Re: [PATCH 2/9] tpm: Allow PCR 23 to be restricted to kernel-only use

[Matthew Garrett]

On Wed, Feb 24, 2021 at 10:00:53AM -0800, James Bottomley wrote:
> On Sat, 2021-02-20 at 01:32 +0000, Matthew Garrett wrote:
> > Under certain circumstances it might be desirable to enable the
> > creation of TPM-backed secrets that are only accessible to the
> > kernel. In an ideal world this could be achieved by using TPM
> > localities, but these don't appear to be available on consumer
> > systems.
> 
> I don't understand this ... the localities seem to work fine on all the
> systems I have ... is this some embedded thing?

I haven't made it work on an HP Z440 or a Lenovo P520. So now I'm
wondering whether having chipsets with TXT support (even if it's turned
off) confuse this point. Sigh. I'd really prefer to use localities than
a PCR, so if it works on client platforms I'd be inclined to say we'll
do a self-test and go for that, and workstation vendors can just
recommend their customers use UPSes or something.

> >  An alternative is to simply block userland from modifying one of the
> > resettable PCRs, leaving it available to the kernel. If the kernel
> > ensures that no userland can access the TPM while it is carrying out
> > work, it can reset PCR 23, extend it to an arbitrary value, create or
> > load a secret, and then reset the PCR again. Even if userland somehow
> > obtains the sealed material, it will be unable to unseal it since PCR
> > 23 will never be in the appropriate state.
> 
> This seems a bit arbitrary: You're removing this PCR from user space
> accessibility, but PCR 23 is defined as "Application Support" how can
> we be sure no application will actually want to use it (and then fail)?

Absolutely no way of guaranteeing that, and enabling this option is
certainly an ABI break.

> Since PCRs are very scarce, why not use a NV index instead.  They're
> still a bounded resource, but most TPMs have far more of them than they
> do PCRs, and the address space is much bigger so picking a nice
> arbitrary 24 bit value reduces the chance of collisions.

How many write cycles do we expect the NV to survive? But I'll find a
client system with a TPM and play with locality support there - maybe we
can just avoid this problem anyway.