KVM: Reduce atomic operations on vcpu->requests