[PATCH 1/7] KVM: api: add kvm_irq_routing_extended_msi

[Andre Przywara]

Hi Paolo,

thanks for looking at this!

On 06/07/15 12:07, Paolo Bonzini wrote:
> 
> 
> On 06/07/2015 12:37, Christoffer Dall wrote:
>> I don't view it as 'the kernel requires this' but as 'the kernel will
>> not complain with arbitrary error code if you set the devid flag'
>> capability, and it's up to userspace (as usual) to provide the correct
>> arguments for things to work, and up to the kernel to ensure we don't
>> crash the system etc.
>>
>> Thus, if you want to advertise it as a capability, I would rather call
>> it KVM_CAP_MSI_DEVID.
> 
> I agree.  Does userspace know that ITS guests always require devid?

Well, as we are about to implement this: yes. But the issue is that MSI
injection and GSI routing code is generic PCI code in userland (at least
in kvmtool, guess in QEMU, too), so I don't want to pull in any kind of
ARM specific code in there. The idea is to always provide the device ID
from the PCI code (for PCI devices it's just the B/D/F triplet), but
only send it to the kernel if needed. Querying a KVM capability is
perfectly fine for this IMO.

> I
> guess it's okay to return -EINVAL if the userspace doesn't set the flag
> but the virtual hardware requires it.

Yes, that is what I do in the kernel implementation. And that is
perfectly fine: the ITS emulation does not work without a device ID, the
ITS driver in the guest assigns the very same payload (and address) to
different devices, so there is no way to tell the MSIs apart without a
unique device ID.

Thanks,
Andre.

> 
> Paolo
> 
>> The question is if userspace code that sets the devid flag will anyway
>> depend on some discovery mechanism of whether or not the kernel supports
>> arm64 irqfd etc. and if so, can we be sure to add the required support
>> at once in the kernel so that EINVAL never means 'you set the flags
>> field on the ioctl on an old kernel'?
>>
>> This smells an awful lot like a capability to me.
>