src/shared/switch-root.c

   1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
   2
   3 #include <errno.h>
   4 #include <fcntl.h>
   5 #include <limits.h>
   6 #include <stdbool.h>
   7 #include <sys/mount.h>
   8 #include <sys/stat.h>
   9 #include <unistd.h>
  10
  11 #include "base-filesystem.h"
  12 #include "chase.h"
  13 #include "creds-util.h"
  14 #include "fd-util.h"
  15 #include "initrd-util.h"
  16 #include "log.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"
  22 #include "rm-rf.h"
  23 #include "stdio-util.h"
  24 #include "string-util.h"
  25 #include "strv.h"
  26 #include "switch-root.h"
  27 #include "user-util.h"
  28
  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) {
  32
  33         struct {
  34                 const char *path;
  35                 unsigned long mount_flags;
  36         } transfer_table[] = {
  37                 { "/dev",                                 MS_BIND|MS_REC }, /* Recursive, because we want to save the original /dev/shm + /dev/pts and similar */
  38                 { "/sys",                                 MS_BIND|MS_REC }, /* Similar, we want to retain various API VFS, or the cgroupv1 /sys/fs/cgroup/ tree */
  39                 { "/proc",                                MS_BIND|MS_REC }, /* Similar */
  40                 { "/run",                                 MS_BIND        }, /* Stuff mounted below this we don't save, as it might have lost its relevance, i.e. credentials, removable media and such, we rather want that the new boot mounts this fresh */
  41                 { SYSTEM_CREDENTIALS_DIRECTORY,           MS_BIND        }, /* Credentials passed into the system should survive */
  42                 { ENCRYPTED_SYSTEM_CREDENTIALS_DIRECTORY, MS_BIND        }, /* Similar */
  43                 { "/run/host",                            MS_BIND|MS_REC }, /* Host supplied hierarchy should also survive */
  44         };
  45
  46         _cleanup_close_ int old_root_fd = -EBADF, new_root_fd = -EBADF;
  47         _cleanup_free_ char *resolved_old_root_after = NULL;
  48         int r, istmp;
  49
  50         assert(new_root);
  51
  52         /* Check if we shall remove the contents of the old root */
  53         old_root_fd = open("/", O_DIRECTORY|O_CLOEXEC);
  54         if (old_root_fd < 0)
  55                 return log_error_errno(errno, "Failed to open root directory: %m");
  56
  57         new_root_fd = open(new_root, O_DIRECTORY|O_CLOEXEC);
  58         if (new_root_fd < 0)
  59                 return log_error_errno(errno, "Failed to open target directory '%s': %m", new_root);
  60
  61         r = inode_same_at(old_root_fd, "", new_root_fd, "", AT_EMPTY_PATH);
  62         if (r < 0)
  63                 return log_error_errno(r, "Failed to determine if old and new root directory are the same: %m");
  64         if (r > 0) {
  65                 log_debug("Skipping switch root, as old and new root directory are the same.");
  66                 return 0;
  67         }
  68
  69         /* Make the new root directory a mount point if it isn't */
  70         r = fd_make_mount_point(new_root_fd);
  71         if (r < 0)
  72                 return log_error_errno(r, "Failed to make new root directory a mount point: %m");
  73
  74         if (FLAGS_SET(flags, SWITCH_ROOT_DESTROY_OLD_ROOT)) {
  75                 istmp = fd_is_temporary_fs(old_root_fd);
  76                 if (istmp < 0)
  77                         return log_error_errno(istmp, "Failed to stat root directory: %m");
  78                 if (istmp > 0)
  79                         log_debug("Root directory is on tmpfs, will do cleanup later.");
  80         } else
  81                 istmp = -1; /* don't know */
  82
  83         if (old_root_after) {
  84                 /* Determine where we shall place the old root after the transition */
  85                 r = chase(old_root_after, new_root, CHASE_PREFIX_ROOT|CHASE_NONEXISTENT, &resolved_old_root_after, NULL);
  86                 if (r < 0)
  87                         return log_error_errno(r, "Failed to resolve %s/%s: %m", new_root, old_root_after);
  88                 if (r == 0) /* Doesn't exist yet. Let's create it */
  89                         (void) mkdir_p_label(resolved_old_root_after, 0755);
  90         }
  91
  92         /* We are about to unmount various file systems with MNT_DETACH (either explicitly via umount() or
  93          * indirectly via pivot_root()), and thus do not synchronously wait for them to be fully sync'ed —
  94          * all while making them invisible/inaccessible in the file system tree for later code. That makes
  95          * sync'ing them then difficult. Let's hence issue a manual sync() here, so that we at least can
  96          * guarantee all file systems are an a good state before entering this state. */
  97         if (!FLAGS_SET(flags, SWITCH_ROOT_DONT_SYNC))
  98                 sync();
  99
 100         /* Work-around for kernel design: the kernel refuses MS_MOVE if any file systems are mounted
 101          * MS_SHARED. Hence remount them MS_PRIVATE here as a work-around.
 102          *
 103          * https://bugzilla.redhat.com/show_bug.cgi?id=847418 */
 104         if (mount(NULL, "/", NULL, MS_REC|MS_PRIVATE, NULL) < 0)
 105                 return log_error_errno(errno, "Failed to set \"/\" mount propagation to private: %m");
 106
 107         /* Do not fail if base_filesystem_create() fails. Not all switch roots are like base_filesystem_create() wants
 108          * them to look like. They might even boot, if they are RO and don't have the FS layout. Just ignore the error
 109          * and switch_root() nevertheless. */
 110         (void) base_filesystem_create_fd(new_root_fd, new_root, UID_INVALID, GID_INVALID);
 111
 112         FOREACH_ARRAY(transfer, transfer_table, ELEMENTSOF(transfer_table)) {
 113                 _cleanup_free_ char *chased = NULL;
 114
 115                 if (access(transfer->path, F_OK) < 0) {
 116                         log_debug_errno(errno, "Path '%s' to move to target root directory, not found, ignoring: %m", transfer->path);
 117                         continue;
 118                 }
 119
 120                 r = chase(transfer->path, new_root, CHASE_PREFIX_ROOT, &chased, NULL);
 121                 if (r < 0)
 122                         return log_error_errno(r, "Failed to resolve %s/%s: %m", new_root, transfer->path);
 123
 124                 /* Let's see if it is a mount point already. */
 125                 r = path_is_mount_point(chased, NULL, 0);
 126                 if (r < 0)
 127                         return log_error_errno(r, "Failed to determine whether %s is a mount point: %m", chased);
 128                 if (r > 0) /* If it is already mounted, then do nothing */
 129                         continue;
 130
 131                 r = mount_nofollow_verbose(LOG_ERR, transfer->path, chased, NULL, transfer->mount_flags, NULL);
 132                 if (r < 0)
 133                         return r;
 134         }
 135
 136         if (fchdir(new_root_fd) < 0)
 137                 return log_error_errno(errno, "Failed to change directory to %s: %m", new_root);
 138
 139         /* We first try a pivot_root() so that we can umount the old root dir. In many cases (i.e. where rootfs is /),
 140          * that's not possible however, and hence we simply overmount root */
 141         if (resolved_old_root_after)
 142                 r = RET_NERRNO(pivot_root(".", resolved_old_root_after));
 143         else {
 144                 r = RET_NERRNO(pivot_root(".", "."));
 145                 if (r >= 0) {
 146                         /* Now unmount the upper of the two stacked file systems */
 147                         if (umount2(".", MNT_DETACH) < 0)
 148                                 return log_error_errno(errno, "Failed to unmount the old root: %m");
 149                 }
 150         }
 151         if (r < 0) {
 152                 log_debug_errno(r, "Pivoting root file system failed, moving mounts instead: %m");
 153
 154                 /* If we have to use MS_MOVE let's first try to get rid of *all* mounts we can, with the
 155                  * exception of the path we want to switch to, plus everything leading to it and within
 156                  * it. This is necessary because unlike pivot_root() just moving the mount to the root via
 157                  * MS_MOVE won't magically unmount anything below it. Once the chroot() succeeds the mounts
 158                  * below would still be around but invisible to us, because not accessible via
 159                  * /proc/self/mountinfo. Hence, let's clean everything up first, as long as we still can. */
 160                 (void) umount_recursive_full(NULL, MNT_DETACH, STRV_MAKE(new_root));
 161
 162                 if (mount(".", "/", NULL, MS_MOVE, NULL) < 0)
 163                         return log_error_errno(errno, "Failed to move %s to /: %m", new_root);
 164
 165                 if (chroot(".") < 0)
 166                         return log_error_errno(errno, "Failed to change root: %m");
 167
 168                 if (chdir(".") < 0)
 169                         return log_error_errno(errno, "Failed to change directory: %m");
 170         }
 171
 172         if (istmp > 0) {
 173                 struct stat rb;
 174
 175                 if (fstat(old_root_fd, &rb) < 0)
 176                         return log_error_errno(errno, "Failed to stat old root directory: %m");
 177
 178                 /* Note: the below won't operate on non-memory file systems (i.e. only on tmpfs, ramfs), and
 179                  * it will stop at mount boundaries */
 180                 (void) rm_rf_children(TAKE_FD(old_root_fd), 0, &rb); /* takes possession of the dir fd, even on failure */
 181         }
 182
 183         return 0;
 184 }