Re: SVSM vTPM specification

[Christophe de Dinechin <dinechin@redhat.com>]

On 2022-10-19 at 09:08 +01, Daniel P. Berrangé <berrange@redhat.com> wrote...
> On Wed, Oct 19, 2022 at 07:47:41AM +0200, Christophe de Dinechin wrote:
>>
>>
>> > On 18 Oct 2022, at 22:22, Dov Murik <dovmurik@linux.ibm.com> wrote:
>> >
>> >
>> >
>> > On 13/10/2022 18:30, James Bottomley wrote:
>> >> On Thu, 2022-10-13 at 10:14 -0500, Tom Lendacky wrote:
>> >>> On 10/12/22 13:44, James Bottomley wrote:
>> >>>> On Wed, 2022-10-12 at 18:33 +0100, Dr. David Alan Gilbert wrote:
>> >>>>> I think one of the vTPMs keys should be in the SNP attestation
>> >>>>> report (the EK???) - I think that would allow you to attest that
>> >>>>> the vTPM you're talking to is a vTPM running in an SNP protected
>> >>>>> firmware.
>> >>>>
>> >>>> Traditionally the TPM identity is the public EK, so that should
>> >>>> definitely be in the report.  Ideally, I think the public storage
>> >>>> root key (key derived from the owner seed) should be in there two
>> >>>> because it makes it easy to create keys that can only be read by
>> >>>> the TPM (keys should be in the owner hierarchy which means they
>> >>>> have to be encrypted to the storage seed, so we need to know what a
>> >>>> public key corresponding to it is).
>> >>>>
>> >>>> One wrinkle of the above is that, when provisioned, the TPM will
>> >>>> only have the seeds, not the keys (the keys are derived from the
>> >>>> seeds via a TPM2_CreatePrimary command).  The current TPM
>> >>>> provisioning guidance:
>> >>>>
>> >>>> https://trustedcomputinggroup.org/resource/tcg-tpm-v2-0-provisioning-guidance/
>> >>>>
>> >>>> Says that the EK should be at permanent handle
>> >>>>
>> >>>> 81010001
>> >>>>
>> >>>> And there's an update saying that should be the RSA-2048 key and
>> >>>> there should be an EC (NIST-P256) one at 81010002.  The
>> >>>> corresponding storage keys should be at 81000001 and 81000002
>> >>>> respectively.  I think when the SVSM provisions the TPM, it should
>> >>>> run TPM2_CreatePrimary for those four keys and put them into the
>> >>>> persistent indexes, then insert the EC keys only for EK and SRK
>> >>>> into the attestation report.
>> >>>
>> >>> We only have 512 bits to work with for the SVSM-provided data, so
>> >>> would hashes of the keys be ok?
>> >>
>> >> Well, if you put the hashes in, the consuming entity would then have
>> >> to find out via an additional channel what the actual keys were
>> >> because you can't reverse the hash (it's possible, just more effort).
>> >> If you use point compression, an EC key (for the NIST p-256 curve) is
>> >> only 32 bytes anyway, so it's the same size as a sha256 hash, so I'd
>> >> say place the actual public keys into the report to give complete and
>> >> usable information
>> >>
>> >
>> > Do we need to leave room for a Guest-Owner-provided nonce?  Guest owner
>> > will provide it to the guest OS which will provide it to the SVSM to be
>> > included in REPORT_DATA of the VMPL0 attestation report.
>> >
>> I think it’s a good idea, but I’m not clear on which component in the
>> guest would be responsible for that exchange.
>>
>> You need a functioning networking stack, and you need a way to know what
>> attestation server to talk to.
>>
>> IIUC, at this stage, we have no valid storage yet, since that would
>> require having received the response from the attestation server.  So the
>> only storage we have is host storage managed by the hypervisor, plus
>> whatever fits in your SVSM-provided data.
>>
>> Can we send this data without additional encryption since it’s already
>> cyphertext? Or do we want additional mechanisms, and if so, what root
>> certificates do we use? Where do we get them from without guest storage?
>>
>> Do we rely on guest networking being up at that point? Or do we need to
>> provide additional mechanisms in the hypervisor to perform this initial
>> exchange on behalf of the guest?
>
> I'd be inclined to not rely on guest networking, and probably even
> strictly decouple what the SVSM does to communicate, from any specific
> attestation server connection protocol or details.
>
> Implementing a network stack in SVSM looks like a very large piece of
> work. It is conceivable that the guest network device is connected to an
> entirely private network that can only communicate with other VMs
> belonging to that guest owner. IOW, we can't assume the guest NIC has the
> ability to route to the attestation server. I don't think we want to be
> adding a 2nd NIC just for firmware phase attestation either.

Yes, I was thinking about this scenario, but also wondering if the opposite
can happen, i.e. a guest having networking to the attestation server that
the host does not have access to. I can think of two cases that seem to make
sense (though not necessarily in public clouds):

- The guest has a dedicated hardware port to some private network
  e.g. through a virtual function
- The guest only networks to other VMs on the same host, without a phsyical
  NIC attached. Another (confidential) VM could be the attestation server.

So this is a bit of a can of worms.

>
> Further if SVSM uses networking, and talks directly to the remote
> attestation service, we can expect that attestation server will be
> using normal best practice and exposing its service over TLS with
> x509 server certs to be validated. That brings even more software
> complexity into the SVSM. It would need to be told what attestation
> server URL to use, though that is likely the easiest bit of the
> problem since fw_cfg can supply that. If any aspect of the
> attestation service protocol ever needs to change, then we also
> need to update SVSM to match.

Exactly.

>
>
> AFAICS, what SVSM needs to do for attestation is very straightforward.
> It sends a blob of data, and then blocks further execution until it
> gets another blob of data back in response.
>
> From the POV of SVSM, conceptually this does not even need connection
> based semantics in its communication method. So while we could use
> virtio-vsock instead of a NIC, even that could be considered overkill.
> An even simpler option to implement could be merely virtio-serial.
> There's a possible question of whether we want a mechanism that is
> cleanly separated from any device intended for guest OS usage. That
> might motivate doing something separate from any existing virtual
> device, that is dedicated just to SVSM attestation needs.

Bringing another architecture into the picture, if I understand the TDX
philosophy correctly, they would typically expect another enclave on the
same host to provide this kind of service.


>
> Whatever comms mechanism is exposed to the SVSM for early attestation
> I think should only communicate with the hypervisor. A proxy on the
> hypervisor would be responsible for receiving the attestation report
> from SVSM, and opening up a connection to whatever attestation server
> is required, getting the response and feeding it back to SVSM. We
> thus rely entirely on existing hypervisor network stack, and all
> knowledge about the attestation server is isolated from SVSM.
>

So I think that a configurable vsock-based or virtio-serial mechanism in
qemu that facilitates this kind of data exchange might be the right
approach. By "configurable", I mean that you should be able to pass it at
least the IP address to talk to, and possibly some TLS certificates to use
for the communication. And ideally, it should also be able to use a socket
to talk to another VM if that's the setup the guest owner chooses.

As pointed out above, such a mechanism could probably be designed to also be
usable by other architectures.

>
> With regards,
> Daniel


--
Cheers,
Christophe de Dinechin (https://c3d.github.io)
Theory of Incomplete Measurements (https://c3d.github.io/TIM)