Re: [PATCH 1/6] KVM: x86: add VCPU stat for KVM_REQ_EVENT processing

[Paolo Bonzini]

On 20/12/2016 10:21, Roman Kagan wrote:
> On Mon, Dec 19, 2016 at 10:47:13AM +0100, Paolo Bonzini wrote:
>> Signed-off-by: Paolo Bonzini <pbonzini@xxxxxxxxxx>
>> ---
>>  arch/x86/include/asm/kvm_host.h | 1 +
>>  arch/x86/kvm/x86.c              | 2 ++
>>  2 files changed, 3 insertions(+)
>>
>> diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
>> index 994d8ed9fc6c..08cfd45a9452 100644
>> --- a/arch/x86/include/asm/kvm_host.h
>> +++ b/arch/x86/include/asm/kvm_host.h
>> @@ -850,6 +850,7 @@ struct kvm_vcpu_stat {
>>  	u64 hypercalls;
>>  	u64 irq_injections;
>>  	u64 nmi_injections;
>> +	u64 req_event;
>>  };
>>  
>>  struct x86_instruction_info;
>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
>> index e95d94edbdc3..e9b512090865 100644
>> --- a/arch/x86/kvm/x86.c
>> +++ b/arch/x86/kvm/x86.c
>> @@ -180,6 +180,7 @@ struct kvm_stats_debugfs_item debugfs_entries[] = {
>>  	{ "insn_emulation_fail", VCPU_STAT(insn_emulation_fail) },
>>  	{ "irq_injections", VCPU_STAT(irq_injections) },
>>  	{ "nmi_injections", VCPU_STAT(nmi_injections) },
>> +	{ "req_event", VCPU_STAT(req_event) },
>>  	{ "mmu_shadow_zapped", VM_STAT(mmu_shadow_zapped) },
>>  	{ "mmu_pte_write", VM_STAT(mmu_pte_write) },
>>  	{ "mmu_pte_updated", VM_STAT(mmu_pte_updated) },
>> @@ -6691,6 +6692,7 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
>>  	}
>>  
>>  	if (kvm_check_request(KVM_REQ_EVENT, vcpu) || req_int_win) {
>> +		++vcpu->stat.req_event;
>>  		kvm_apic_accept_events(vcpu);
>>  		if (vcpu->arch.mp_state == KVM_MP_STATE_INIT_RECEIVED) {
>>  			r = 1;
> 
> Just curious what kind of information you expect to be able to extract
> from this counter?  I mean, I'm not opposed to introducing it, I just
> want to figure out how to use it properly.  Compare against
> irq_injections?

Yes, exactly.  Since vmexit cycle counts are always a bit noisy, I could
compare irq_injections against req_event and get an idea of the expected
improvement.

Also for example the stat shows that the sti_hlt case has a higher # of
req_event than sti_nop.  I have a patch to fix that, but it's left for
later because I don't know of a workload that triggers it.

I don't really need this patch of course.

> Also I guess it may be interesting to count branching to
> cancel_injection label in vcpu_enter_guest.

Interesting one too.

Paolo