Re: [patch 4/4] KVM-trace port to tracepoints

[Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>]

* Anthony Liguori (anthony@codemonkey.ws) wrote:
> Hi Mathieu,
>

Hi Anthony,

> It's difficult to review your patches because they aren't inlined.
>
> At any rate, this patches are unusable with SVM.  They try to execute VT 
> instructions unconditionally.  For instance, you changed:
>>
>> -    KVMTRACE_1D(INTR, vcpu, vmcs_read32(VM_EXIT_INTR_INFO), handler);
>> +    trace_kvm_intr(vcpu);
>
> Which lived in VT-specific code (vmx.c)
>
> To:
>
>> +static void probe_kvm_intr(struct kvm_vcpu *vcpu)
>> +{
>> +    kvm_add_trace(KVM_TRC_INTR, vcpu, 1,
>> +        (u32 []){ vmcs_read32(VM_EXIT_INTR_INFO) });
>> +}
>> +
>
> Which lives in common code (kvm_trace.c).  But vmcs_read32() is VT-specific 
> and should not be used in common code so this motion is wrong.  Why not 
> just pass more arguments to probe_kvm_intr()?  Then your first two patches 
> can be dropped completely.
>

Yes, I just noticed that I made a small mistake : the probe code should
actually go in arch/x86/kvm/kvm_trace_probes.c, which is x86-specific.

The reason why I would try to move the vmcs_read32(VM_EXIT_INTR_INFO) to
the probe code is because, unlike the Markers, when a function call with
potential side-effects is put within the arguments, it's really an
argument to a static inline function. In the markers, since it was a
parameter passed to a macro, I could shuffle it into the branch and it
would not be executed when markers were disabled.

However, we don't have this with tracepoints.

kvm-move-register-read-write-to-system-headers.patch becomes
obsolete since I put the probe code in arch/x86/kvm/.

But it would still be required to move vmcs read and encodings to
headers, either to include/asm-x86 or arch/x86/kvm.

Mathieu

> Regards,
>
> Anthony Liguori
>
> Mathieu Desnoyers wrote:

-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68