]>
git.ipfire.org Git - thirdparty/systemd.git/blob - src/shared/switch-root.c
1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
11 #include "base-filesystem.h"
14 #include "initrd-util.h"
16 #include "missing_syscall.h"
17 #include "mkdir-label.h"
18 #include "mount-util.h"
19 #include "mountpoint-util.h"
20 #include "path-util.h"
22 #include "stdio-util.h"
23 #include "string-util.h"
25 #include "switch-root.h"
26 #include "user-util.h"
28 int switch_root(const char *new_root
,
29 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 */
30 unsigned long mount_flags
) { /* MS_MOVE or MS_BIND used for /proc/, /dev/, /run/, /sys/ */
32 _cleanup_close_
int old_root_fd
= -EBADF
, new_root_fd
= -EBADF
;
33 _cleanup_free_
char *resolved_old_root_after
= NULL
;
37 assert(IN_SET(mount_flags
, MS_MOVE
, MS_BIND
));
39 if (path_equal(new_root
, "/"))
42 /* Check if we shall remove the contents of the old root */
43 old_root_fd
= open("/", O_DIRECTORY
|O_CLOEXEC
);
45 return log_error_errno(errno
, "Failed to open root directory: %m");
47 istmp
= fd_is_temporary_fs(old_root_fd
);
49 return log_error_errno(istmp
, "Failed to stat root directory: %m");
51 log_debug("Root directory is on tmpfs, will do cleanup later.");
53 new_root_fd
= open(new_root
, O_DIRECTORY
|O_CLOEXEC
);
55 return log_error_errno(errno
, "Failed to open target directory '%s': %m", new_root
);
58 /* Determine where we shall place the old root after the transition */
59 r
= chase(old_root_after
, new_root
, CHASE_PREFIX_ROOT
|CHASE_NONEXISTENT
, &resolved_old_root_after
, NULL
);
61 return log_error_errno(r
, "Failed to resolve %s/%s: %m", new_root
, old_root_after
);
62 if (r
== 0) /* Doesn't exist yet. Let's create it */
63 (void) mkdir_p_label(resolved_old_root_after
, 0755);
66 /* We are about to unmount various file systems with MNT_DETACH (either explicitly via umount() or
67 * indirectly via pivot_root()), and thus do not synchronously wait for them to be fully sync'ed —
68 * all while making them invisible/inaccessible in the file system tree for later code. That makes
69 * sync'ing them then difficult. Let's hence issue a manual sync() here, so that we at least can
70 * guarantee all file systems are an a good state before entering this state. */
73 /* Work-around for kernel design: the kernel refuses MS_MOVE if any file systems are mounted
74 * MS_SHARED. Hence remount them MS_PRIVATE here as a work-around.
76 * https://bugzilla.redhat.com/show_bug.cgi?id=847418 */
77 if (mount(NULL
, "/", NULL
, MS_REC
|MS_PRIVATE
, NULL
) < 0)
78 return log_error_errno(errno
, "Failed to set \"/\" mount propagation to private: %m");
80 FOREACH_STRING(path
, "/sys", "/dev", "/run", "/proc") {
81 _cleanup_free_
char *chased
= NULL
;
83 r
= chase(path
, new_root
, CHASE_PREFIX_ROOT
|CHASE_NONEXISTENT
, &chased
, NULL
);
85 return log_error_errno(r
, "Failed to resolve %s/%s: %m", new_root
, path
);
87 /* Already exists. Let's see if it is a mount point already. */
88 r
= path_is_mount_point(chased
, NULL
, 0);
90 return log_error_errno(r
, "Failed to determine whether %s is a mount point: %m", chased
);
91 if (r
> 0) /* If it is already mounted, then do nothing */
94 /* Doesn't exist yet? */
95 (void) mkdir_p_label(chased
, 0755);
97 if (mount(path
, chased
, NULL
, mount_flags
, NULL
) < 0)
98 return log_error_errno(errno
, "Failed to mount %s to %s: %m", path
, chased
);
101 /* Do not fail if base_filesystem_create() fails. Not all switch roots are like base_filesystem_create() wants
102 * them to look like. They might even boot, if they are RO and don't have the FS layout. Just ignore the error
103 * and switch_root() nevertheless. */
104 (void) base_filesystem_create_fd(new_root_fd
, new_root
, UID_INVALID
, GID_INVALID
);
106 if (fchdir(new_root_fd
) < 0)
107 return log_error_errno(errno
, "Failed to change directory to %s: %m", new_root
);
109 /* We first try a pivot_root() so that we can umount the old root dir. In many cases (i.e. where rootfs is /),
110 * that's not possible however, and hence we simply overmount root */
111 if (resolved_old_root_after
)
112 r
= RET_NERRNO(pivot_root(".", resolved_old_root_after
));
114 r
= RET_NERRNO(pivot_root(".", "."));
116 /* Now unmount the upper of the two stacked file systems */
117 if (umount2(".", MNT_DETACH
) < 0)
118 return log_error_errno(errno
, "Failed to unmount the old root: %m");
122 log_debug_errno(r
, "Pivoting root file system failed, moving mounts instead: %m");
124 /* If we have to use MS_MOVE let's first try to get rid of *all* mounts we can, with the
125 * exception of the path we want to switch to, plus everything leading to it and within
126 * it. This is necessary because unlike pivot_root() just moving the mount to the root via
127 * MS_MOVE won't magically unmount anything below it. Once the chroot() succeeds the mounts
128 * below would still be around but invisible to us, because not accessible via
129 * /proc/self/mountinfo. Hence, let's clean everything up first, as long as we still can. */
130 (void) umount_recursive_full(NULL
, MNT_DETACH
, STRV_MAKE(new_root
));
132 if (mount(".", "/", NULL
, MS_MOVE
, NULL
) < 0)
133 return log_error_errno(errno
, "Failed to move %s to /: %m", new_root
);
136 return log_error_errno(errno
, "Failed to change root: %m");
139 return log_error_errno(errno
, "Failed to change directory: %m");
145 if (fstat(old_root_fd
, &rb
) < 0)
146 return log_error_errno(errno
, "Failed to stat old root directory: %m");
148 /* Note: the below won't operate on non-memory file systems (i.e. only on tmpfs, ramfs), and
149 * it will stop at mount boundaries */
150 (void) rm_rf_children(TAKE_FD(old_root_fd
), 0, &rb
); /* takes possession of the dir fd, even on failure */