Re: [PATCH v3 04/16] kvm: x86: Add support for activate/de-activate APICv at runtime

[Paolo Bonzini]

On 13/09/19 21:00, Suthikulpanit, Suravee wrote:
> +	kvm_for_each_vcpu(i, v, kvm)
> +		kvm_clear_request(KVM_REQ_APICV_DEACTIVATE, v);
> +
> +	if (kvm_x86_ops->pre_update_apicv_exec_ctrl)
> +		kvm_x86_ops->pre_update_apicv_exec_ctrl(vcpu, true);
> +
> +	kvm->arch.apicv_state = APICV_ACTIVATED;
> +
> +	kvm_make_all_cpus_request(kvm, KVM_REQ_APICV_ACTIVATE);
> +
> +	mutex_unlock(&kvm->arch.apicv_lock);

I think a lot of this logic can be simplified.  In particular, I would have:

- a single KVM_REQ_APICV_TOGGLE request that unifies
kvm_vcpu_activate_apicv and kvm_vcpu_deactivate_apicv.  The invariant is
that, at the end of the function, (vcpu->arch.apicv_active ==
(kvm->arch.apicv_state = APICV_ACTIVATED)).  The apicv_lock then becomes
an rwsem, so that kvm_activate_apic and kvm_deactivate_apic will
down_write it, while the processing of KVM_REQ_APICV_TOGGLE can
down_read it.

- the srcu_read_unlock/srcu_read_lock should be hidden in
svm_request_activate_avic/svm_request_deactivate_avic.  Everything else
should only take struct kvm*, following what you've started with patch
1.  In particular kvm_vcpu_deactivate_apicv should be changed to take a
struct kvm*, so that Hyper-V can do kvm_deactivate_apic(kvm,
APIC_DISABLED).  Hyper-V should not care about
srcu_read_lock/srcu_read_unlock.  avic_setup_access_page and
avic_destroy_access_page also can be changed to take struct kvm*.

Paolo