Re: [PATCH v6 40/42] virt: Add SEV-SNP guest driver

[Peter Gonda <pgonda@google.com>]

On Wed, Oct 27, 2021 at 3:13 PM Brijesh Singh <brijesh.singh@amd.com> wrote:
>
>
>
> On 10/27/21 4:05 PM, Peter Gonda wrote:
> ....
>
> >>>>>
> >>>>> Thanks for updating this sequence number logic. But I still have some
> >>>>> concerns. In verify_and_dec_payload() we check the encryption header
> >>>>> but all these fields are accessible to the hypervisor, meaning it can
> >>>>> change the header and cause this sequence number to not get
> >>>>> incremented. We then will reuse the sequence number for the next
> >>>>> command, which isn't great for AES GCM. It seems very hard to tell if
> >>>>> the FW actually got our request and created a response there by
> >>>>> incrementing the sequence number by 2, or if the hypervisor is acting
> >>>>> in bad faith. It seems like to be safe we need to completely stop
> >>>>> using this vmpck if we cannot confirm the PSP has gotten our request
> >>>>> and created a response. Thoughts?
> >>>>>
> >>>>
> >>>> Very good point, I think we can detect this condition by rearranging the
> >>>> checks. The verify_and_dec_payload() is called only after the command is
> >>>> succesful and does the following checks
> >>>>
> >>>> 1) Verifies the header
> >>>> 2) Decrypts the payload
> >>>> 3) Later we increment the sequence
> >>>>
> >>>> If we arrange to the below order then we can avoid this condition.
> >>>> 1) Decrypt the payload
> >>>> 2) Increment the sequence number
> >>>> 3) Verify the header
> >>>>
> >>>> The descryption will succeed only if PSP constructed the payload.
> >>>>
> >>>> Does this make sense ?
> >>>
> >>> Either ordering seems fine to me. I don't think it changes much though
> >>> since the header (bytes 30-50 according to the spec) are included in
> >>> the authenticated data of the encryption. So any hypervisor modictions
> >>> will lead to a decryption failure right?
> >>>
> >>> Either case if we do fail the decryption, what are your thoughts on
> >>> not allowing further use of that VMPCK?
> >>>
> >>
> >> We have limited number of VMPCK (total 3). I am not sure switching to
> >> different will change much. HV can quickly exaust it. Once we have SVSM
> >> in-place then its possible that SVSM may use of the VMPCK. If the
> >> decryption failed, then maybe its safe to erase the key from the secrets
> >> page (in other words guest OS cannot use that key for any further
> >> communication). A guest can reload the driver will different VMPCK id
> >> and try again.
> >
> > SNP cannot really cover DOS at all since the VMM could just never
> > schedule the VM. In this case we know that the hypervisor is trying to
> > mess with the guest, so my preference would be to stop sending guest
> > messages to prevent that duplicated IV usage. If one caller gets an
> > EBADMSG it knows its in this case but the rest of userspace has no
> > idea. Maybe log an error?
> >
> >>
>
> Yap, we cannot protect against the DOS. This is why I was saying that we
> zero the key from secrets page so that guest cannot use that key for any
> future communication (whether its from rest of userspace or kexec
> kernel). I can update the driver to log the message and ensure that
> future messages will *not* use that key. The VMPCK ID is a module
> params, so a guest can reload the driver to use different VMPCK.

Duh! Sorry I thought you said we needed a VMPL0 SVSM to do that. That
sounds great.

>
>
> >> thanks