Re: [Qemu-devel] [PATCH v3 8/8] tpm: Added support for TPM emulator

[Stefan Berger <stefanb@linux.vnet.ibm.com>]

On 05/03/2017 07:29 AM, Daniel P. Berrange wrote:
> On Wed, May 03, 2017 at 11:24:42AM +0000, Marc-André Lureau wrote:
>> Hi
>>
>> On Wed, May 3, 2017 at 3:17 PM Dr. David Alan Gilbert <dgilbert@redhat.com>
>> wrote:
>>
>>> * Daniel P. Berrange (berrange@redhat.com) wrote:
>>>> On Tue, May 02, 2017 at 03:35:48PM -0400, Stefan Berger wrote:
>>>>> On 05/02/2017 02:50 PM, Marc-André Lureau wrote:
>>>>>> Hi
>>>>>>
>>>>>> On Tue, May 2, 2017 at 10:25 PM Patrick Ohly <patrick.ohly@intel.com
>>>>>> <mailto:patrick.ohly@intel.com>> wrote:
>>>>>>
>>>>>>      On Tue, 2017-05-02 at 13:19 -0400, Stefan Berger wrote:
>>>>>>      > On 05/02/2017 01:09 PM, Marc-André Lureau wrote:
>>>>>>      > > On Tue, May 2, 2017 at 8:59 PM Stefan Berger
>>>>>>      <stefanb@linux.vnet.ibm.com <mailto:stefanb@linux.vnet.ibm.com>>
>>>>>>      > > wrote:
>>>>>>      > >
>>>>>>      > >> And who is going to implement that qemu-swtpm? Obviously
>>> this
>>>>>>      discussion
>>>>>>      > >> doesn't contribute to progress if nobody is doing that in
>>> the
>>>>>>      end.
>>>>>>      > >>
>>>>>>      > > The same persons who try to push for that emulated TPM code.
>>>>>>      The easiest
>>>>>>      > > approach would be to copy/adapt the swtpm code in qemu, if
>>> the
>>>>>>      licence is
>>>>>>      > > compatible. I can help with that if there is a consensus it's
>>>>>>      a better
>>>>>>      > > approach.
>>>>>>      >
>>>>>>      >
>>>>>>      > It's a matter of time and at least I don't have time for that.
>>>>>>
>>>>>>      Neither do I, and nor (I believe) does Amarnath. The approach
>>> with
>>>>>>      using
>>>>>>      the existing swtpm project seemed attractive to us exactly
>>> because it
>>>>>>      avoids having to write and maintain more than just the glue code
>>>>>>      between
>>>>>>      the two projects.
>>>>>>
>>>>>>
>>>>>> The main argument is not about having more or less code in qemu to
>>>>>> maintain, but yes this is a concern (although giving up that
>>> maintenance
>>>>>> to a seperate project with mostly Stefan-alone isn't a much better
>>>>>> alternative). btw, is the project actually used by something else
>>> than
>>>>>> qemu? (I am not talking about developpers/testing). If not, then it
>>>>>> makes sense to make it part of qemu.
>>>>> The intention would be to use it for RunC as well (plus higher layers
>>>>> afterwards): https://github.com/opencontainers/runc/pull/1082
>>>>>
>>>>>> But it's mostly a technical reason, to avoid having to rely on a
>>> foreign
>>>>>> protocol and project with all the compatibility constrains.
>>>>> I understand. Ideally swtpm-0.1 would be equivalent to 1.0 with all
>>> features
>>>>> available and no further protocol extensions necessary. In practice
>>> that may
>>>>> look different.
>>>>>
>>>>>> In the end, we may decide to start with a separate project, and
>>> change
>>>>>> it in the future if it's problematic (that would break some cases,
>>> such
>>>>>> as being able to freely switch the helper). Tbh, I am not so happy
>>> with
>>>>>> the code quality of swtpm, and I haven't looked closely at libtpms.
>>>>>> Having a qemu-swtpm as part of qemu would probably help improve it
>>> too,
>>>>>> and bring a few more developers for maintainance...
>>>>> libtpms combines a few source codes with some glue around it. The
>>> coding
>>>>> style is different for TPM 1.2 and TPM 2 code for example and the code
>>> bases
>>>>> are in the 10s of thousands of line. In the case of TPM 2 it 'lives
>>> from'
>>>>> TCG code drops and thus there is no reformatting of source code etc.
>>>>>
>>>>> If someone wants to get started on qemu-swtpm that's certainly cool
>>> but over
>>>>> the years it's just been quite difficult to find developers for it to
>>> share
>>>>> the burden. All that said, someone should state whether this series is
>>> a go
>>>>> or no-go because of the external project it requires.
>>>> I think it is *good* that it uses the external swtpm project and do not
>>>> want to see it reimplemented inside QEMU, particularly with the interest
>>>> for swtpm to be used in container contexts via RunC. Such common
>>> infrastructure
>>>> for both containers & QEMU will be important given the increasing
>>> convergance
>>>> of technology across containers & VMs.
>>> I agree; there aren't that many people who understand the details of TPMs,
>>> reimplementing one in QEMU isn't something you'd want to do.
>>>
>> It's not about reimplementing TPM emulation. swtpm is a small utility
>> talking to libtpms doing the heavy work. swtpm could quite easily be
>> copied/adapted to fit in qemu if it's only meant to be used by qemu.
>> However, it seems the helper is going to be used by other projects, so it
>> make sense to make it a seperate project. Nevertheless, I think we should
>> carefully review the protocol and the compatibility situation before
>> committing this work, it's a major burden ahead..
> Yes, I agree that reviewing the protocol is a very neccessary step, to
> ensure it is going to be forwards compatible in a manner suitable for
> QEMU.
>
> In particular we need to understand what, if any, relationship there
> needs to be wrt machine types & live migration stability. eg if a
> VM is booted and the protocol negotiates version X, and we then live
> migrate to a host with a newer swtpm, we need to ensure that the
> protocol doesn't negotiate something that leads to guest OS incompatiblity
> problems in accessing the TPM.
>
> basically similar scenario to that which we have had with vhost-user
> and version compatibility with external vhost-user server implementations
> across live migration.

Ok, please have a look at the protocol.

As for the state of the TPM implementations:
  - the TPM 1.2 part is stable and besides openssl changes in the future 
I would not expect any more changes to the core code. That means that 
the state blobs the code is writing out and we are migrating are 
'stable'. They are written in big endian format just like any other QEMU 
device and thus are migratable between endianess.
  - the TPM 2 part , as stated before, is still somewhat in flux. I am 
not sure when there will be a final TPM 2 from TCG. There the 
possibility exists that the state blobs the TPM 2 is writing out still 
change. I have added a version tag in front of the blobs so in case 
something else gets added that that can be accommodated. Besides that 
it's also adapted to write the state blobs in big endian format for the 
same reason as above. Maybe at some point I'll just freeze the code and 
don't follow the ongoing TPM 2 development anymore besides bug fixes to 
exsting code, which then freezes the state blobs as well.

The issue on the swtpm and QEMU side then is the protocols to transfer 3 
opaque (possibly encrypted) swtpm state blobs when suspending/resuming 
or migrating the swtpm. I currently have 1 command to pull them out by 
their 3 identifies (so run that command 3 times) and 1 commands to put 
them back in by their identifier. When is comes to migration, this is 
probably one of the most critical parts to look at.

    Stefan

>
> Regards,
> Daniel