Re: [RFC PATCH 00/12] x86: Trenchboot secure late launch Linux kernel support

[Matthew Garrett]

On Thu, Mar 26, 2020 at 3:52 PM Andy Lutomirski <luto@xxxxxxxxxx> wrote:
>
> On Thu, Mar 26, 2020 at 2:28 PM Matthew Garrett <mjg59@xxxxxxxxxx> wrote:
> > https://trustedcomputinggroup.org/wp-content/uploads/TCG_PlatformResetAttackMitigationSpecification_1.10_published.pdf
> > - you want to protect in-memory secrets from a physically present
> > attacker hitting the reset button, booting something else and just
> > dumping RAM. This is avoided by setting a variable at boot time (in
> > the boot stub), and then clearing it on reboot once the secrets have
> > been cleared from RAM. If the variable isn't cleared, the firmware
> > overwrites all RAM contents before booting anything else.
>
> I admit my information is rather dated, but I'm pretty sure that at
> least some and possibly all TXT implementations solve this more
> directly.  In particular, as I understand it, when you TXT-launch
> anything, a nonvolatile flag in the chipset is set.  On reboot, the
> chipset will not allow access to memory *at all* until an
> authenticated code module wipes memory and clears that flag.

Mm, yes, this one might be something we can just ignore in the TXT case.

> > When you say "re-launch", you mean perform a second secure launch? I
> > think that would work, as long as we could reconstruct an identical
> > state to ensure that the PCR17 values matched - and that seems like a
> > hard problem.
>
> Exactly.  I would hope that performing a second secure launch would
> reproduce the same post-launch PCRs as the first launch.  If the
> kernel were wise enough to record all PCR extensions, it could replay
> them.

That presumably depends on how much state is in the measured region -
we can't just measure the code in order to assert that we're secure.

> In any case, I'm kind of with Daniel here.  We survived for quite a
> long time without EFI variables at all.  The ability to write them is
> nice, and we certainly need some way, however awkward, to write them
> on rare occasions, but I don't think we really need painless runtime
> writes to EFI variables.

I'm fine with a solution that involves jumping through some hoops, but
it feels like simply supporting measuring and passing through the
runtime services would be fine - if you want to keep them outside the
TCB, build a kernel that doesn't have EFI runtime service support and
skip that measurement?