[PATCH RFC] KVM: arm64: vgic: Decouple the check of the EnableLPIs bit from the ITS LPI translation

[Shenming Lu]

On 2020/12/31 20:22, Marc Zyngier wrote:
> On 2020-12-31 11:58, Shenming Lu wrote:
>> On 2020/12/31 16:57, Marc Zyngier wrote:
>>> Hi Shemming,
>>>
>>> On 2020-12-31 06:28, Shenming Lu wrote:
>>>> When the EnableLPIs bit is set to 0, any ITS LPI requests in the
>>>> Redistributor would be ignored. And this check is independent from
>>>> the ITS LPI translation. So it might be better to move the check
>>>> of the EnableLPIs bit out of the LPI resolving, and also add it
>>>> to the path that uses the translation cache.
>>>
>>> But by doing that, you are moving the overhead of checking for
>>> EnableLPIs from the slow path (translation walk) to the fast
>>> path (cache hit), which seems counter-productive.
>>
>> Oh, I didn't notice the overhead of the checking, I thought it would
>> be negligible...
> 
> It probably doesn't show on a modern box, but some of the slower
> systems might see it. Overall, this is a design decision to keep
> the translation cache as simple and straightforward as possible:
> if anything affects the output of the cache, we invalidate it,
> and that's it.

Ok, get it.

> 
>>
>>>
>>>> Besides it seems that
>>>> by this the invalidating of the translation cache caused by the LPI
>>>> disabling is unnecessary.
>>>>
>>>> Not sure if I have missed something... Thanks.
>>>
>>> I am certainly missing the purpose of this patch.
>>>
>>> The effect of EnableLPIs being zero is to drop the result of any
>>> translation (a new pending bit) on the floor. Given that, it is
>>> immaterial whether this causes a new translation or hits in the
>>> cache, as the result is still to not pend a new interrupt.
>>>
>>> I get the feeling that you are trying to optimise for the unusual
>>> case where EnableLPIs is 0 *and* you have a screaming device
>>> injecting tons of interrupt. If that is the case, I don't think
>>> this is worth it.
>>
>> In fact, I just found (imagining) that if the EnableLPIs bit is 0,
>> the kvm_vgic_v4_set_forwarding() would fail when performing the LPI
>> translation, but indeed we don't try to pend any interrupts there...
>>
>> By the way, it seems that the LPI disabling would not affect the
>> injection of VLPIs...
> 
> Yes, good point. We could unmap the VPE from all ITS, which would result
> in all translations to be discarded, but this has the really bad side
> effect of *also* preventing the delivery of vSGIs, which isn't what
> you'd expect.
> 
> Overall, I don't think there is a good way to support this, and maybe
> we should just prevent EnableLPIs to be turned off when using direct
> injection. After all, the architecture does allow that for GICv3
> implementations, which is what we emulate.

Agreed, if there is no good way, we could just make the EnableLPIs clearing
unsupported...

Thanks(Happy 2021),
Shenming

> 
> Thanks,
> 
>         M.