Re: [RFC PATCH v2 03/13] ima: invalidate unsupported PCR banks

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Sun, 2025-03-23 at 15:09 +0100, Nicolai Stange wrote:
> Normally IMA would extend a template hash of each bank's associated
> algorithm into a PCR. However, if a bank's hash algorithm is
> unavailable to the kernel at IMA init time, it would fallback to
> extending padded SHA1 hashes instead.
> 
> That is, if e.g. SHA-256 was missing at IMA init, it would extend
> padded SHA1 template hashes into a PCR's SHA-256 bank.
> 
> The ima_measurement command (marked as experimental) from ima-evm-
> utils would accordingly try both variants when attempting to verify a
> measurement list against PCRs. keylime OTOH doesn't seem to -- it
> expects the template hash type to match the PCR bank algorithm. I
> would argue that for the latter case, the fallback scheme could
> potentially cause hard to debug verification failures.
> 
> There's another problem with the fallback scheme: right now, SHA-1
> availability is a hard requirement for IMA, and it would be good for
> a number of reasons to get rid of that. However, if SHA-1 is not
> available to the kernel, it can hardly provide padded SHA-1 template
> hashes for PCR banks with unsupported algos.

I think this was done against the day IMA only supported sha1 and the
TPM sha256 and beyond so there'd at least be a record that could be
replayed.  I think today with most distros defaulting IMAs hash to
sha256 that's much less of a problem.

> There are several more or less reasonable alternatives possible,
> among them are:
> a.) Instead of padded SHA-1, use padded/truncated ima_hash template
>     hashes.
> b.) Don't extend unsupported banks at all.
> c.) Record every event as a violation, i.e. extend unsupported banks
>     with 0xffs.
> d.) Invalidate unsupported banks at least once by extending with a
> unique
>     constant (e.g. with 0xfes).

Instead of any of that, why not do what the TCG tells us to do for
unsupported banks and simply cap them with 0xffffffff record
EV_SEPARATOR and stop extending to them? (note this would probably
require defining a separator event for IMA)

Regards,

James