On 23/09/2016 12:12, Peter Xu wrote:
> On Fri, Sep 23, 2016 at 12:03:26PM +0200, Paolo Bonzini wrote:
>>
>>
>> On 23/09/2016 12:02, Peter Xu wrote:
>>> On Fri, Sep 23, 2016 at 11:27:09AM +0200, Paolo Bonzini wrote:
>>>>> Since the whole IOMMU feature is new and somewhat experimental, I think
>>>>> it's okay to just make EIM the default for >=2.8 machine types if KVM is
>>>>> on (using DEFINE_PROP_ON_OFF_AUTO; auto means true if KVM is on and
>>>>> false otherwise, and pc-2.7 would set eim=off). It means requiring
>>>>> kernel 4.8 by default, but I don't think it's a big deal.
>>> I think the problem is, even we have KVM support for x2apic, we are
>>> still losing QEMU part. And guests with cluster x2apic and >8 vcpus
>>> will not working properly on device interrupts, which can be very
>>> confusing to people (it can boot, but some devices just don't work
>>> properly, and they won't see useful information in guest dmesg).
>>
>> Yes, that's why I suggested EIM=on by default.
> 
> I am confused. :(
> 
> Why not we just keep people from that wrong configuration by default,
> until we have x2apic in QEMU?

Do you mean Igor's patches?  I expect that they will go in pretty much
at the same time as Radim's.

Paolo

Reply via email to