Re: [PATCH] KVM: X86: correct trace_kvm_pv_tlb_flush stats

From: Jon Kohler
Date: Wed May 04 2022 - 20:27:57 EST




> On May 4, 2022, at 5:47 PM, Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
>
> On Wed, May 04, 2022, Jon Kohler wrote:
>> The trace point in record_steal_time() is above the conditional
>> that fires kvm_vcpu_flush_tlb_guest(), so even when we might
>> not be flushing tlb, we still record that we are.
>
> No, it records whether not a TLB flush is needed.

Sure, the trace does, but the stat makes it seem like the host is going
nuts with doing pv tlb flushes when in reality it may not really be
doing all that much work.

>
>> Fix by nestling trace_kvm_pv_tlb_flush() under appropriate
>> conditional. This results in the stats for kvm:kvm_pv_tlb_flush,
>> as trivially observable by perf stat -e "kvm:*" -a sleep Xs, in
>> reporting the amount of times we actually do a pv tlb flush,
>> instead of just the amount of times we happen to call
>> record_steal_time().
>>
>> Signed-off-by: Jon Kohler <jon@xxxxxxxxxxx>
>> ---
>> arch/x86/kvm/x86.c | 4 ++--
>> 1 file changed, 2 insertions(+), 2 deletions(-)
>>
>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
>> index 4790f0d7d40b..8d4e0e58ec34 100644
>> --- a/arch/x86/kvm/x86.c
>> +++ b/arch/x86/kvm/x86.c
>> @@ -3410,9 +3410,9 @@ static void record_steal_time(struct kvm_vcpu *vcpu)
>>
>> vcpu->arch.st.preempted = 0;
>>
>> - trace_kvm_pv_tlb_flush(vcpu->vcpu_id,
>> - st_preempted & KVM_VCPU_FLUSH_TLB);
>> if (st_preempted & KVM_VCPU_FLUSH_TLB)
>> + trace_kvm_pv_tlb_flush(vcpu->vcpu_id,
>> + st_preempted & KVM_VCPU_FLUSH_TLB);
>
> If you're going to trace only when a flush is needed, this should simply be:
>
> trace_kvm_pv_tlb_flush(vcpu->vcpu_id);
>
> I haven't used this tracepoint often (at all?) so I don't have a strong preference,
> but I can see the "no TLB flush needed" information being extremely valuable when
> debugging a supsected TLB flushing bug.

Yea thats fair; however, this is only calling into some other function that is
actually doing the work. Those other flush TLB areas do not have traces
AFAIK, so even that is a bit in complete.

The net problem here is really that the stat is likely incorrect; however,
one other oddity I didn’t quite understand after looking into this is that
the call site for all of this is in record_steal_time(), which is only called
from vcpu_enter_guest(), and that is called *after*
kvm_service_local_tlb_flush_requests(), which also calls
kvm_vcpu_flush_tlb_guest() if request == KVM_REQ_TLB_FLUSH_GUEST

That request may be there set from a few different places.

I don’t have any proof of this, but it seems to me like we might have a
situation where we double flush?

Put another way, I wonder if there is any sense behind maybe hoisting
if (kvm_check_request(KVM_REQ_STEAL_UPDATE, vcpu)) up before
Other tlb flushes, and have it clear the FLUSH_GUEST if it was set?

Just thinking aloud, seemed fishy.

Regardless, this pv tlb flush stat needs some love, open to suggestions
on the most appropriate way to tackle it?

>
>> kvm_vcpu_flush_tlb_guest(vcpu);
>>
>> if (!user_access_begin(st, sizeof(*st)))
>> --
>> 2.30.1 (Apple Git-130)