[PATCH] x86/mtrr: only administrator can read the configurations.

[Kees Cook]

On Tue, Nov 12, 2019 at 06:49:56PM +0100, Borislav Petkov wrote:
> On Mon, Nov 11, 2019 at 09:56:16AM -0800, Kees Cook wrote:
> > Some recap from being accidentally offlist:
> > 
> > - this patch should check capabilities at open time (or retain the
> >   checks on the opener's permissions for later checks).
> > 
> > - changing the DAC permissions might break something that expects to
> >   read mtrr when not uid 0.
> > 
> > - if we leave the DAC permissions alone and just move the capable check
> >   to the opener, we should get the intent of the original patch. (i.e.
> >   check against CAP_SYS_ADMIN not just the wider uid 0.)
> > 
> > - *this may still break things* if userspace expects to be able to
> >   read other parts of the file as non-uid-0 and non-CAP_SYS_ADMIN.
> >   If *that* is the case, then we need to censor the contents using
> >   the opener's permissions (as done in other /proc cases).
> > 
> > I think the most cautious way forward is something like
> > 51d7b120418e ("/proc/iomem: only expose physical resource addresses to
> > privileged users"). Untested (and should likely be expanded to know
> > about read vs write for lockdown interaction):
> 
> I'm back'n'forth on this.
> 
> So tglx and I agree that it doesn't make a whole lotta sense for
> non-privileged luserspace to be able to read /proc/mtrr because it is a
> small leak and normal users shouldn't care about the caching attributes
> of memory regions in the first place.
> 
> So maybe we should do the second variant.
> 
> But then we're not supposed to break luserspace.
> 
> But then we can revert it if we do...
> 
> Ugh.

Shall I send a patch for just moving the capable() checks into open()
and if someone yells we switch to the other option on the assumption
that then we'll have a real-world case we can test the other solution
against?

-- 
Kees Cook