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         /* 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.  But on
  35          * the switch from initrd we do use MS_REC, as it is expected that mounts set up in /run/ are
  36          * maintained. */
  37         static const struct {
  38                 const char *path;
  39                 unsigned long mount_flags;                 /* Flags to apply if SWITCH_ROOT_RECURSIVE_RUN is unset */
  40                 unsigned long mount_flags_recursive_run;   /* Flags to apply if SWITCH_ROOT_RECURSIVE_RUN is set (0 if shall be skipped) */
  41         } transfer_table[] = {
  42                 { "/dev",                                 MS_BIND|MS_REC,  MS_BIND|MS_REC }, /* Recursive, because we want to save the original /dev/shm/ + /dev/pts/ and similar */
  43                 { "/sys",                                 MS_BIND|MS_REC,  MS_BIND|MS_REC }, /* Similar, we want to retain various API VFS, or the cgroupv1 /sys/fs/cgroup/ tree */
  44                 { "/proc",                                MS_BIND|MS_REC,  MS_BIND|MS_REC }, /* Similar */
  45                 { "/run",                                 MS_BIND,         MS_BIND|MS_REC }, /* Recursive except on soft reboot, see above */
  46                 { SYSTEM_CREDENTIALS_DIRECTORY,           MS_BIND,         0 /* skip! */  }, /* Credentials passed into the system should survive */
  47                 { ENCRYPTED_SYSTEM_CREDENTIALS_DIRECTORY, MS_BIND,         0 /* skip! */  },  /* Similar */
  48                 { "/run/host",                            MS_BIND|MS_REC,  0 /* skip! */  },  /* Host supplied hierarchy should also survive */
  49         };
  50
  51         _cleanup_close_ int old_root_fd = -EBADF, new_root_fd = -EBADF;
  52         _cleanup_free_ char *resolved_old_root_after = NULL;
  53         int r, istmp;
  54
  55         assert(new_root);
  56
  57         /* Check if we shall remove the contents of the old root */
  58         old_root_fd = open("/", O_DIRECTORY|O_CLOEXEC);
  59         if (old_root_fd < 0)
  60                 return log_error_errno(errno, "Failed to open root directory: %m");
  61
  62         new_root_fd = open(new_root, O_DIRECTORY|O_CLOEXEC);
  63         if (new_root_fd < 0)
  64                 return log_error_errno(errno, "Failed to open target directory '%s': %m", new_root);
  65
  66         r = fds_are_same_mount(old_root_fd, new_root_fd);
  67         if (r < 0)
  68                 return log_error_errno(r, "Failed to check if old and new root directory/mount are the same: %m");
  69         if (r > 0) {
  70                 log_debug("Skipping switch root, as old and new root directories/mounts are the same.");
  71                 return 0;
  72         }
  73
  74         /* Make the new root directory a mount point if it isn't */
  75         r = fd_make_mount_point(new_root_fd);
  76         if (r < 0)
  77                 return log_error_errno(r, "Failed to make new root directory a mount point: %m");
  78         if (r > 0) {
  79                 int fd;
  80
  81                 /* When the path was not a mount point, then we need to reopen the path, otherwise, it still
  82                  * points to the underlying directory. */
  83
  84                 fd = open(new_root, O_DIRECTORY|O_CLOEXEC);
  85                 if (fd < 0)
  86                         return log_error_errno(errno, "Failed to reopen target directory '%s': %m", new_root);
  87
  88                 close_and_replace(new_root_fd, fd);
  89         }
  90
  91         if (FLAGS_SET(flags, SWITCH_ROOT_DESTROY_OLD_ROOT)) {
  92                 istmp = fd_is_temporary_fs(old_root_fd);
  93                 if (istmp < 0)
  94                         return log_error_errno(istmp, "Failed to stat root directory: %m");
  95                 if (istmp > 0)
  96                         log_debug("Root directory is on tmpfs, will do cleanup later.");
  97         } else
  98                 istmp = -1; /* don't know */
  99
 100         if (old_root_after) {
 101                 /* Determine where we shall place the old root after the transition */
 102                 r = chase(old_root_after, new_root, CHASE_PREFIX_ROOT|CHASE_NONEXISTENT, &resolved_old_root_after, NULL);
 103                 if (r < 0)
 104                         return log_error_errno(r, "Failed to resolve %s/%s: %m", new_root, old_root_after);
 105                 if (r == 0) /* Doesn't exist yet. Let's create it */
 106                         (void) mkdir_p_label(resolved_old_root_after, 0755);
 107         }
 108
 109         /* We are about to unmount various file systems with MNT_DETACH (either explicitly via umount() or
 110          * indirectly via pivot_root()), and thus do not synchronously wait for them to be fully sync'ed —
 111          * all while making them invisible/inaccessible in the file system tree for later code. That makes
 112          * sync'ing them then difficult. Let's hence issue a manual sync() here, so that we at least can
 113          * guarantee all file systems are an a good state before entering this state. */
 114         if (!FLAGS_SET(flags, SWITCH_ROOT_DONT_SYNC))
 115                 sync();
 116
 117         /* Work-around for kernel design: the kernel refuses MS_MOVE if any file systems are mounted
 118          * MS_SHARED. Hence remount them MS_PRIVATE here as a work-around.
 119          *
 120          * https://bugzilla.redhat.com/show_bug.cgi?id=847418 */
 121         if (mount(NULL, "/", NULL, MS_REC|MS_PRIVATE, NULL) < 0)
 122                 return log_error_errno(errno, "Failed to set \"/\" mount propagation to private: %m");
 123
 124         /* Do not fail if base_filesystem_create() fails. Not all switch roots are like base_filesystem_create() wants
 125          * them to look like. They might even boot, if they are RO and don't have the FS layout. Just ignore the error
 126          * and switch_root() nevertheless. */
 127         (void) base_filesystem_create_fd(new_root_fd, new_root, UID_INVALID, GID_INVALID);
 128
 129         FOREACH_ARRAY(transfer, transfer_table, ELEMENTSOF(transfer_table)) {
 130                 _cleanup_free_ char *chased = NULL;
 131                 unsigned long mount_flags;
 132
 133                 mount_flags = FLAGS_SET(flags, SWITCH_ROOT_RECURSIVE_RUN) ? transfer->mount_flags_recursive_run : transfer->mount_flags;
 134                 if (mount_flags == 0) /* skip if zero */
 135                         continue;
 136
 137                 if (access(transfer->path, F_OK) < 0) {
 138                         log_debug_errno(errno, "Path '%s' to move to target root directory, not found, ignoring: %m", transfer->path);
 139                         continue;
 140                 }
 141
 142                 r = chase(transfer->path, new_root, CHASE_PREFIX_ROOT, &chased, NULL);
 143                 if (r < 0)
 144                         return log_error_errno(r, "Failed to resolve %s/%s: %m", new_root, transfer->path);
 145
 146                 /* Let's see if it is a mount point already. */
 147                 r = path_is_mount_point(chased, NULL, 0);
 148                 if (r < 0)
 149                         return log_error_errno(r, "Failed to determine whether %s is a mount point: %m", chased);
 150                 if (r > 0) /* If it is already mounted, then do nothing */
 151                         continue;
 152
 153                 r = mount_nofollow_verbose(LOG_ERR, transfer->path, chased, NULL, mount_flags, NULL);
 154                 if (r < 0)
 155                         return r;
 156         }
 157
 158         if (fchdir(new_root_fd) < 0)
 159                 return log_error_errno(errno, "Failed to change directory to %s: %m", new_root);
 160
 161         /* We first try a pivot_root() so that we can umount the old root dir. In many cases (i.e. where rootfs is /),
 162          * that's not possible however, and hence we simply overmount root */
 163         if (resolved_old_root_after)
 164                 r = RET_NERRNO(pivot_root(".", resolved_old_root_after));
 165         else {
 166                 r = RET_NERRNO(pivot_root(".", "."));
 167                 if (r >= 0) {
 168                         /* Now unmount the upper of the two stacked file systems */
 169                         if (umount2(".", MNT_DETACH) < 0)
 170                                 return log_error_errno(errno, "Failed to unmount the old root: %m");
 171                 }
 172         }
 173         if (r < 0) {
 174                 log_debug_errno(r, "Pivoting root file system failed, moving mounts instead: %m");
 175
 176                 if (resolved_old_root_after) {
 177                         r = mount_nofollow_verbose(LOG_ERR, "/", resolved_old_root_after, NULL, MS_BIND|MS_REC, NULL);
 178                         if (r < 0)
 179                                 return r;
 180                 }
 181
 182                 /* If we have to use MS_MOVE let's first try to get rid of *all* mounts we can, with the
 183                  * exception of the path we want to switch to, plus everything leading to it and within
 184                  * it. This is necessary because unlike pivot_root() just moving the mount to the root via
 185                  * MS_MOVE won't magically unmount anything below it. Once the chroot() succeeds the mounts
 186                  * below would still be around but invisible to us, because not accessible via
 187                  * /proc/self/mountinfo. Hence, let's clean everything up first, as long as we still can. */
 188                 (void) umount_recursive_full(NULL, MNT_DETACH, STRV_MAKE(new_root));
 189
 190                 if (mount(".", "/", NULL, MS_MOVE, NULL) < 0)
 191                         return log_error_errno(errno, "Failed to move %s to /: %m", new_root);
 192
 193                 if (chroot(".") < 0)
 194                         return log_error_errno(errno, "Failed to change root: %m");
 195
 196                 if (chdir(".") < 0)
 197                         return log_error_errno(errno, "Failed to change directory: %m");
 198         }
 199
 200         if (istmp > 0) {
 201                 struct stat rb;
 202
 203                 if (fstat(old_root_fd, &rb) < 0)
 204                         return log_error_errno(errno, "Failed to stat old root directory: %m");
 205
 206                 /* Note: the below won't operate on non-memory file systems (i.e. only on tmpfs, ramfs), and
 207                  * it will stop at mount boundaries */
 208                 (void) rm_rf_children(TAKE_FD(old_root_fd), 0, &rb); /* takes possession of the dir fd, even on failure */
 209         }
 210
 211         return 0;
 212 }