[PATCH RFC 1/1] qemu: capabilities: disable csske for host cpu

Fri Mar 11 15:24:22 UTC 2022

Am 11.03.22 um 15:56 schrieb Daniel P. Berrangé:
> On Fri, Mar 11, 2022 at 03:52:57PM +0100, Christian Borntraeger wrote:
>>
>>
>> Am 11.03.22 um 14:08 schrieb Daniel P. Berrangé:
>>> On Fri, Mar 11, 2022 at 12:37:46PM +0000, Daniel P. Berrangé wrote:
>>>> On Fri, Mar 11, 2022 at 01:12:35PM +0100, Christian Borntraeger wrote:
>>>>>
>>>>>
>>>>> Am 11.03.22 um 10:23 schrieb David Hildenbrand:
>>>>>> On 11.03.22 10:17, Daniel P. Berrangé wrote:
>>>>>>> On Thu, Mar 10, 2022 at 11:17:38PM -0500, Collin Walling wrote:
>>>>>>>> CPU models past gen16a will no longer support the csske feature. In
>>>>>>>> order to secure migration of guests running on machines that still
>>>>>>>> support this feature to machines that do not, let's disable csske
>>>>>>>> in the host-model.
>>>>>>
>>>>>> Sorry to say, removing CPU features is a no-go when wanting to guarantee
>>>>>> forward migration without taking care about CPU model details manually
>>>>>> and simply using the host model. Self-made HW vendor problem.
>>>>>
>>>>> And this simply does not reflect reality. Intel and Power have removed TX
>>>>> for example. We can now sit back and please ourselves how we live in our
>>>>> world of dreams. Or we can try to define an interface that deals with
>>>>> reality and actually solves problems.
>>>>
>>>> This proposal wouldn't have helped in the case of Intel removing
>>>> TSX, because it was removed without prior warning in the middle
>>>> of the product lifecycle. At that time there were already millions
>>>> of VMs in existance using the removed feature.
>>>>
>>>>>>> The problem scenario you describe is the intended semantics of
>>>>>>> host-model though. It enables all features available in the host
>>>>>>> that you launched on. It lets you live migrate to a target host
>>>>>>> with the same, or a greater number of features. If the target has
>>>>>>> a greater number of features, it should restrict the VM to the
>>>>>>> subset of features that were present on the original source CPU.
>>>>>>> If the target has fewer features, then you simply can't live
>>>>>>> migrate a VM using host-model.
>>>>>>>
>>>>>>> To get live migration in both directions across CPUs with differing
>>>>>>> featuresets, then the VM needs to be configured with a named CPU
>>>>>>> model that is a subset of both, rather than host-model.
>>>>>>
>>>>>> Right, and cpu-model-baseline does that job for you if you're lazy to
>>>>>> lookup the proper model.
>>>>>
>>>>> Yes baseline will work, but this requires tooling like openstack. The normal
>>>>> user will just use the default and this is host-model.
>>>>>
>>>>> Let me explain the usecase for this feature. Migration between different versins
>>>>> baseline: always works
>>>>> host-passthrough: you get what you deserve
>>>>> default model: works
>>>>> We have disabled CSSKE from our default models (-cpu gen15a will not present csske).
>>>>> So that works as well.
>>>>> host-model: Also works for all machines that have csske.
>>>>> Now: Lets say gen17 will no longer support this. That means that we can not migrate
>>>>> host-model from gen16 or gen15 because those will have csske.
>>>>> What options do we have? If we disable csske in the host capabilities that would mean
>>>>> that a host compare against an xml from an older QEMU would fail (even if you move
>>>>> from gen14 to gen14). So this is not a good option.
>>>>>
>>>>> By disabling deprecated features ONLY for the _initial_ expansion of model-model, but
>>>>> keeping it in the host capabilities you can migrate existing guests (with the
>>>>> feature) as we only disable in the expansion, but manually asking for it still works.
>>>>> AND it will allow to move this instantiation of the guest to future machines without
>>>>> the feature. Basically everything works.
>>>>
>>>> The change you proposal works functionally, but none the less it is
>>>> changing the semantics of host-model. It is defined to expose all the
>>>> features in the host, and the proposal changes yet. If an app actually
>>>> /wants/ to use the deprecated feature and it exists in the host, then
>>>> host-model should be allowing that as it does today.
>>>>
>>>> The problem scenario you describe is ultimately that OpenStack does
>>>> not have a future proof default CPU choice. Libvirt and QEMU provide
>>>> a mechanism for them to pick other CPU models that would address the
>>>> problem, but they're not using that. The challenge is that OpenStack
>>>> defaults currently are a zero-interaction thing.
>>>>
>>>> They could retain their zero-interaction defaults, if at install time
>>>> they queried the libvirt capabilities to learn which named CPU models
>>>> are available, whereupon they could decide to use gen15a.  The main
>>>> challenge here is that the list of named CPU models is an unordered
>>>> set, so it is hard to programatically figure out which of the available
>>>> named CPU models is the newest/best/recommended.
>>>>
>>>> IOW, what's missing is a way for apps to easily identify that 'gen15a'
>>>> is the best CPU to use on the host, without needing human interaction.
>>>
>>> I think this could be solved with a change to query-cpu-definitions
>>> in QEMU, to add an extra 'recommended: bool' attribute to the
>>> CpuDefinitionInfo struct.  This would be defined to be only set for
>>> 1 CPU model in the list, and would reflect the recommended CPU model
>>> given the current version of QEMU, kernel and hardware. Or we could
>>> allow 'recommended' to be set for more than 1 CPU, provided we define
>>> an explicit ordering of returned CPU models.
>>
>> I like the recommended: bool attribute. It should provide what we need.
>>
>> Would you then also suggest to use this for host-model or only for a new
>> type like "host-recommended" ?
> 
> Neither of those. Libvirt would simply report this attribute in
> the information it exposes about CPUs.
> 
> OpenStack would explicitly extract this and set it in the XML
> for the guest, so that each guest's view of "recommended" is
> fixed from the time that guest is first created, rather than
> potentially changing on each later boots.

Openstack is one thing, but I think this flag would really be useful
for instantiation without open stack.