blob: 6034485d7424873d06044c86ecc94d9c3ca57468
| 1 | /* vi: set sw=4 ts=4: */ |
| 2 | /* Copyright 2005 Rob Landley <rob@landley.net> |
| 3 | * |
| 4 | * Switch from rootfs to another filesystem as the root of the mount tree. |
| 5 | * |
| 6 | * Licensed under GPLv2, see file LICENSE in this source tree. |
| 7 | */ |
| 8 | //config:config SWITCH_ROOT |
| 9 | //config: bool "switch_root" |
| 10 | //config: default y |
| 11 | //config: select PLATFORM_LINUX |
| 12 | //config: help |
| 13 | //config: The switch_root utility is used from initramfs to select a new |
| 14 | //config: root device. Under initramfs, you have to use this instead of |
| 15 | //config: pivot_root. (Stop reading here if you don't care why.) |
| 16 | //config: |
| 17 | //config: Booting with initramfs extracts a gzipped cpio archive into rootfs |
| 18 | //config: (which is a variant of ramfs/tmpfs). Because rootfs can't be moved |
| 19 | //config: or unmounted*, pivot_root will not work from initramfs. Instead, |
| 20 | //config: switch_root deletes everything out of rootfs (including itself), |
| 21 | //config: does a mount --move that overmounts rootfs with the new root, and |
| 22 | //config: then execs the specified init program. |
| 23 | //config: |
| 24 | //config: * Because the Linux kernel uses rootfs internally as the starting |
| 25 | //config: and ending point for searching through the kernel's doubly linked |
| 26 | //config: list of active mount points. That's why. |
| 27 | |
| 28 | //applet:IF_SWITCH_ROOT(APPLET(switch_root, BB_DIR_SBIN, BB_SUID_DROP)) |
| 29 | |
| 30 | //kbuild:lib-$(CONFIG_SWITCH_ROOT) += switch_root.o |
| 31 | |
| 32 | //usage:#define switch_root_trivial_usage |
| 33 | //usage: "[-c /dev/console] NEW_ROOT NEW_INIT [ARGS]" |
| 34 | //usage:#define switch_root_full_usage "\n\n" |
| 35 | //usage: "Free initramfs and switch to another root fs:\n" |
| 36 | //usage: "chroot to NEW_ROOT, delete all in /, move NEW_ROOT to /,\n" |
| 37 | //usage: "execute NEW_INIT. PID must be 1. NEW_ROOT must be a mountpoint.\n" |
| 38 | //usage: "\n -c DEV Reopen stdio to DEV after switch" |
| 39 | |
| 40 | #include <sys/vfs.h> |
| 41 | #include <sys/mount.h> |
| 42 | #include "libbb.h" |
| 43 | // Make up for header deficiencies |
| 44 | #ifndef RAMFS_MAGIC |
| 45 | # define RAMFS_MAGIC ((unsigned)0x858458f6) |
| 46 | #endif |
| 47 | #ifndef TMPFS_MAGIC |
| 48 | # define TMPFS_MAGIC ((unsigned)0x01021994) |
| 49 | #endif |
| 50 | #ifndef MS_MOVE |
| 51 | # define MS_MOVE 8192 |
| 52 | #endif |
| 53 | |
| 54 | // Recursively delete contents of rootfs |
| 55 | static void delete_contents(const char *directory, dev_t rootdev) |
| 56 | { |
| 57 | DIR *dir; |
| 58 | struct dirent *d; |
| 59 | struct stat st; |
| 60 | |
| 61 | // Don't descend into other filesystems |
| 62 | if (lstat(directory, &st) || st.st_dev != rootdev) |
| 63 | return; |
| 64 | |
| 65 | // Recursively delete the contents of directories |
| 66 | if (S_ISDIR(st.st_mode)) { |
| 67 | dir = opendir(directory); |
| 68 | if (dir) { |
| 69 | while ((d = readdir(dir))) { |
| 70 | char *newdir = d->d_name; |
| 71 | |
| 72 | // Skip . and .. |
| 73 | if (DOT_OR_DOTDOT(newdir)) |
| 74 | continue; |
| 75 | |
| 76 | // Recurse to delete contents |
| 77 | newdir = concat_path_file(directory, newdir); |
| 78 | delete_contents(newdir, rootdev); |
| 79 | free(newdir); |
| 80 | } |
| 81 | closedir(dir); |
| 82 | |
| 83 | // Directory should now be empty, zap it |
| 84 | rmdir(directory); |
| 85 | } |
| 86 | } else { |
| 87 | // It wasn't a directory, zap it |
| 88 | unlink(directory); |
| 89 | } |
| 90 | } |
| 91 | |
| 92 | int switch_root_main(int argc, char **argv) MAIN_EXTERNALLY_VISIBLE; |
| 93 | int switch_root_main(int argc UNUSED_PARAM, char **argv) |
| 94 | { |
| 95 | char *newroot, *console = NULL; |
| 96 | struct stat st; |
| 97 | struct statfs stfs; |
| 98 | dev_t rootdev; |
| 99 | |
| 100 | // Parse args (-c console) |
| 101 | opt_complementary = "-2"; // minimum 2 params |
| 102 | getopt32(argv, "+c:", &console); // '+': stop at first non-option |
| 103 | argv += optind; |
| 104 | newroot = *argv++; |
| 105 | |
| 106 | // Change to new root directory and verify it's a different fs |
| 107 | xchdir(newroot); |
| 108 | xstat("/", &st); |
| 109 | rootdev = st.st_dev; |
| 110 | xstat(".", &st); |
| 111 | if (st.st_dev == rootdev || getpid() != 1) { |
| 112 | // Show usage, it says new root must be a mountpoint |
| 113 | // and we must be PID 1 |
| 114 | bb_show_usage(); |
| 115 | } |
| 116 | |
| 117 | // Additional sanity checks: we're about to rm -rf /, so be REALLY SURE |
| 118 | // we mean it. I could make this a CONFIG option, but I would get email |
| 119 | // from all the people who WILL destroy their filesystems. |
| 120 | if (stat("/init", &st) != 0 || !S_ISREG(st.st_mode)) { |
| 121 | bb_error_msg_and_die("/init is not a regular file"); |
| 122 | } |
| 123 | statfs("/", &stfs); // this never fails |
| 124 | if ((unsigned)stfs.f_type != RAMFS_MAGIC |
| 125 | && (unsigned)stfs.f_type != TMPFS_MAGIC |
| 126 | ) { |
| 127 | bb_error_msg_and_die("root filesystem is not ramfs/tmpfs"); |
| 128 | } |
| 129 | |
| 130 | // Zap everything out of rootdev |
| 131 | delete_contents("/", rootdev); |
| 132 | |
| 133 | // Overmount / with newdir and chroot into it |
| 134 | if (mount(".", "/", NULL, MS_MOVE, NULL)) { |
| 135 | // For example, fails when newroot is not a mountpoint |
| 136 | bb_perror_msg_and_die("error moving root"); |
| 137 | } |
| 138 | xchroot("."); |
| 139 | // The chdir is needed to recalculate "." and ".." links |
| 140 | /*xchdir("/"); - done in xchroot */ |
| 141 | |
| 142 | // If a new console specified, redirect stdin/stdout/stderr to it |
| 143 | if (console) { |
| 144 | close(0); |
| 145 | xopen(console, O_RDWR); |
| 146 | xdup2(0, 1); |
| 147 | xdup2(0, 2); |
| 148 | } |
| 149 | |
| 150 | // Exec real init |
| 151 | execv(argv[0], argv); |
| 152 | bb_perror_msg_and_die("can't execute '%s'", argv[0]); |
| 153 | } |
| 154 | |
| 155 | /* |
| 156 | From: Rob Landley <rob@landley.net> |
| 157 | Date: Tue, Jun 16, 2009 at 7:47 PM |
| 158 | Subject: Re: switch_root... |
| 159 | |
| 160 | ... |
| 161 | ... |
| 162 | ... |
| 163 | |
| 164 | If you're _not_ running out of init_ramfs (if for example you're using initrd |
| 165 | instead), you probably shouldn't use switch_root because it's the wrong tool. |
| 166 | |
| 167 | Basically what the sucker does is something like the following shell script: |
| 168 | |
| 169 | find / -xdev | xargs rm -rf |
| 170 | cd "$1" |
| 171 | shift |
| 172 | mount --move . / |
| 173 | exec chroot . "$@" |
| 174 | |
| 175 | There are a couple reasons that won't work as a shell script: |
| 176 | |
| 177 | 1) If you delete the commands out of your $PATH, your shell scripts can't run |
| 178 | more commands, but you can't start using dynamically linked _new_ commands |
| 179 | until after you do the chroot because the path to the dynamic linker is wrong. |
| 180 | So there's a step that needs to be sort of atomic but can't be as a shell |
| 181 | script. (You can work around this with static linking or very carefully laid |
| 182 | out paths and sequencing, but it's brittle, ugly, and non-obvious.) |
| 183 | |
| 184 | 2) The "find | rm" bit will acually delete everything because the mount points |
| 185 | still show up (even if their contents don't), and rm -rf will then happily zap |
| 186 | that. So the first line is an oversimplification of what you need to do _not_ |
| 187 | to descend into other filesystems and delete their contents. |
| 188 | |
| 189 | The reason we do this is to free up memory, by the way. Since initramfs is a |
| 190 | ramfs, deleting its contents frees up the memory it uses. (We leave it with |
| 191 | one remaining dentry for the new mount point, but that's ok.) |
| 192 | |
| 193 | Note that you cannot ever umount rootfs, for approximately the same reason you |
| 194 | can't kill PID 1. The kernel tracks mount points as a doubly linked list, and |
| 195 | the pointer to the start/end of that list always points to an entry that's |
| 196 | known to be there (rootfs), so it never has to worry about moving that pointer |
| 197 | and it never has to worry about the list being empty. (Back around 2.6.13 |
| 198 | there _was_ a bug that let you umount rootfs, and the system locked hard the |
| 199 | instant you did so endlessly looping to find the end of the mount list and |
| 200 | never stopping. They fixed it.) |
| 201 | |
| 202 | Oh, and the reason we mount --move _and_ do the chroot is due to the way "/" |
| 203 | works. Each process has two special symlinks, ".", and "/". Each of them |
| 204 | points to the dentry of a directory, and give you a location paths can start |
| 205 | from. (Historically ".." was also special, because you could enter a |
| 206 | directory via a symlink so backing out to the directory you came from doesn't |
| 207 | necessarily mean the one physically above where "." points to. These days I |
| 208 | think it's just handed off to the filesystem.) |
| 209 | |
| 210 | Anyway, path resolution starts with "." or "/" (although the "./" at the start |
| 211 | of the path may be implicit), meaning it's relative to one of those two |
| 212 | directories. Your current directory, and your current root directory. The |
| 213 | chdir() syscall changes where "." points to, and the chroot() syscall changes |
| 214 | where "/" points to. (Again, both are per-process which is why chroot only |
| 215 | affects your current process and its child processes.) |
| 216 | |
| 217 | Note that chroot() does _not_ change where "." points to, and back before they |
| 218 | put crazy security checks into the kernel your current directory could be |
| 219 | somewhere you could no longer access after the chroot. (The command line |
| 220 | chroot does a cd as well, the chroot _syscall_ is what I'm talking about.) |
| 221 | |
| 222 | The reason mounting something new over / has no obvious effect is the same |
| 223 | reason mounting something over your current directory has no obvious effect: |
| 224 | the . and / links aren't recalculated after a mount, so they still point to |
| 225 | the same dentry they did before, even if that dentry is no longer accessible |
| 226 | by other means. Note that "cd ." is a NOP, and "chroot /" is a nop; both look |
| 227 | up the cached dentry and set it right back. They don't re-parse any paths, |
| 228 | because they're what all paths your process uses would be relative to. |
| 229 | |
| 230 | That's why the careful sequencing above: we cd into the new mount point before |
| 231 | we do the mount --move. Moving the mount point would otherwise make it |
| 232 | totally inaccessible to us because cd-ing to the old path wouldn't give it to |
| 233 | us anymore, and cd "/" just gives us the cached dentry from when the process |
| 234 | was created (in this case the old initramfs one). But the "." symlink gives |
| 235 | us the dentry of the filesystem we just moved, so we can then "chroot ." to |
| 236 | copy that dentry to "/" and get the new filesystem. If we _didn't_ save that |
| 237 | dentry in "." we couldn't get it back after the mount --move. |
| 238 | |
| 239 | (Yes, this is all screwy and I had to email questions to Linus Torvalds to get |
| 240 | it straight myself. I keep meaning to write up a "how mount actually works" |
| 241 | document someday...) |
| 242 | */ |
| 243 |