1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
11 #include "base-filesystem.h"
13 #include "creds-util.h"
15 #include "initrd-util.h"
17 #include "missing_syscall.h"
18 #include "mkdir-label.h"
19 #include "mount-util.h"
20 #include "mountpoint-util.h"
21 #include "path-util.h"
23 #include "stdio-util.h"
24 #include "string-util.h"
26 #include "switch-root.h"
27 #include "user-util.h"
29 int switch_root(const char *new_root
,
30 const char *old_root_after
, /* path below the new root, where to place the old root after the transition; may be NULL to unmount it */
31 SwitchRootFlags flags
) {
33 /* Stuff mounted below /run we don't save on soft reboot, as it might have lost its relevance, i.e.
34 * credentials, removable media and such, we rather want that the new boot mounts this fresh.
35 * But on the switch from initrd we do use MS_REC, as it is expected that mounts set up in /run
37 unsigned long run_mount_flags
= MS_BIND
|(!FLAGS_SET(flags
, SWITCH_ROOT_SKIP_RECURSIVE_RUN
) ? MS_REC
: 0);
40 unsigned long mount_flags
;
41 bool skip_if_run_is_rec
; /* For child mounts of /run, if it's moved recursively no need to handle */
42 } transfer_table
[] = {
43 { "/dev", MS_BIND
|MS_REC
, false }, /* Recursive, because we want to save the original /dev/shm + /dev/pts and similar */
44 { "/sys", MS_BIND
|MS_REC
, false }, /* Similar, we want to retain various API VFS, or the cgroupv1 /sys/fs/cgroup/ tree */
45 { "/proc", MS_BIND
|MS_REC
, false }, /* Similar */
46 { "/run", run_mount_flags
, false }, /* Recursive except on soft reboot, see above */
47 { SYSTEM_CREDENTIALS_DIRECTORY
, MS_BIND
, true }, /* Credentials passed into the system should survive */
48 { ENCRYPTED_SYSTEM_CREDENTIALS_DIRECTORY
, MS_BIND
, true }, /* Similar */
49 { "/run/host", MS_BIND
|MS_REC
, true }, /* Host supplied hierarchy should also survive */
52 _cleanup_close_
int old_root_fd
= -EBADF
, new_root_fd
= -EBADF
;
53 _cleanup_free_
char *resolved_old_root_after
= NULL
;
58 /* Check if we shall remove the contents of the old root */
59 old_root_fd
= open("/", O_DIRECTORY
|O_CLOEXEC
);
61 return log_error_errno(errno
, "Failed to open root directory: %m");
63 new_root_fd
= open(new_root
, O_DIRECTORY
|O_CLOEXEC
);
65 return log_error_errno(errno
, "Failed to open target directory '%s': %m", new_root
);
67 r
= inode_same_at(old_root_fd
, "", new_root_fd
, "", AT_EMPTY_PATH
);
69 return log_error_errno(r
, "Failed to determine if old and new root directory are the same: %m");
71 log_debug("Skipping switch root, as old and new root directory are the same.");
75 /* Make the new root directory a mount point if it isn't */
76 r
= fd_make_mount_point(new_root_fd
);
78 return log_error_errno(r
, "Failed to make new root directory a mount point: %m");
80 if (FLAGS_SET(flags
, SWITCH_ROOT_DESTROY_OLD_ROOT
)) {
81 istmp
= fd_is_temporary_fs(old_root_fd
);
83 return log_error_errno(istmp
, "Failed to stat root directory: %m");
85 log_debug("Root directory is on tmpfs, will do cleanup later.");
87 istmp
= -1; /* don't know */
90 /* Determine where we shall place the old root after the transition */
91 r
= chase(old_root_after
, new_root
, CHASE_PREFIX_ROOT
|CHASE_NONEXISTENT
, &resolved_old_root_after
, NULL
);
93 return log_error_errno(r
, "Failed to resolve %s/%s: %m", new_root
, old_root_after
);
94 if (r
== 0) /* Doesn't exist yet. Let's create it */
95 (void) mkdir_p_label(resolved_old_root_after
, 0755);
98 /* We are about to unmount various file systems with MNT_DETACH (either explicitly via umount() or
99 * indirectly via pivot_root()), and thus do not synchronously wait for them to be fully sync'ed —
100 * all while making them invisible/inaccessible in the file system tree for later code. That makes
101 * sync'ing them then difficult. Let's hence issue a manual sync() here, so that we at least can
102 * guarantee all file systems are an a good state before entering this state. */
103 if (!FLAGS_SET(flags
, SWITCH_ROOT_DONT_SYNC
))
106 /* Work-around for kernel design: the kernel refuses MS_MOVE if any file systems are mounted
107 * MS_SHARED. Hence remount them MS_PRIVATE here as a work-around.
109 * https://bugzilla.redhat.com/show_bug.cgi?id=847418 */
110 if (mount(NULL
, "/", NULL
, MS_REC
|MS_PRIVATE
, NULL
) < 0)
111 return log_error_errno(errno
, "Failed to set \"/\" mount propagation to private: %m");
113 /* Do not fail if base_filesystem_create() fails. Not all switch roots are like base_filesystem_create() wants
114 * them to look like. They might even boot, if they are RO and don't have the FS layout. Just ignore the error
115 * and switch_root() nevertheless. */
116 (void) base_filesystem_create_fd(new_root_fd
, new_root
, UID_INVALID
, GID_INVALID
);
118 FOREACH_ARRAY(transfer
, transfer_table
, ELEMENTSOF(transfer_table
)) {
119 _cleanup_free_
char *chased
= NULL
;
121 if (transfer
->skip_if_run_is_rec
&& !FLAGS_SET(flags
, SWITCH_ROOT_SKIP_RECURSIVE_RUN
))
124 if (access(transfer
->path
, F_OK
) < 0) {
125 log_debug_errno(errno
, "Path '%s' to move to target root directory, not found, ignoring: %m", transfer
->path
);
129 r
= chase(transfer
->path
, new_root
, CHASE_PREFIX_ROOT
, &chased
, NULL
);
131 return log_error_errno(r
, "Failed to resolve %s/%s: %m", new_root
, transfer
->path
);
133 /* Let's see if it is a mount point already. */
134 r
= path_is_mount_point(chased
, NULL
, 0);
136 return log_error_errno(r
, "Failed to determine whether %s is a mount point: %m", chased
);
137 if (r
> 0) /* If it is already mounted, then do nothing */
140 r
= mount_nofollow_verbose(LOG_ERR
, transfer
->path
, chased
, NULL
, transfer
->mount_flags
, NULL
);
145 if (fchdir(new_root_fd
) < 0)
146 return log_error_errno(errno
, "Failed to change directory to %s: %m", new_root
);
148 /* We first try a pivot_root() so that we can umount the old root dir. In many cases (i.e. where rootfs is /),
149 * that's not possible however, and hence we simply overmount root */
150 if (resolved_old_root_after
)
151 r
= RET_NERRNO(pivot_root(".", resolved_old_root_after
));
153 r
= RET_NERRNO(pivot_root(".", "."));
155 /* Now unmount the upper of the two stacked file systems */
156 if (umount2(".", MNT_DETACH
) < 0)
157 return log_error_errno(errno
, "Failed to unmount the old root: %m");
161 log_debug_errno(r
, "Pivoting root file system failed, moving mounts instead: %m");
163 /* If we have to use MS_MOVE let's first try to get rid of *all* mounts we can, with the
164 * exception of the path we want to switch to, plus everything leading to it and within
165 * it. This is necessary because unlike pivot_root() just moving the mount to the root via
166 * MS_MOVE won't magically unmount anything below it. Once the chroot() succeeds the mounts
167 * below would still be around but invisible to us, because not accessible via
168 * /proc/self/mountinfo. Hence, let's clean everything up first, as long as we still can. */
169 (void) umount_recursive_full(NULL
, MNT_DETACH
, STRV_MAKE(new_root
));
171 if (mount(".", "/", NULL
, MS_MOVE
, NULL
) < 0)
172 return log_error_errno(errno
, "Failed to move %s to /: %m", new_root
);
175 return log_error_errno(errno
, "Failed to change root: %m");
178 return log_error_errno(errno
, "Failed to change directory: %m");
184 if (fstat(old_root_fd
, &rb
) < 0)
185 return log_error_errno(errno
, "Failed to stat old root directory: %m");
187 /* Note: the below won't operate on non-memory file systems (i.e. only on tmpfs, ramfs), and
188 * it will stop at mount boundaries */
189 (void) rm_rf_children(TAKE_FD(old_root_fd
), 0, &rb
); /* takes possession of the dir fd, even on failure */