2016-03-07 16:36+0100, Paolo Bonzini:
> On 04/03/2016 21:46, Suravee Suthikulpanit wrote:
>> +static void svm_deliver_avic_intr(struct kvm_vcpu *vcpu, int vec)
>> +{
>> + struct vcpu_svm *svm = to_svm(vcpu);
>> +
>> + kvm_lapic_set_vector(vec, avic_get_bk_page_entry(svm, APIC_IRR));
(I think that smp_mb here would make sense, even though we're fine now
thanks to re-checking vcpu->mode in kvm_vcpu_kick.
A comment explaining this optimization would be nice. I'm thinking
about a race where we don't send the doorbell even though the VCPU is
in guest mode, because vcpu->mode was read before writing APIC_IRR.)
>> +
>> + if (vcpu->mode == IN_GUEST_MODE) {
>> + wrmsrl(SVM_AVIC_DOORBELL,
>> + __default_cpu_present_to_apicid(vcpu->cpu));
>> + } else {
>> + kvm_vcpu_kick(vcpu);
>> + }
>
> You also need to add
>
> kvm_make_request(KVM_REQ_EVENT, vcpu);
>
> before the "if", similar to vmx_deliver_posted_interrupt.
KVM won't do anything in KVM_REQ_EVENT and I think that the request can
be avoided because KVM already has to handle IRR writes from AVIC.
And what about
[...]
else if (!vcpu->...->is_running)
kvm_vcpu_kick(vcpu);
?
The kick isn't needed unless the VCPU is scheduled out.
Or maybe just
if (vcpu->...->is_running)
wrmsrl()
else
kvm_vcpu_kick();
?
Which doesn't use the information we have on top AVIC, making our logic
a bit simpler.
