Re: [PATCH] [RFC] mnt: add ability to clone mntns starting with the current root

[Andy Lutomirski <luto-kltTT9wpgjJwATOyAt5JVQ@public.gmane.org>]

On Tue, Oct 7, 2014 at 1:30 PM, Eric W. Biederman <ebiederm-aS9lmoZGLiVWk0Htik3J/w@public.gmane.org> wrote:
> Al Viro <viro-3bDd1+5oDREiFSDQTTA3OLVCufUGDwFn@public.gmane.org> writes:
>
> 2> On Tue, Oct 07, 2014 at 02:30:40PM +0100, Al Viro wrote:
>>> On Tue, Oct 07, 2014 at 04:12:57PM +0400, Andrey Vagin wrote:
>>> > Another problem is that rootfs can't be hidden from a container, because
>>> > rootfs can't be moved or umounted.
>>>
>>> ... which is a bug in mntns_install(), AFAICS.
>>
>> Ability to get to exposed rootfs, that is.
>
> The container side of this argument is pretty bogus.  It only applies
> if user namespaces are not used for the container.
>
> So it is only root (and not root in a container) who can get to the
> exposed rootfs.
>
> I have a vague memory someone actually had a real use in miminal systems
> for being able to get back to the rootfs and being able to use rootfs as
> the rootfs.  There was even a patch at that time that Andrew Morton was
> carrying for a time to allow unmounting root and get at rootfs, and to
> prevent the oops on rootfs unmount in some way.
>
> So not only do I not think it is a bug to get back too rootfs, I think
> it is a feature that some people have expressed at least half-way sane
> uses for.
>
>>> > Here is an example how to get access to rootfs:
>>> > fd = open("/proc/self/ns/mnt", O_RDONLY)
>>> > umount2("/", MNT_DETACH);
>>> > setns(fd, CLONE_NEWNS)
>>> >
>>> > rootfs may contain data, which should not be avaliable in CT-s.
>>>
>>> Indeed.
>>
>> ... and it looks like the above is what your mangled reproducer in previous
>> patch had been made of -
>>       fd = open("/proc/self/ns/mnt", O_RDONLY)
>>       umount2("/", MNT_DETACH);
>>       setns(fd, CLONE_NEWNS)
>>       umount2("/", MNT_DETACH);
>>
>> IMO what it shows is setns() bug.  This "switch root/cwd, no matter what"
>> is wrong.
>
> IMO the bug is allowing us to unmount things that should never be unmounted.
>
> In a mount namespace created with just user namespace permissions we
> can't get at rootfs because MNT_LOCKED is set on the root directory
> and thus it can not be mounted.
>
> Further if anyone has permission to call chroot and chdir on any mount
> in a mount namespace (that isn't currently covered) they can get at all
> of them that are not currently covered.  A mount namespace where no one
> can get at any uncovered filesystem seems to be the definition of
> useless and ridiculous.
>
>
> Now there is a bug in that MNT_DETACH today does not currently enforce
> MNT_LOCKED on submounts of the mount point that is detached. I am
> currently looking at how to construct the appropriate permission check
> to prevent that.  Unfortunately I can not disallow MNT_DETACH with
> submounts all together as that breaks too many legitimate uses.

Why should MNT_LOCKED on submounts be enforced?

Is it because, if you retain a reference to the detached tree, then
you can see under the submounts?  If so, let's fix *that*.  Because
otherwise the whole model of pivot_root + detach will break.

Also, damn it, we need change_the_ns_root instead of pivot_root.  I
doubt that any container programs actually want to keep the old root
attached after pivot_root.

--Andy