[PATCH 3/4] conf: add s390-pv as launch security type

Thu May 20 15:26:36 UTC 2021

On 5/20/21 12:10 PM, Boris Fiuczynski wrote:
> On 5/20/21 2:08 PM, Daniel Henrique Barboza wrote:
>>
>>
>> On 5/20/21 4:05 AM, Boris Fiuczynski wrote:
>>> On 5/19/21 9:34 PM, Daniel Henrique Barboza wrote:
>>>>
>>>>
>>>> On 5/19/21 2:40 PM, Boris Fiuczynski wrote:
>>>>> Add launch security type 's390-pv' as well as some tests.
>>>>>
>>>>> Signed-off-by: Boris Fiuczynski <fiuczy at linux.ibm.com>
>>>>> ---
>>>>>   docs/schemas/domaincommon.rng                 |  1 +
>>>>>   src/conf/domain_conf.c                        |  8 +++++
>>>>>   src/conf/domain_conf.h                        |  1 +
>>>>>   src/qemu/qemu_command.c                       | 26 ++++++++++++++
>>>>>   src/qemu/qemu_namespace.c                     |  1 +
>>>>>   src/qemu/qemu_process.c                       |  1 +
>>>>>   src/qemu/qemu_validate.c                      |  8 +++++
>>>>>   .../launch-security-s390-pv-ignore-policy.xml | 24 +++++++++++++
>>>>>   .../launch-security-s390-pv.xml               | 18 ++++++++++
>>>>>   .../launch-security-s390-pv-ignore-policy.xml |  1 +
>>>>>   tests/genericxml2xmltest.c                    |  2 ++
>>>>>   ...ty-s390-pv-ignore-policy.s390x-latest.args | 35 +++++++++++++++++++
>>>>>   .../launch-security-s390-pv-ignore-policy.xml | 33 +++++++++++++++++
>>>>>   .../launch-security-s390-pv.s390x-latest.args | 35 +++++++++++++++++++
>>>>>   .../launch-security-s390-pv.xml               | 30 ++++++++++++++++
>>>>>   tests/qemuxml2argvtest.c                      |  3 ++
>>>>>   16 files changed, 227 insertions(+)
>>>>>   create mode 100644 tests/genericxml2xmlindata/launch-security-s390-pv-ignore-policy.xml
>>>>>   create mode 100644 tests/genericxml2xmlindata/launch-security-s390-pv.xml
>>>>>   create mode 120000 tests/genericxml2xmloutdata/launch-security-s390-pv-ignore-policy.xml
>>>>>   create mode 100644 tests/qemuxml2argvdata/launch-security-s390-pv-ignore-policy.s390x-latest.args
>>>>>   create mode 100644 tests/qemuxml2argvdata/launch-security-s390-pv-ignore-policy.xml
>>>>>   create mode 100644 tests/qemuxml2argvdata/launch-security-s390-pv.s390x-latest.args
>>>>>   create mode 100644 tests/qemuxml2argvdata/launch-security-s390-pv.xml
>>>>>
>>>>> diff --git a/docs/schemas/domaincommon.rng b/docs/schemas/domaincommon.rng
>>>>> index 3df13a0cf1..7c92e4c812 100644
>>>>> --- a/docs/schemas/domaincommon.rng
>>>>> +++ b/docs/schemas/domaincommon.rng
>>>>> @@ -485,6 +485,7 @@
>>>>>         <attribute name="type">
>>>>>           <choice>
>>>>>             <value>sev</value>
>>>>> +          <value>s390-pv</value>
>>>>>           </choice>
>>>>>         </attribute>
>>>>>         <interleave>
>>>>
>>>> You added a new 's390-pv' security type, but down there you're using
>>>> the new confidential-guest-support feature from QEMU 6.0 which is also
>>>> valid for AMD and pSeries. I think you can do a little change in the idea
>>>> of these patches while keeping most of it. Instead of calling this new
>>>> support 's390-pv', call it 'confidential-guest-support' or 'CGS'.
>>>>
>>>> My reasoning is that the QEMU community (namely David Gibson, qemu-ppc
>>>> maintainer) went into a lot of discussions back and forth to develop the
>>>> confidential-guest-support machine option, based on what was at first AMD-SEV
>>>> specific code, with the intention of make it easier for users to enable
>>>> secure guests across machine types. I believe Libvirt should follow suit
>>>> and do the same - a single option to enable secure guest supports for
>>>> all guests, with any differences in the support being handled by each arch
>>>> deep down in the driver.
>>>>
>>>> Otherwise, what will end up happening is that when someone (probably myself)
>>>> come along with the secure guest support for pSeries (PEF), I will need to
>>>> create yet another launch type 'ppc64-pef' to do basically the same thing you're
>>>> already doing for s390x, which is adding '-machine confidential-guest-support=<>'
>>>> in the QEMU command line. Same thing with AMD SEV, and with any other
>>>> arch that QEMU might support with the confidential-guest-support option. We're
>>>> going to add extra XML parsing code and docs to handle the same thing.
>>>>
>>>> Note that I'm not asking you to go ahead and implement the Libvirt support for
>>>> all the 3 archs. What I'm asking is to change the name of the launch security
>>>> type in the domain XML and docs to reflect that this will be the same type
>>>> that all other archs that has confidential-guest-support will end up using.
>>>>
>>>>
>>>> Thanks,
>>>>
>>>>
>>>> Daniel
>>>
>>> Daniel,
>>> thanks for your review and feedback.
>>> When I looked at the QEMU commit 590466f056c4f
>>> https://git.qemu.org/?p=qemu.git;a=commit;h=590466f056c4f2a7ff87ed751cece4f4ff02fd57
>>> I did not get the impression that there is a common type for confidential guest support possible as the requiered data per type differs. Also I got the impression that the different types are not necessarily architecture bound. I may have gotten the wrong impression.
>>
>>
>> The point I tried to make is that, although there are different types of secure
>> VMs in QEMU, this doesn't mean that we need different launch security types in
>> Libvirt since they're all tied in the same QEMU machine option. If you check
>> the QEMU confidential-guest-support docs:
>>
>> https://gitlab.com/qemu-project/qemu/-/blob/master/docs/confidential-guest-support.txt
>>
>> ----
>> To run a confidential guest you need to add two command line parameters:
>>
>> 1. Use "-object" to create a "confidential guest support" object.  The
>>     type and parameters will vary with the specific mechanism to be
>>     used
>> 2. Set the "confidential-guest-support" machine parameter to the ID of
>>     the object from (1).
>> ----
>>
>> This is true for all 3 currently supported types. For s390x PV:
>>
>>      -object s390-pv-guest,id=pv0 \
>>      -machine confidential-guest-support=pv0 \
>>
>> AMD SEV:
>>
>>      -object sev-guest,id=sev0,cbitpos=47,reduced-phys-bits=1 \
>>      -machine confidential-guest-support=sev0 \
>>
>> PAPR PEF:
>>
>>      -object pef-guest,id=pef0 \
>>      -machine confidential-guest-support=pef0 \
>>
>>
> 
> All the above requires libvirt code (capability checks, validation, command line generation etc) to exist per individual type that you try to hide from the user.
> Below the user is required to know that when he wants to use SEV he has to specify 'confidential-guest-support' as type AND in addition is also required to provide a policy for SEV. So it is not transparent for users maybe even more confusing since the old type sev must not be eliminated for backward compatibility reasons.

Yeah, I'm not sure how to handle the existing AMD SEV support with the newest
stuff from QEMU. Whether a new launcher would be created while keeping the
existing one with the old approach, or the existing launcher would be
re-utilized with the confidential-guest-support option for new domains,
or whatever else.

> 
> So I agree that there is one "knob" but you need to potentially provide different data and pre-checks when "pressing" it... just looking at sev, pef and pv and not knowing what Intel's TDX requires.
> 
> But I agree with you that at it's core its a matter of taste/design.

I'm still more fond of my idea (perhaps I have a persona bias going on since
that would cut my work short for the papr-pef support :D), but I'm fine with
your approach of one launcher for each technology as well. I'd say we can
wait for more people to comment on this and go with your approach if no
one else have a strong option about it.

Thanks,

Daniel

> 
>> The difference between them, aside from the object types, is that AMD has
>> configurable arguments to deal with. So my proposal is to, instead of creating
>> a s390-pv launch type, you create a confidential-guest-support type:
>>
>>    <launchSecurity type='confidential-guest-support' />
>>
>> This will translate, for s390x guests, into s390-pv. All validations and
>> constraints specific to s390-pv applies. This same XML for a PAPR guest will
>> trigger papr-pef code, and AMD-SEV code for AMD guests. AMD will probably want
>> to add more arguments, so they would want to extend it like it is already
>> done with their current 'sev' launcher:
>>
>>    <launchSecurity type='confidential-guest-support'>
>>      <policy>0x0001</policy>
>>      <cbitpos>47</cbitpos>
>>      <reducedPhysBits>1</reducedPhysBits>
>>      <dhCert>RBBBSDDD=FDDCCCDDDG</dhCert>
>>      <session>AAACCCDD=FFFCCCDSDS</session>
>>    </launchSecurity>
>>
>>
>> The advantage is that the same launchSecurity type will enable
>> confidential-guest-support for everyone, sparing a bit of XML parsing and
>> docs we would have if each arch adds a new launch type that maps to its own
>> secure VM technology. Note that you can (and should) clarify in the
>> docs that the confidential-guest-support for s390x guests is the s390-pf
>> technology and mention all applicable features/limitations of it. This also
>> will make for an easier time for other Libvirt users that has to deal with
>> multiple archs (not sure there are many, but ...) since the same option enables
>> secure VM for all archs.
>>
>>
>> I think you can argue that it's a matter of taste/design and that we should
>> instead add a launcherSecurity type for each secure VM technology, regardless
>> of how QEMU see all of them as a 'secure VM knob'. I would like to hear
>> other opinions (specially from AMD) about what they feel about it. I'd
>> rather have the same launcherSecurity type for everyone.
>>
>>
>>
>> Thanks,
>>
>>
>> Daniel
>>
>>
>>
>>
>>>
>>
> 
> 