On 16/03/2021 20:04, Jann Horn wrote:
> On Tue, Mar 16, 2021 at 6:02 PM Mickaël Salaün <mic@xxxxxxxxxxx> wrote:
>> One could argue that chroot(2) is useless without a properly populated
>> root hierarchy (i.e. without /dev and /proc). However, there are
>> multiple use cases that don't require the chrooting process to create
>> file hierarchies with special files nor mount points, e.g.:
>> * A process sandboxing itself, once all its libraries are loaded, may
>> not need files other than regular files, or even no file at all.
>> * Some pre-populated root hierarchies could be used to chroot into,
>> provided for instance by development environments or tailored
>> distributions.
>> * Processes executed in a chroot may not require access to these special
>> files (e.g. with minimal runtimes, or by emulating some special files
>> with a LD_PRELOADed library or seccomp).
>>
>> Unprivileged chroot is especially interesting for userspace developers
>> wishing to harden their applications. For instance, chroot(2) and Yama
>> enable to build a capability-based security (i.e. remove filesystem
>> ambient accesses) by calling chroot/chdir with an empty directory and
>> accessing data through dedicated file descriptors obtained with
>> openat2(2) and RESOLVE_BENEATH/RESOLVE_IN_ROOT/RESOLVE_NO_MAGICLINKS.
>
> I don't entirely understand. Are you writing this with the assumption
> that a future change will make it possible to set these RESOLVE flags
> process-wide, or something like that?
No, this scenario is for applications willing to sandbox themselves and
only use the FDs to access legitimate data.
>
>
> As long as that doesn't exist, I think that to make this safe, you'd
> have to do something like the following - let a child process set up a
> new mount namespace for you, and then chroot() into that namespace's
> root:
>
> struct shared_data {
> int root_fd;
> };
> int helper_fn(void *args) {
> struct shared_data *shared = args;
> mount("none", "/tmp", "tmpfs", MS_NOSUID|MS_NODEV, "");
> mkdir("/tmp/old_root", 0700);
> pivot_root("/tmp", "/tmp/old_root");
> umount("/tmp/old_root", "");
> shared->root_fd = open("/", O_PATH);
> }
> void setup_chroot() {
> struct shared_data shared = {};
> prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0);
> clone(helper_fn, my_stack,
> CLONE_VFORK|CLONE_VM|CLONE_FILES|CLONE_NEWUSER|CLONE_NEWNS|SIGCHLD,
> NULL);
> fchdir(shared.root_fd);
> chroot(".");
> }
What about this?
chdir("/proc/self/fdinfo");
chroot(".");
close(all unnecessary FDs);
>
> [...]
>> diff --git a/fs/open.c b/fs/open.c
> [...]
>> +static inline int current_chroot_allowed(void)
>> +{
>> + /*
>> + * Changing the root directory for the calling task (and its future
>> + * children) requires that this task has CAP_SYS_CHROOT in its
>> + * namespace, or be running with no_new_privs and not sharing its
>> + * fs_struct and not escaping its current root (cf. create_user_ns()).
>> + * As for seccomp, checking no_new_privs avoids scenarios where
>> + * unprivileged tasks can affect the behavior of privileged children.
>> + */
>> + if (task_no_new_privs(current) && current->fs->users == 1 &&
>
> this read of current->fs->users should be using READ_ONCE()
Right, I'll fix this.
>
>> + !current_chrooted())
>> + return 0;
>> + if (ns_capable(current_user_ns(), CAP_SYS_CHROOT))
>> + return 0;
>> + return -EPERM;
>> +}
> [...]
>
> Overall I think this change is a good idea.
>
