1 /* vi: set sw=4 ts=4: */
2 /* Copyright 2005 Rob Landley <rob@landley.net>
4 * Switch from rootfs to another filesystem as the root of the mount tree.
6 * Licensed under GPLv2, see file LICENSE in this source tree.
8 //config:config SWITCH_ROOT
9 //config: bool "switch_root (5.2 kb)"
11 //config: select PLATFORM_LINUX
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.)
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.
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.
28 //applet:IF_SWITCH_ROOT(APPLET(switch_root, BB_DIR_SBIN, BB_SUID_DROP))
29 // APPLET_ODDNAME:name main location suid_type help
30 //applet:IF_RUN_INIT( APPLET_ODDNAME(run-init, switch_root, BB_DIR_SBIN, BB_SUID_DROP, run_init))
32 //kbuild:lib-$(CONFIG_SWITCH_ROOT) += switch_root.o
33 //kbuild:lib-$(CONFIG_RUN_INIT) += switch_root.o
35 //usage:#define switch_root_trivial_usage
36 //usage: "[-c CONSOLE_DEV] NEW_ROOT NEW_INIT [ARGS]"
37 //usage:#define switch_root_full_usage "\n\n"
38 //usage: "Free initramfs and switch to another root fs:\n"
39 //usage: "chroot to NEW_ROOT, delete all in /, move NEW_ROOT to /,\n"
40 //usage: "execute NEW_INIT. PID must be 1. NEW_ROOT must be a mountpoint.\n"
41 //usage: "\n -c DEV Reopen stdio to DEV after switch"
43 //usage:#define run_init_trivial_usage
44 //usage: "[-d CAP,CAP...] [-c CONSOLE_DEV] NEW_ROOT NEW_INIT [ARGS]"
45 //usage:#define run_init_full_usage "\n\n"
46 //usage: "Free initramfs and switch to another root fs:\n"
47 //usage: "chroot to NEW_ROOT, delete all in /, move NEW_ROOT to /,\n"
48 //usage: "execute NEW_INIT. PID must be 1. NEW_ROOT must be a mountpoint.\n"
49 //usage: "\n -c DEV Reopen stdio to DEV after switch"
50 //usage: "\n -d CAPS Drop capabilities"
53 #include <sys/mount.h>
55 # include <sys/prctl.h>
56 # include <linux/capability.h>
57 // #include <sys/capability.h>
58 // This header is in libcap, but the functions are in libc.
59 // Comment in the header says this above capset/capget:
60 /* system calls - look to libc for function to system call mapping */
61 extern int capset(cap_user_header_t header, cap_user_data_t data);
62 extern int capget(cap_user_header_t header, const cap_user_data_t data);
63 // so for bbox, let's just repeat the declarations.
64 // This way, libcap needs not be installed in build environment.
69 // Make up for header deficiencies
71 # define RAMFS_MAGIC ((unsigned)0x858458f6)
74 # define TMPFS_MAGIC ((unsigned)0x01021994)
80 // Recursively delete contents of rootfs
81 static void delete_contents(const char *directory, dev_t rootdev)
87 // Don't descend into other filesystems
88 if (lstat(directory, &st) || st.st_dev != rootdev)
91 // Recursively delete the contents of directories
92 if (S_ISDIR(st.st_mode)) {
93 dir = opendir(directory);
95 while ((d = readdir(dir))) {
96 char *newdir = d->d_name;
99 if (DOT_OR_DOTDOT(newdir))
102 // Recurse to delete contents
103 newdir = concat_path_file(directory, newdir);
104 delete_contents(newdir, rootdev);
109 // Directory should now be empty, zap it
113 // It wasn't a directory, zap it
121 static void drop_capset(int cap_idx)
125 /* Get the current capability mask */
129 caps.data[CAP_TO_INDEX(cap_idx)].inheritable &= ~CAP_TO_MASK(cap_idx);
131 /* And drop the capability. */
132 if (capset(&caps.header, caps.data) != 0)
133 bb_perror_msg_and_die("capset");
136 static void drop_bounding_set(int cap_idx)
140 ret = prctl(PR_CAPBSET_READ, cap_idx, 0, 0, 0);
142 bb_perror_msg_and_die("prctl: %s", "PR_CAPBSET_READ");
145 ret = prctl(PR_CAPBSET_DROP, cap_idx, 0, 0, 0);
147 bb_perror_msg_and_die("prctl: %s", "PR_CAPBSET_DROP");
151 static void drop_usermodehelper(const char *filename, int cap_idx)
154 char buf[sizeof(int)*3 * 2 + 8];
158 ret = open_read_close(filename, buf, sizeof(buf) - 1);
160 return; /* assuming files do not exist */
163 ret = sscanf(buf, "%u %u", &lo, &hi);
165 bb_perror_msg_and_die("can't parse file '%s'", filename);
168 lo &= ~(1 << cap_idx);
170 hi &= ~(1 << (cap_idx - 32));
172 fd = xopen(filename, O_WRONLY);
173 fdprintf(fd, "%u %u", lo, hi);
177 static void drop_capabilities(char *string)
181 cap = strtok(string, ",");
185 cap_idx = cap_name_to_number(cap);
186 drop_usermodehelper("/proc/sys/kernel/usermodehelper/bset", cap_idx);
187 drop_usermodehelper("/proc/sys/kernel/usermodehelper/inheritable", cap_idx);
188 drop_bounding_set(cap_idx);
189 drop_capset(cap_idx);
190 bb_error_msg("dropped capability: %s", cap);
191 cap = strtok(NULL, ",");
196 int switch_root_main(int argc, char **argv) MAIN_EXTERNALLY_VISIBLE;
197 int switch_root_main(int argc UNUSED_PARAM, char **argv)
199 char *newroot, *console = NULL;
204 // Parse args (-c console). '+': stop at first non-option
205 if (ENABLE_SWITCH_ROOT && (!ENABLE_RUN_INIT || applet_name[0] == 's')) {
208 "\0" "-2" /* minimum 2 args */,
213 char *cap_list = NULL;
216 "\0" "-2" /* minimum 2 args */,
221 drop_capabilities(cap_list);
227 // Change to new root directory and verify it's a different fs
232 if (st.st_dev == rootdev || getpid() != 1) {
233 // Show usage, it says new root must be a mountpoint
234 // and we must be PID 1
238 // Additional sanity checks: we're about to rm -rf /, so be REALLY SURE
239 // we mean it. I could make this a CONFIG option, but I would get email
240 // from all the people who WILL destroy their filesystems.
241 if (stat("/init", &st) != 0 || !S_ISREG(st.st_mode)) {
242 bb_error_msg_and_die("/init is not a regular file");
244 statfs("/", &stfs); // this never fails
245 if ((unsigned)stfs.f_type != RAMFS_MAGIC
246 && (unsigned)stfs.f_type != TMPFS_MAGIC
248 bb_error_msg_and_die("root filesystem is not ramfs/tmpfs");
251 // Zap everything out of rootdev
252 delete_contents("/", rootdev);
254 // Overmount / with newdir and chroot into it
255 if (mount(".", "/", NULL, MS_MOVE, NULL)) {
256 // For example, fails when newroot is not a mountpoint
257 bb_perror_msg_and_die("error moving root");
260 // The chdir is needed to recalculate "." and ".." links
261 /*xchdir("/"); - done in xchroot */
263 // If a new console specified, redirect stdin/stdout/stderr to it
265 int fd = open_or_warn(console, O_RDWR);
274 execv(argv[0], argv);
275 bb_perror_msg_and_die("can't execute '%s'", argv[0]);
279 From: Rob Landley <rob@landley.net>
280 Date: Tue, Jun 16, 2009 at 7:47 PM
281 Subject: Re: switch_root...
287 If you're _not_ running out of init_ramfs (if for example you're using initrd
288 instead), you probably shouldn't use switch_root because it's the wrong tool.
290 Basically what the sucker does is something like the following shell script:
292 find / -xdev | xargs rm -rf
298 There are a couple reasons that won't work as a shell script:
300 1) If you delete the commands out of your $PATH, your shell scripts can't run
301 more commands, but you can't start using dynamically linked _new_ commands
302 until after you do the chroot because the path to the dynamic linker is wrong.
303 So there's a step that needs to be sort of atomic but can't be as a shell
304 script. (You can work around this with static linking or very carefully laid
305 out paths and sequencing, but it's brittle, ugly, and non-obvious.)
307 2) The "find | rm" bit will actually delete everything because the mount points
308 still show up (even if their contents don't), and rm -rf will then happily zap
309 that. So the first line is an oversimplification of what you need to do _not_
310 to descend into other filesystems and delete their contents.
312 The reason we do this is to free up memory, by the way. Since initramfs is a
313 ramfs, deleting its contents frees up the memory it uses. (We leave it with
314 one remaining dentry for the new mount point, but that's ok.)
316 Note that you cannot ever umount rootfs, for approximately the same reason you
317 can't kill PID 1. The kernel tracks mount points as a doubly linked list, and
318 the pointer to the start/end of that list always points to an entry that's
319 known to be there (rootfs), so it never has to worry about moving that pointer
320 and it never has to worry about the list being empty. (Back around 2.6.13
321 there _was_ a bug that let you umount rootfs, and the system locked hard the
322 instant you did so endlessly looping to find the end of the mount list and
323 never stopping. They fixed it.)
325 Oh, and the reason we mount --move _and_ do the chroot is due to the way "/"
326 works. Each process has two special symlinks, ".", and "/". Each of them
327 points to the dentry of a directory, and give you a location paths can start
328 from. (Historically ".." was also special, because you could enter a
329 directory via a symlink so backing out to the directory you came from doesn't
330 necessarily mean the one physically above where "." points to. These days I
331 think it's just handed off to the filesystem.)
333 Anyway, path resolution starts with "." or "/" (although the "./" at the start
334 of the path may be implicit), meaning it's relative to one of those two
335 directories. Your current directory, and your current root directory. The
336 chdir() syscall changes where "." points to, and the chroot() syscall changes
337 where "/" points to. (Again, both are per-process which is why chroot only
338 affects your current process and its child processes.)
340 Note that chroot() does _not_ change where "." points to, and back before they
341 put crazy security checks into the kernel your current directory could be
342 somewhere you could no longer access after the chroot. (The command line
343 chroot does a cd as well, the chroot _syscall_ is what I'm talking about.)
345 The reason mounting something new over / has no obvious effect is the same
346 reason mounting something over your current directory has no obvious effect:
347 the . and / links aren't recalculated after a mount, so they still point to
348 the same dentry they did before, even if that dentry is no longer accessible
349 by other means. Note that "cd ." is a NOP, and "chroot /" is a nop; both look
350 up the cached dentry and set it right back. They don't re-parse any paths,
351 because they're what all paths your process uses would be relative to.
353 That's why the careful sequencing above: we cd into the new mount point before
354 we do the mount --move. Moving the mount point would otherwise make it
355 totally inaccessible to us because cd-ing to the old path wouldn't give it to
356 us anymore, and cd "/" just gives us the cached dentry from when the process
357 was created (in this case the old initramfs one). But the "." symlink gives
358 us the dentry of the filesystem we just moved, so we can then "chroot ." to
359 copy that dentry to "/" and get the new filesystem. If we _didn't_ save that
360 dentry in "." we couldn't get it back after the mount --move.
362 (Yes, this is all screwy and I had to email questions to Linus Torvalds to get
363 it straight myself. I keep meaning to write up a "how mount actually works"