Re: [PATCH 3/3] x86/sev: add a SVSM vTPM platform device

["Jarkko Sakkinen" <jarkko@kernel.org>]

On Wed Jan 15, 2025 at 12:48 AM EET, Jarkko Sakkinen wrote:
> On Wed Jan 15, 2025 at 12:46 AM EET, Jarkko Sakkinen wrote:
> > On Tue Jan 14, 2025 at 12:42 PM EET, Stefano Garzarella wrote:
> > > Hi Jarkko,
> > >
> > > On Thu, 19 Dec 2024 at 17:07, Stefano Garzarella <sgarzare@redhat.com> wrote:
> > > >
> > > > On Thu, Dec 19, 2024 at 05:40:58PM +0200, Jarkko Sakkinen wrote:
> > > > >On Thu Dec 19, 2024 at 5:35 PM EET, Stefano Garzarella wrote:
> > > > >> So to use them directly in sev, we would have to move these definitions
> > > > >> into include/linux/tpm.h or some other file in inlcude/. Is this
> > > > >> acceptable for TPM maintainers?
> > > > >
> > > > >There's only me.
> > > > >
> > > > >I don't know.
> > > > >
> > > > >What you want to put to include/linux/tpm.h anyway?
> > > >
> > > > At least tpmm_chip_alloc(), tpm2_probe(), and tpm_chip_register()
> > > >
> > > > >I have not followed this discussion.
> > > >
> > > > Let me try to summarize what we are doing: We are writing a small TPM
> > > > driver to support AMD SEV-SNP SVSM. Basically SVSM defines some sort of
> > > > hypercalls, which the guest OS can call to talk to the emulated vTPM.
> > > >
> > > > In the current version of this series, based on James' RFC, we have an
> > > > intermediate module (tpm_platform) and then another small driver
> > > > (platform_device) in arch/x86/coco/sev/core.c that registers the
> > > > callback to use.
> > > >
> > > > To avoid the intermediate driver (Jason correct me if I misunderstood),
> > > > we want to register the `tpm_chip` with its `tpm_class_ops` directly in
> > > > arch/x86/coco/sev/core.c where it's easy to use "SVSM calls" (i.e.
> > > > svsm_perform_call_protocol()).
> > > >
> > > > And here I have this problem, so I was proposing to expose these APIs.
> > > > BTW, we do have an alternative though that I proposed in the previous
> > > > email that might avoid this.
> > >
> > > Any thought on this?
> >
> > A redundant super low-quality TPM stack driver implemtation to support
> > only single vendor's vTPM with speculative generalization.
> >
> > It's a formula for destruction really.
> >
> > I don't know if I event want to comment on this. Figure out a better
> > solution I guess that works together sound with existing stack.
> >
> > If that helps we could make the main TPM driver only Y/N (instead of
> > tristate).
>
> Also e.g. James' hmac encryption: not a single bug fixed by the author,
> which does further reduce my ability to have any possible trust on this.
>
> I do care quality over features, sorry.

One more rant.

It's engineering problem to find **a fit** for the existing art. For
You can set the constraint here as "no two TPM stacks".

I know also almost nothing about SVSM. E.g. I don't understand why a
vTPM cannot be seen as fTPM by the guest, and why this needs user
space exported device (please do not answer here, do a better job
instead).

Even if I wanted to say how this should be changed, I could not
because it too far away to make any possible sense to begin with.
And I don't want to take the risk of those words being used as an
argument later on, when I don't even know what I'm looking.

BR, Jarkko