On Wed, 2024-05-22 at 17:11 +0300, Jarkko Sakkinen wrote: > For tpm_crb we should actually disable HMAC at some point. It is > essentially a performance regression for it. You'd think that, because of the shared buffer and no bus, but you never quite know. For instance several confidential computing early implementations used the crb interface set up by qemu (i.e. over shared buffers which are readable by the host). For them the only way to get security is with sessions. Even with the default Intel CRB, the TPM transaction isn't handled directly by the main CPU, it's offloaded to the ME (which we all know google loves because of its tight security boundary). It is entirely possible to imagine a software interposer running in the ME snooping the CRB buffer. A very different type of attack from the LPB interposer, but plausible non the less. James
