Re: [PATCH v3 4/4] KVM: Add flush_on_enter before guest enter

[Paolo Bonzini <pbonzini@redhat.com>]

On 10/11/2017 11:31, Peter Zijlstra wrote:
> On Fri, Nov 10, 2017 at 11:15:06AM +0100, Paolo Bonzini wrote:
>> On 10/11/2017 11:08, Peter Zijlstra wrote:
>>> On Fri, Nov 10, 2017 at 01:49:47AM -0800, Wanpeng Li wrote:
>>>> @@ -2887,7 +2899,7 @@ static void kvm_steal_time_set_preempted(struct kvm_vcpu *vcpu)
>>>>  	if (!(vcpu->arch.st.msr_val & KVM_MSR_ENABLED))
>>>>  		return;
>>>>  
>>>> -	vcpu->arch.st.steal.preempted = KVM_VCPU_PREEMPTED;
>>>> +	vcpu->arch.st.steal.preempted |= KVM_VCPU_PREEMPTED;
>>>
>>> I don't understand this one... If there is concurrency its wrong, if
>>> there is no concurrency it still doesn't make sense as there should not
>>> be any FLUSH flags to preserve..
>>
>> There is no concurrency, foreign VCPUs are not going to write to the
>> location unless PREEMPTED is set.  So indeed the "|=" is pointless.
>>
>> However, I wonder if it'd be useful for a VCPU to set the bit _on
>> itself_ before going to sleep.  Like
>>
>>     set KVM_VCPU_SHOULD_FLUSH
>>     hlt
>>     /* Automagic TLB flush!  */
>>
>> This would not work currently, but if it *is* useful, we should make it
>> work and document it as legal.  Peter, do you think it would make any sense?
> 
> Almost but not quite I think.. So there is no guarantee HLT (or even
> MWAIT with a state that has CPUILDE_FLAG_TLB_FLUSHED set) will actually
> do the TLB flush.

Well, for virt you could always guarantee it if it's useful.  But from
the rest of your message it looks like it would be a separate PV feature
than this one.

Thanks,

Paolo

> And if we preempt the vCPU to run a kernel thread we will not in fact
> invalidate the TLBs either.
> 
> Also, you're confusing the SHOULD_FLUSH with the HAS_FLUSHED concept.
> Because if we didn't flush and we should have we should still issue it
> on VMENTER.
> 
> So if we could somehow tell if a HLT or preemption did indeed flush the
> TLBs post fact (reading back the attained C state is possible but really
> rather expensive IIRC), then we could set a HAS_FLUSHED flag and avoid
> issuing when SHOULD_FLUSH is also set.
> 
>