Re: [PATCH] arm64/kvm: Add generic v8 KVM target

[Marc Zyngier <marc.zyngier@arm.com>]

On 17/07/15 11:15, Christoffer Dall wrote:
> On Fri, Jul 17, 2015 at 10:56:39AM +0100, Marc Zyngier wrote:
>> On 17/07/15 10:33, Christoffer Dall wrote:
>>> On Fri, Jul 03, 2015 at 11:10:09AM +0100, Marc Zyngier wrote:
>>>> On 03/07/15 10:34, Peter Maydell wrote:
>>>>> On 3 July 2015 at 09:28, Marc Zyngier <marc.zyngier@arm.com> wrote:
>>>>>> On 03/07/15 09:12, Peter Maydell wrote:
>>>>>>> I would still like to see the proponents of this patch say
>>>>>>> what their model is for userspace support of cross-host migration,
>>>>>>> if we're abandoning the model the current API envisages.
>>>>>>
>>>>>> I thought we had discussed this above, and don't really see this as a
>>>>>> departure from the current model:
>>>>>>
>>>>>> - "-cpu host" results in "GENERIC" being used: VM can only be migrated
>>>>>> to the exact same HW (no cross-host migration). MIDR should probably
>>>>>> become RO.
>>>>>> - "-cpu host" results in "A57" (for example): VM can be migrated to a
>>>>>> variety of A57 platforms, and allow for some fuzzing on the revision (or
>>>>>> accept any revision).
>>>>>> - "-cpu a57" forces an A57 model to be emulated, always. It is always
>>>>>> possible to migrate such a VM on any host.
>>>>>>
>>>>>> I think only the first point is new, but the last two are what we have
>>>>>> (or what we should have).
>>>>>
>>>>> Right, but the implicit idea of this GENERIC patch seems to
>>>>> be that new host CPU types don't get their own KVM_ARM_TARGET_*
>>>>> constant, and are thus forever unable to do cross-host migration.
>>>>> It's not clear to me why we'd want to have new CPUs be second
>>>>> class citizens like that.
>>>>
>>>> I certainly don't want to see *any* CPU be a second class citizen. But
>>>> let's face it, we're adding more and more targets that don't implement
>>>> anything new, and just satisfy themselves with the generic implementation.
>>>>
>>>> I see it as an incentive to provide something useful (tables of all the
>>>> registers with default values?) so that cross-host migration becomes a
>>>> reality instead of the figment of our imagination (as it is now). If it
>>>> wasn't already ABI, I'd have removed the existing targets until we have
>>>> something meaningful to put there.
>>>
>>> What we're doing now certainly seems silly, because we're adding kernel
>>> patches without bringing anything to the table...
>>>
>>>>
>>>> Now, I also have my own doubts about cross-host migration (timers
>>>> anyone?). But I don't see the above as a change in policy. More as a way
>>>> to outline the fact that we currently don't have the right level of
>>>> information/infrastructure to support it at all.
>>>>
>>> The one thing that I've lost track of here (sorry) is whether we're
>>> enforcing the inability to do cross-host migration with the generic
>>> target when this patch is merged or do we leave this up to the graces of
>>> userspace?
>>
>> The jury is still out on that one.
>>
>> I was initially not going to enforce anything (after all, this isn't
>> that different from the whole CNTVOFF story where we allow userspace to
>> shoot itself in the foot), but I'm equally happy to make MIDR_EL1
>> read-only if we're creating a generic guest...
>>
> Looking at the code, midr_el1 is already an invariant register, so isn't
> this automagically enforced already?

Ah, you're perfectly right, I has already in that fantasy world where we
can actually migrate VMs across implementations.

> In that case, I'm fine with merging this patch.

Cool. I'll queue that for 4.3.

Thanks,

	M.
-- 
Jazz is not dead. It just smells funny...