[PATCH v1 3/3] KEYS: trusted: Introduce support for NXP CAAM-based trusted keys

[Jarkko Sakkinen]

On Sat, Mar 27, 2021 at 01:41:24PM +0100, David Gstir wrote:
> Hi!
> 
> > On 25.03.2021, at 06:26, Sumit Garg <sumit.garg at linaro.org> wrote:
> > 
> > On Wed, 24 Mar 2021 at 19:37, Ahmad Fatoum <a.fatoum at pengutronix.de> wrote:
> >> 
> >> Hello Sumit,
> >> 
> >> On 24.03.21 11:47, Sumit Garg wrote:
> >>> On Wed, 24 Mar 2021 at 14:56, Ahmad Fatoum <a.fatoum at pengutronix.de> wrote:
> >>>> 
> >>>> Hello Mimi,
> >>>> 
> >>>> On 23.03.21 19:07, Mimi Zohar wrote:
> >>>>> On Tue, 2021-03-23 at 17:35 +0100, Ahmad Fatoum wrote:
> >>>>>> On 21.03.21 21:48, Horia Geantă wrote:
> >>>>>>> caam has random number generation capabilities, so it's worth using that
> >>>>>>> by implementing .get_random.
> >>>>>> 
> >>>>>> If the CAAM HWRNG is already seeding the kernel RNG, why not use the kernel's?
> >>>>>> 
> >>>>>> Makes for less code duplication IMO.
> >>>>> 
> >>>>> Using kernel RNG, in general, for trusted keys has been discussed
> >>>>> before.   Please refer to Dave Safford's detailed explanation for not
> >>>>> using it [1].
> >>>> 
> >>>> The argument seems to boil down to:
> >>>> 
> >>>> - TPM RNG are known to be of good quality
> >>>> - Trusted keys always used it so far
> >>>> 
> >>>> Both are fine by me for TPMs, but the CAAM backend is new code and neither point
> >>>> really applies.
> >>>> 
> >>>> get_random_bytes_wait is already used for generating key material elsewhere.
> >>>> Why shouldn't new trusted key backends be able to do the same thing?
> >>>> 
> >>> 
> >>> Please refer to documented trusted keys behaviour here [1]. New
> >>> trusted key backends should align to this behaviour and in your case
> >>> CAAM offers HWRNG so we should be better using that.
> >> 
> >> Why is it better?
> >> 
> >> Can you explain what benefit a CAAM user would have if the trusted key
> >> randomness comes directly out of the CAAM instead of indirectly from
> >> the kernel entropy pool that is seeded by it?
> > 
> > IMO, user trust in case of trusted keys comes from trusted keys
> > backend which is CAAM here. If a user doesn't trust that CAAM would
> > act as a reliable source for RNG then CAAM shouldn't be used as a
> > trust source in the first place.
> > 
> > And I think building user's trust for kernel RNG implementation with
> > multiple entropy contributions is pretty difficult when compared with
> > CAAM HWRNG implementation.
> 
> Generally speaking, I’d say trusting the CAAM RNG and trusting in it’s
> other features are two separate things. However, reading through the CAAM
> key blob spec I’ve got here, CAAM key blob keys (the keys that secure a blob’s
> content) are generated using its internal RNG. So I’d save if the CAAM RNG
> is insecure, so are generated key blobs. Maybe somebody with more insight
> into the CAAM internals can verify that, but I don’t see any point in using
> the kernel’s RNG as long as we let CAAM generate the key blob keys for us.

Here's my long'ish analysis. Please read it to the end if by ever means
possible, and apologies, I usually try to keep usually my comms short, but
this requires some more meat than the usual.

The Bad News
============

Now that we add multiple hardware trust sources for trusted keys, will
there ever be a scenario where a trusted key is originally sealed with a
backing hardware A, unsealed, and resealed with hardware B?

The hardware and vendor neutral way to generate the key material would be
unconditionally always just the kernel RNG.

CAAM is actually worse than TCG because it's not even a standards body, if
I got it right. Not a lot but at least a tiny fraction.

This brings an open item in TEE patches: trusted_tee_get_random() is an
issue in generating kernel material. I would rather replace that with
kernel RNG *for now*, because the same open question applies also to ARM
TEE. It's also a single company controlled backing technology.

By all practical means, I do trust ARM TEE in my personal life but this is
not important.

CAAM *and* TEE backends break the golden rule of putting as little trust as
possible to anything, even not anything weird is clear at sight, as
security is essentially a game of known unknowns and unknown unknowns.

Unfortunately, TPM trusted keys started this bad security practice, and
obviously it cannot be fixed without breaking uapi backwards compatibility.

This leaves me exactly two rational options:

A. Add a patch to remove trusted_tee_get_random() and use kernel RNG
   instead.
B. Drop the whole TEE patch set up until I have good reasons to believe
   that it's the best possible idea ever to use TEE RNG.

Doing does (A) does not disclude of doing (B) later on, if someone some
day sends a patch with sound reasoning.

It's also good to understand that when some day a vendor D, other than TCG,
CAAM or ARM, comes up, we need to go again this lenghty and messy
discussion. Now this already puts an already accepted patch set into a
risk, because by being a responsible maintainer I would have legit reasons
just simply to drop it.

OK, but....

The GOOD News
=============

So there's actually option (C) that also fixes the TPM trustd keys issue:

Add a new kernel patch, which:

1. Adds the use of kernel RNG as a boot option.
2. If this boot option is not active, the subsystem will print a warning
   to klog denoting this.
3. Default is of course vendor RNG given the bad design issue in the TPM
   trusted keys, but the warning in klog will help to address it at least
   a bit.
4. Document all this to Documentation/security/keys/trusted-encrypted.rst.

I'd prefer the choice between A, B and C be concluded rather sooner than
later.

/Jarkko