Re: [Qemu-devel] [PATCH v6 2/2] target: arm: Add support for VCPU event states

[gengdongjiu <gengdongjiu@huawei.com>]

On 2018/8/24 18:38, Peter Maydell wrote:
> On 24 August 2018 at 11:28, gengdongjiu <gengdongjiu@huawei.com> wrote:
>> On 2018/8/24 0:52, Peter Maydell wrote:
>>> On 23 August 2018 at 16:45, Dongjiu Geng <gengdongjiu@huawei.com> wrote:
>>>> +static int kvm_put_vcpu_events(ARMCPU *cpu)
>>>> +{
>>>> +    CPUARMState *env = &cpu->env;
>>>> +    struct kvm_vcpu_events events = {};
>>>> +
>>>> +    if (!kvm_has_vcpu_events()) {
>>>> +        return 0;
>>>> +    }
>>>> +
>>>> +    memset(&events, 0, sizeof(events));
>>>> +    events.exception.serror_pending = env->serror.pending;
>>>> +
>>>> +    if (have_inject_serror_esr) {
>>>> +        events.exception.serror_has_esr = env->serror.has_esr;
>>>> +        events.exception.serror_esr = env->serror.esr;
>>>> +    }
>>>
>>> I realised that the effect of this condition is that
>>> if we migrate a VM from a machine which supports specifying the
>>> SError ESR to one which does not, and at the point of migration
>>> there is a pending SError with an ESR value, then we will
>>> silently drop the specified ESR value. The other alternative
>>> would be to fail the migration (by dropping the if() check,
>>> and letting the host kernel fail the ioctl if that meant that
>>> we asked it to set an SError ESR it couldn't manage.)
>>>
>>> I guess that's OK? It's all hypothetical currently since
>>> we don't support migration between different host CPU types.
>>
>> Peter,
>>    there are two status needed to migrate, one is serror_pending, another is SError ESR value.
>>
>> If A migrates to B, A can set an SError ESR, but B does not support to set.
>> when A is pending a SError and need to migrate to B, I think it should support to migrate the serror_pending status without the ESR value(the ESR value is 0).
>> That is to say,  if A is pending a SError, when migrate to B, B should also pend a SError.
>>
>> or do you think we should refused this migration?
> 
> I don't know, that's why I asked. If we have a pending
> SError with an ESR, and we end up on a destination machine
> where we can pend the SError but not the ESR, does that
> make sense (ie will the guest still be able to usefully
  I agree this suggestion that can pend the SError but without the ESR in the destination machine.

> continue), or have we thrown away information that the
> guest requires to be able to usefully use the SError ?
> Presumably the ESR is important, or we could just never
> bother to set it when pending SErrors.
> 
> thanks
> -- PMM
> 
> .
>