On 26 July 2015 at 14:46, Pavel Fedin <p.fe...@samsung.com> wrote: > Unfortunately, this is how qemu is designed. kvm_irqchip_create() is > called long before machine model's init code (machvirt_init() in our > case) is called. So, if we want to check for the right thing, we have > to know it before machine model actually starts to instantiate its > components. > > I don't see how to change this in a more or less simple way. This > routine, except capability testing, does one more important thing > on old kernels. There it actually creates vGICv2 using > KVM_CREATE_IRQCHIP ioctl.
Right, but all we really need to do in kvm_irqchip_create() is confirm whether we can use the new API later; we don't actually need to care whether we will later be creating a GICv2 or a GICv3 (or whether the kernel supports GICv3). Rather than having kvm_irqchip_create() pile up a list of every irqchip type we might want later, I think it's probably better to just have it do the check for KVM_CAP_DEVICE_CTRL. If that check fails, we are on an old kernel, and should return 0 so we do the old-style KVM_CREATE_IRQCHIP. If it succeeds, then we're on a new kernel, and can postpone doing any GIC init until the GIC device object is created. Then the GIC device object should just try to do the kvm_create_device() for the relevant kind of GIC. If we're trying to create a GICv3 on a kernel that only knows about a GICv2 then this will fail at this point and we can display a suitable error message. Christoffer: was there a particular reason you wrote kvm_irqchip_create() to do a test device-create as well as check the capability bit? thanks -- PMM
