Re: [PATCH RFC V4 3/3] kvm: Choose better candidate for directed yield

[Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>]

On 07/17/2012 01:59 PM, Avi Kivity wrote:
> On 07/16/2012 07:10 PM, Rik van Riel wrote:
>> On 07/16/2012 06:07 AM, Avi Kivity wrote:
>>
>>>> +{
>>>> +    bool eligible;
>>>> +
>>>> +    eligible = !vcpu->ple.cpu_relax_intercepted ||
>>>> +            (vcpu->ple.cpu_relax_intercepted&&
>>>> +             vcpu->ple.dy_eligible);
>>>> +
>>>> +    if (vcpu->ple.cpu_relax_intercepted)
>>>> +        vcpu->ple.dy_eligible = !vcpu->ple.dy_eligible;
>>>
>>> Probably should assign 'true', since the previous value is essentially
>>> random.
>>
>> I suspect the intended purpose of this conditional is to
>> flip the eligibility of a vcpu for being selected as a
>> direct yield target.
>>
>> In other words, that bit of the code is correct.
>
> If vcpu A is in a long spin loop and is preempted away, and vcpu B dips
> several times in kvm_vcpu_on_spin(), then it will act as intended.

Yes, true.

But
> if vcpu A is spinning for x% of its time and processing on the other,
> then vcpu B will flip its dy_eligible for those x%, and not flip it when
> it's processing.  I don't understand how this is useful.

Suppose A is doing really good job and and has not done pause loop
exit, we will not touch it's dy_eligible flag. Also dy_eligible flag
will not prevent B doing yield_to to A.

Suppose A has started spinning in the beginning itself, it will do pause 
loop exit if it crosses threshold, and we will now start toggling
dy_eligible.

Was that you were referring?

And it seems we may still have to set dy_eligible flag to false at the 
beginning of vcpu_on_spin along with cpu_relax_intercepted = true, like 
below, so that we do not have spill-over status from previous PL exits.

vcpu_on_spin()
{
  cpu_relax_intercepted = true;
  dy_eligible = false;
.
.
.

cpu_relax_intercepted = false;
}

Let me know if that addresses your concern.

>
> I guess this is an attempt to impose fairness on yielding, and it makes
> sense to do this, but I don't know if this is the best way to achieve it.
>