Does MNT_LOCKED work as expected?

Does MNT_LOCKED work as expected?

[Andrey Wagin]

Hi All,

I think I found a case, when MNT_LOCKED (5ff9d8a65ce8 "vfs: Lock in
place mounts from more privileged users") doesn't help to hide
overmounted parts from unprivileged users.
The problem exists for mounts, which are not overmounted entirely. In
such cases we can open a directory from a target mount, which is not
overmounted. Then we do pivot_root to move all mounts in a temporary
directory. At the final step we deatch all mounts from the temporary
directory. After that all children mounts are umounted from the target
mount and we can use our file descriptor to open files, which have
been overmount before.

Here is a example https://github.com/avagin/userns_vs_mntns.

Re: Does MNT_LOCKED work as expected?

[Eric W. Biederman]

Andrey Wagin <avagin@gmail.com> writes:

> Hi All,
>
> I think I found a case, when MNT_LOCKED (5ff9d8a65ce8 "vfs: Lock in
> place mounts from more privileged users") doesn't help to hide
> overmounted parts from unprivileged users.

> The problem exists for mounts, which are not overmounted entirely. In
> such cases we can open a directory from a target mount, which is not
> overmounted. Then we do pivot_root to move all mounts in a temporary
> directory. At the final step we deatch all mounts from the temporary
> directory. After that all children mounts are umounted from the target
> mount and we can use our file descriptor to open files, which have
> been overmount before.
>
> Here is a example https://github.com/avagin/userns_vs_mntns.

I am looking at this now to make certain I am understand what you are
seeing but so far it looks like lazy unmounts allow you to bypass
MNT_LOCKED :(

Grumble.

Eric