On Sun, Nov 12, 2023 at 3:14 PM Christian Brauner <brauner@xxxxxxxxxx> wrote:
>
> On Fri, Nov 10, 2023 at 12:06:14PM -0500, Charles Mirabile wrote:
> > This is a one line change that makes `linkat` aware of namespaces when
> > checking for capabilities.
> >
> > As far as I can tell, the call to `capable` in this code dates back to
> > before the `ns_capable` function existed, so I don't think the author
> > specifically intended to prefer regular `capable` over `ns_capable`,
> > and no one has noticed or cared to change it yet... until now!
> >
> > It is already hard enough to use `linkat` to link temporarily files
> > into the filesystem without the `/proc` workaround, and when moving
> > a program that was working fine on bare metal into a container,
> > I got hung up on this additional snag due to the lack of namespace
> > awareness in `linkat`.
>
> I agree that it would be nice to relax this a bit to make this play
> nicer with containers.
>
> The current checks want to restrict scenarios where an application is
> able to create a hardlink for an arbitrary file descriptor it has
> received via e.g., SCM_RIGHTS or that it has inherited.
Makes sense.
>
> So we want to somehow get a good enough approximation to the question
> whether the caller would have been able to open the source file.
>
> When we check for CAP_DAC_READ_SEARCH in the caller's namespace we
> presuppose that the file is accessible in the current namespace and that
> CAP_DAC_READ_SEARCH would have been enough to open it. Both aren't
> necessarily true. Neither need the file be accessible, e.g., due to a
> chroot or pivot_root nor need CAP_DAC_READ_SEARCH be enough. For
> example, the file could be accessible in the caller's namespace but due
> to uid/gid mapping the {g,u}id of the file doesn't have a mapping in the
> caller's namespace. So that doesn't really cut it imho.
Good point.
>
> However, if we check for CAP_DAC_READ_SEARCH in the namespace the file
> was opened in that could work. We know that the file must've been
> accessible in the namespace the file was opened in and we
> know that the {g,u}id of the file must have been mapped in the namespace
> the file was opened in. So if we check that the caller does have
> CAP_DAC_READ_SEARCH in the opener's namespace we can approximate that
> the caller could've opened the file.
Would that be the namespace pointed to by `->f_cred->user_ns` on the
struct file corresponding to the fd?
If so is there a better way to surface that struct file for checking than this?
error=-ENOENT;
if(flags & AT_EMPTY_PATH && !old->name[0]) {
struct file *file = fget(oldfd);
bool capable = ns_capable(file->f_cred->user_ns, CAP_DAC_READ_SEARCH);
fput(file);
if(!capable)
goto out_putnames;
}
>
> So that should allow us to enabled this for containers.
>
Best - Charlie
