Re: [PATCH 2/2] KVM: x86: optimize delivery of TSC deadline timer interrupt

[Paolo Bonzini <pbonzini@redhat.com>]

On 06/02/2015 21:51, Marcelo Tosatti wrote:
> On Fri, Feb 06, 2015 at 01:16:59PM +0100, Paolo Bonzini wrote:
>> The newly-added tracepoint shows the following results on
>> the tscdeadline_latency test:
>>
>>         qemu-kvm-8387  [002]  6425.558974: kvm_vcpu_wakeup:      poll time 10407 ns
>>         qemu-kvm-8387  [002]  6425.558984: kvm_vcpu_wakeup:      poll time 0 ns
>>         qemu-kvm-8387  [002]  6425.561242: kvm_vcpu_wakeup:      poll time 10477 ns
>>         qemu-kvm-8387  [002]  6425.561251: kvm_vcpu_wakeup:      poll time 0 ns
>>
>> and so on.  This is because we need to go through kvm_vcpu_block again
>> after the timer IRQ is injected.  Avoid it by polling once before
>> entering kvm_vcpu_block.
>>
>> On my machine (Xeon E5 Sandy Bridge) this removes about 500 cycles (7%)
>> from the latency of the TSC deadline timer.
>>
>> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
>> ---
>>  arch/x86/kvm/x86.c | 14 +++++++++-----
>>  1 file changed, 9 insertions(+), 5 deletions(-)
>>
>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
>> index 0b8dd13676ef..1e766033ebff 100644
>> --- a/arch/x86/kvm/x86.c
>> +++ b/arch/x86/kvm/x86.c
>> @@ -6389,11 +6389,15 @@ static inline int __vcpu_run(struct kvm *kvm, struct kvm_vcpu *vcpu)
>>  	    !vcpu->arch.apf.halted)
>>  		return vcpu_enter_guest(vcpu);
>>  
>> -	srcu_read_unlock(&kvm->srcu, vcpu->srcu_idx);
>> -	kvm_vcpu_block(vcpu);
>> -	vcpu->srcu_idx = srcu_read_lock(&kvm->srcu);
>> -	if (!kvm_check_request(KVM_REQ_UNHALT, vcpu))
>> -		return 1;
>> +	if (kvm_arch_vcpu_runnable(vcpu))
>> +		clear_bit(KVM_REQ_UNHALT, &vcpu->requests);
>> +	else {
> 
> Why the clear_bit? Since only kvm_vcpu_block in the below section
> sets it, and that section clears it as well.

You're right.

> Can remove another 300 cycles from do_div when programming LAPIC
> tscdeadline timer.

Do you mean using something like lib/reciprocal_div.c?  Good idea,
though that's not latency, it's just being slow. :)

Paolo