read-only if requested. */
925 r
= mount_nofollow_verbose(LOG_DEBUG
, NULL
, mount_tmp
, NULL
, MS_BIND
|MS_REMOUNT
|MS_RDONLY
, NULL
);
930 /* Fourth, we move the new bind mount into the propagation directory. This way it will appear there read-only
933 mount_outside
= strjoina(propagate_path
, "/XXXXXX");
934 if (is_image
|| S_ISDIR(st
.st_mode
))
935 r
= mkdtemp(mount_outside
) ? 0 : -errno
;
937 r
= mkostemp_safe(mount_outside
);
941 log_debug_errno(r
, "Cannot create propagation file or directory %s: %m", mount_outside
);
945 mount_outside_created
= true;
947 r
= mount_nofollow_verbose(LOG_DEBUG
, mount_tmp
, mount_outside
, NULL
, MS_MOVE
, NULL
);
951 mount_outside_mounted
= true;
952 mount_tmp_mounted
= false;
954 if (is_image
|| S_ISDIR(st
.st_mode
))
955 (void) rmdir(mount_tmp
);
957 (void) unlink(mount_tmp
);
958 mount_tmp_created
= false;
960 (void) umount_verbose(LOG_DEBUG
, mount_slave
, UMOUNT_NOFOLLOW
);
961 mount_slave_mounted
= false;
963 (void) rmdir(mount_slave
);
964 mount_slave_created
= false;
966 if (pipe2(errno_pipe_fd
, O_CLOEXEC
|O_NONBLOCK
) < 0) {
967 log_debug_errno(errno
, "Failed to create pipe: %m");
971 r
= namespace_fork("(sd-bindmnt)", "(sd-bindmnt-inner)", NULL
, 0, FORK_RESET_SIGNALS
|FORK_DEATHSIG
,
972 pidns_fd
, mntns_fd
, -1, -1, root_fd
, &child
);
976 _cleanup_free_
char *mount_outside_fn
= NULL
, *mount_inside
= NULL
;
978 errno_pipe_fd
[0] = safe_close(errno_pipe_fd
[0]);
980 if (make_file_or_directory
) {
982 (void) mkdir_parents(dest
, 0755);
983 (void) make_mount_point_inode_from_stat(&st
, dest
, 0700);
985 (void) mkdir_p(dest
, 0755);
988 /* Fifth, move the mount to the right place inside */
989 r
= path_extract_filename(mount_outside
, &mount_outside_fn
);
991 log_debug_errno(r
, "Failed to extract filename from propagation file or directory '%s': %m", mount_outside
);
995 mount_inside
= path_join(incoming_path
, mount_outside_fn
);
1001 r
= mount_nofollow_verbose(LOG_DEBUG
, mount_inside
, dest
, NULL
, MS_MOVE
, NULL
);
1005 _exit(EXIT_SUCCESS
);
1008 (void) write(errno_pipe_fd
[1], &r
, sizeof(r
));
1009 errno_pipe_fd
[1] = safe_close(errno_pipe_fd
[1]);
1011 _exit(EXIT_FAILURE
);
1014 errno_pipe_fd
[1] = safe_close(errno_pipe_fd
[1]);
1016 r
= wait_for_terminate_and_check("(sd-bindmnt)", child
, 0);
1018 log_debug_errno(r
, "Failed to wait for child: %m");
1021 if (r
!= EXIT_SUCCESS
) {
1022 if (read(errno_pipe_fd
[0], &r
, sizeof(r
)) == sizeof(r
))
1023 log_debug_errno(r
, "Failed to mount: %m");
1025 log_debug("Child failed.");
1030 if (mount_outside_mounted
)
1031 (void) umount_verbose(LOG_DEBUG
, mount_outside
, UMOUNT_NOFOLLOW
);
1032 if (mount_outside_created
) {
1033 if (is_image
|| S_ISDIR(st
.st_mode
))
1034 (void) rmdir(mount_outside
);
1036 (void) unlink(mount_outside
);
1039 if (mount_tmp_mounted
)
1040 (void) umount_verbose(LOG_DEBUG
, mount_tmp
, UMOUNT_NOFOLLOW
);
1041 if (mount_tmp_created
) {
1042 if (is_image
|| S_ISDIR(st
.st_mode
))
1043 (void) rmdir(mount_tmp
);
1045 (void) unlink(mount_tmp
);
1048 if (mount_slave_mounted
)
1049 (void) umount_verbose(LOG_DEBUG
, mount_slave
, UMOUNT_NOFOLLOW
);
1050 if (mount_slave_created
)
1051 (void) rmdir(mount_slave
);
1056 int bind_mount_in_namespace(
1058 const char *propagate_path
,
1059 const char *incoming_path
,
1063 bool make_file_or_directory
) {
1065 return mount_in_namespace(target
, propagate_path
, incoming_path
, src
, dest
, read_only
, make_file_or_directory
, /* options= */ NULL
, /* image_policy= */ NULL
, /* is_image= */ false);
1068 int mount_image_in_namespace(
1070 const char *propagate_path
,
1071 const char *incoming_path
,
1075 bool make_file_or_directory
,
1076 const MountOptions
*options
,
1077 const ImagePolicy
*image_policy
) {
1079 return mount_in_namespace(target
, propagate_path
, incoming_path
, src
, dest
, read_only
, make_file_or_directory
, options
, image_policy
, /* is_image=*/ true);
1082 int make_mount_point(const char *path
) {
1087 /* If 'path' is already a mount point, does nothing and returns 0. If it is not it makes it one, and returns 1. */
1089 r
= path_is_mount_point(path
, NULL
, 0);
1091 return log_debug_errno(r
, "Failed to determine whether '%s' is a mount point: %m", path
);
1095 r
= mount_nofollow_verbose(LOG_DEBUG
, path
, path
, NULL
, MS_BIND
|MS_REC
, NULL
);
1102 int fd_make_mount_point(int fd
) {
1107 r
= fd_is_mount_point(fd
, NULL
, 0);
1109 return log_debug_errno(r
, "Failed to determine whether file descriptor is a mount point: %m");
1113 r
= mount_follow_verbose(LOG_DEBUG
, FORMAT_PROC_FD_PATH(fd
), FORMAT_PROC_FD_PATH(fd
), NULL
, MS_BIND
|MS_REC
, NULL
);
1120 int make_userns(uid_t uid_shift
, uid_t uid_range
, uid_t owner
, RemountIdmapping idmapping
) {
1121 _cleanup_close_
int userns_fd
= -EBADF
;
1122 _cleanup_free_
char *line
= NULL
;
1124 /* Allocates a userns file descriptor with the mapping we need. For this we'll fork off a child
1125 * process whose only purpose is to give us a new user namespace. It's killed when we got it. */
1127 if (!userns_shift_range_valid(uid_shift
, uid_range
))
1130 if (IN_SET(idmapping
, REMOUNT_IDMAPPING_NONE
, REMOUNT_IDMAPPING_HOST_ROOT
)) {
1131 if (asprintf(&line
, UID_FMT
" " UID_FMT
" " UID_FMT
"\n", 0u, uid_shift
, uid_range
) < 0)
1132 return log_oom_debug();
1134 /* If requested we'll include an entry in the mapping so that the host root user can make
1135 * changes to the uidmapped mount like it normally would. Specifically, we'll map the user
1136 * with UID_MAPPED_ROOT on the backing fs to UID 0. This is useful, since nspawn code wants
1137 * to create various missing inodes in the OS tree before booting into it, and this becomes
1138 * very easy and straightforward to do if it can just do it under its own regular UID. Note
1139 * that in that case the container's runtime uidmap (i.e. the one the container payload
1140 * processes run in) will leave this UID unmapped, i.e. if we accidentally leave files owned
1141 * by host root in the already uidmapped tree around they'll show up as owned by 'nobody',
1142 * which is safe. (Of course, we shouldn't leave such inodes around, but always chown() them
1143 * to the container's own UID range, but it's good to have a safety net, in case we
1145 if (idmapping
== REMOUNT_IDMAPPING_HOST_ROOT
)
1146 if (strextendf(&line
,
1147 UID_FMT
" " UID_FMT
" " UID_FMT
"\n",
1148 UID_MAPPED_ROOT
, 0u, 1u) < 0)
1149 return log_oom_debug();
1152 if (idmapping
== REMOUNT_IDMAPPING_HOST_OWNER
) {
1153 /* Remap the owner of the bind mounted directory to the root user within the container. This
1154 * way every file written by root within the container to the bind-mounted directory will
1155 * be owned by the original user. All other user will remain unmapped. */
1156 if (asprintf(&line
, UID_FMT
" " UID_FMT
" " UID_FMT
"\n", owner
, uid_shift
, 1u) < 0)
1157 return log_oom_debug();
1160 /* We always assign the same UID and GID ranges */
1161 userns_fd
= userns_acquire(line
, line
);
1163 return log_debug_errno(userns_fd
, "Failed to acquire new userns: %m");
1165 return TAKE_FD(userns_fd
);
1168 int remount_idmap_fd(
1172 _cleanup_close_
int mount_fd
= -EBADF
;
1176 assert(userns_fd
>= 0);
1178 /* Clone the mount point */
1179 mount_fd
= open_tree(-1, p
, OPEN_TREE_CLONE
| OPEN_TREE_CLOEXEC
);
1181 return log_debug_errno(errno
, "Failed to open tree of mounted filesystem '%s': %m", p
);
1183 /* Set the user namespace mapping attribute on the cloned mount point */
1184 if (mount_setattr(mount_fd
, "", AT_EMPTY_PATH
| AT_RECURSIVE
,
1185 &(struct mount_attr
) {
1186 .attr_set
= MOUNT_ATTR_IDMAP
,
1187 .userns_fd
= userns_fd
,
1188 }, sizeof(struct mount_attr
)) < 0)
1189 return log_debug_errno(errno
, "Failed to change bind mount attributes for '%s': %m", p
);
1191 /* Remove the old mount point */
1192 r
= umount_verbose(LOG_DEBUG
, p
, UMOUNT_NOFOLLOW
);
1196 /* And place the cloned version in its place */
1197 if (move_mount(mount_fd
, "", -1, p
, MOVE_MOUNT_F_EMPTY_PATH
) < 0)
1198 return log_debug_errno(errno
, "Failed to attach UID mapped mount to '%s': %m", p
);
1203 int remount_idmap(const char *p
, uid_t uid_shift
, uid_t uid_range
, uid_t owner
, RemountIdmapping idmapping
) {
1204 _cleanup_close_
int userns_fd
= -EBADF
;
1206 userns_fd
= make_userns(uid_shift
, uid_range
, owner
, idmapping
);
1210 return remount_idmap_fd(p
, userns_fd
);
1213 typedef struct SubMount
{
1218 static void sub_mount_clear(SubMount
*s
) {
1221 s
->path
= mfree(s
->path
);
1222 s
->mount_fd
= safe_close(s
->mount_fd
);
1225 static void sub_mount_array_free(SubMount
*s
, size_t n
) {
1226 assert(s
|| n
== 0);
1228 for (size_t i
= 0; i
< n
; i
++)
1229 sub_mount_clear(s
+ i
);
1234 static int sub_mount_compare(const SubMount
*a
, const SubMount
*b
) {
1240 return path_compare(a
->path
, b
->path
);
1243 static void sub_mount_drop(SubMount
*s
, size_t n
) {
1244 assert(s
|| n
== 0);
1246 for (size_t m
= 0, i
= 1; i
< n
; i
++) {
1247 if (path_startswith(s
[i
].path
, s
[m
].path
))
1248 sub_mount_clear(s
+ i
);
1254 static int get_sub_mounts(const char *prefix
, SubMount
**ret_mounts
, size_t *ret_n_mounts
) {
1255 _cleanup_(mnt_free_tablep
) struct libmnt_table
*table
= NULL
;
1256 _cleanup_(mnt_free_iterp
) struct libmnt_iter
*iter
= NULL
;
1257 SubMount
*mounts
= NULL
;
1261 CLEANUP_ARRAY(mounts
, n
, sub_mount_array_free
);
1265 assert(ret_n_mounts
);
1267 r
= libmount_parse("/proc/self/mountinfo", NULL
, &table
, &iter
);
1269 return log_debug_errno(r
, "Failed to parse /proc/self/mountinfo: %m");
1272 _cleanup_close_
int mount_fd
= -EBADF
;
1273 _cleanup_free_
char *p
= NULL
;
1274 struct libmnt_fs
*fs
;
1278 r
= mnt_table_next_fs(table
, iter
, &fs
);
1282 return log_debug_errno(r
, "Failed to get next entry from /proc/self/mountinfo: %m");
1284 path
= mnt_fs_get_target(fs
);
1288 if (isempty(path_startswith(path
, prefix
)))
1291 id1
= mnt_fs_get_id(fs
);
1292 r
= path_get_mnt_id(path
, &id2
);
1294 log_debug_errno(r
, "Failed to get mount ID of '%s', ignoring: %m", path
);
1298 /* The path may be hidden by another over-mount or already remounted. */
1299 log_debug("The mount IDs of '%s' obtained by libmount and path_get_mnt_id() are different (%i vs %i), ignoring.",
1304 mount_fd
= open_tree(AT_FDCWD
, path
, OPEN_TREE_CLONE
| OPEN_TREE_CLOEXEC
| AT_RECURSIVE
);
1306 if (errno
== ENOENT
) /* The path may be hidden by another over-mount or already unmounted. */
1309 return log_debug_errno(errno
, "Failed to open tree of mounted filesystem '%s': %m", path
);
1314 return log_oom_debug();
1316 if (!GREEDY_REALLOC(mounts
, n
+ 1))
1317 return log_oom_debug();
1319 mounts
[n
++] = (SubMount
) {
1320 .path
= TAKE_PTR(p
),
1321 .mount_fd
= TAKE_FD(mount_fd
),
1325 typesafe_qsort(mounts
, n
, sub_mount_compare
);
1326 sub_mount_drop(mounts
, n
);
1328 *ret_mounts
= TAKE_PTR(mounts
);
1333 static int move_sub_mounts(SubMount
*mounts
, size_t n
) {
1334 assert(mounts
|| n
== 0);
1336 for (size_t i
= 0; i
< n
; i
++) {
1337 if (!mounts
[i
].path
|| mounts
[i
].mount_fd
< 0)
1340 (void) mkdir_p_label(mounts
[i
].path
, 0755);
1342 if (move_mount(mounts
[i
].mount_fd
, "", AT_FDCWD
, mounts
[i
].path
, MOVE_MOUNT_F_EMPTY_PATH
) < 0)
1343 return log_debug_errno(errno
, "Failed to move mount_fd to '%s': %m", mounts
[i
].path
);
1349 int remount_and_move_sub_mounts(
1353 unsigned long flags
,
1354 const char *options
) {
1356 SubMount
*mounts
= NULL
; /* avoid false maybe-uninitialized warning */
1357 size_t n
= 0; /* avoid false maybe-uninitialized warning */
1360 CLEANUP_ARRAY(mounts
, n
, sub_mount_array_free
);
1364 /* This is useful when creating a new network namespace. Unlike procfs, we need to remount sysfs,
1365 * otherwise properties of the network interfaces in the main network namespace are still accessible
1366 * through the old sysfs, e.g. /sys/class/net/eth0. All sub-mounts previously mounted on the sysfs
1367 * are moved onto the new sysfs mount. */
1369 r
= path_is_mount_point(where
, NULL
, 0);
1371 return log_debug_errno(r
, "Failed to determine if '%s' is a mountpoint: %m", where
);
1373 /* Shortcut. Simply mount the requested filesystem. */
1374 return mount_nofollow_verbose(LOG_DEBUG
, what
, where
, type
, flags
, options
);
1376 /* Get the list of sub-mounts and duplicate them. */
1377 r
= get_sub_mounts(where
, &mounts
, &n
);
1381 /* Then, remount the mount and its sub-mounts. */
1382 (void) umount_recursive(where
, 0);
1384 /* Remount the target filesystem. */
1385 r
= mount_nofollow_verbose(LOG_DEBUG
, what
, where
, type
, flags
, options
);
1389 /* Finally, move the all sub-mounts on the new target mount point. */
1390 return move_sub_mounts(mounts
, n
);
1393 int remount_sysfs(const char *where
) {
1394 return remount_and_move_sub_mounts("sysfs", where
, "sysfs", MS_NOSUID
|MS_NOEXEC
|MS_NODEV
, NULL
);
1397 int make_mount_point_inode_from_stat(const struct stat
*st
, const char *dest
, mode_t mode
) {
1401 if (S_ISDIR(st
->st_mode
))
1402 return mkdir_label(dest
, mode
);
1404 return RET_NERRNO(mknod(dest
, S_IFREG
|(mode
& ~0111), 0));
1407 int make_mount_point_inode_from_path(const char *source
, const char *dest
, mode_t mode
) {
1413 if (stat(source
, &st
) < 0)
1416 return make_mount_point_inode_from_stat(&st
, dest
, mode
);