>>> However, adding new modes can silently make this check wrong (e.g.
>>> grepping for KVM_IRQCHIP_KERNEL will no longer identify all users). Do
>>> you think the optimization is worth it?
>>
>> I don't think we want to add new modes.
>
> I would prefer to write it like:
>
> kvm->arch.irqchip_mode > KVM_IRQCHIP_KERNEL_INIT;
Agreed.
>
> Same assembly with simpler code. Setting KVM_IRQCHIP_KERNEL_INIT before
> KVM_IRQCHIP_SPLIT would make it a bit more descriprive (and would allow
> the check below as well).
Okay, I added that.
>
>>>>> /* Matches with wmb after initializing kvm->irq_routing. */
>>>>> smp_rmb();
>>>>> diff --git a/arch/x86/kvm/irq_comm.c b/arch/x86/kvm/irq_comm.c
>>>>> index b96d389..4e4a67a 100644
>>>>> --- a/arch/x86/kvm/irq_comm.c
>>>>> +++ b/arch/x86/kvm/irq_comm.c
>>>>> @@ -282,22 +282,18 @@ int kvm_set_routing_entry(struct kvm *kvm,
>>>>>
>>>>> switch (ue->type) {
>>>>> case KVM_IRQ_ROUTING_IRQCHIP:
>>>>> + if (!irqchip_kernel(kvm) && !irqchip_kernel_init(kvm))
>>>>> + goto out;
>>>>> delta = 0;
>>>>
>>>> This can be irqchip_in_kernel, after which irqchip_kernel_init can be
>>>> removed.
>>>
>>> irqchip_in_kernel in its current form would allow KVM_IRQCHIP_SPLIT,
>>> which is not what we want here, or am I missing something?
>>
>> Hmm, perhaps we can split the checks to rule out KVM_IRQCHIP_NONE
>> outside the switch, and KVM_IRQCHIP_SPLIT here?
>
> Right, we don't want MSI or HV without LAPIC in kernel either.
>
Ok, I will see how this turns out!
Thanks!
--
Thanks,
David
