Re: [Qemu-devel] [RFC][PATCH 14/16] kvm: x86: Add user space part for in-kernel i8259

[Avi Kivity]

On 12/04/2011 03:42 PM, Jan Kiszka wrote:
> On 2011-12-04 14:31, Avi Kivity wrote:
> > On 12/03/2011 01:17 PM, Jan Kiszka wrote:
> >> From: Jan Kiszka <address@hidden>
> >>
> >> Introduce the alternative 'kvm-i8259' device model that exploits KVM
> >> in-kernel acceleration.
> >>
> >> The PIIX3 initialization code is furthermore extended by KVM specific
> >> IRQ route setup. Moreover, GSI injection differs in KVM mode from the
> >> user space model. As we can dispatch ISA-range IRQs to both IOAPIC and
> >> PIC inside the kernel, we do not need to inject them separately. This is
> >> reflected by a KVM-specific GSI handler.
> >>
> >> +
> >> +qemu_irq *kvm_i8259_init(void)
> >> +{
> >> +    ISADevice *dev;
> >> +
> >> +    dev = isa_create("kvm-i8259");
> >>
> > 
> > Same issue.  Is this a different device, or an different implementation
> > of the same device?
>
> They are theoretically the same from guest perspective (therefore you
> can migrate between machines that differ in this).

But the name becomes part of the save/restore ABI, so you can't.

> > 
> > We're forcing migration from 1.0 to 1.1 to disable in-kernel irqchip on
> > the target.  For qemu itself, that's no issue.  But for qemu-kvm, it
> > will result in loss of performance, or hacks to alias the two back together.
>
> We should this happen with qemu-kvm? The vmstates are compatible, thus
> you can migration from old qemu-kvm in-kernel devices to the new kvm-*
> ones (once they are feature-equivalent). Not sure how much hacks this
> may require to qemu-kvm, but I don't think it should make the situation
> worse for that tree.

They aren't compatible due to the name clash.  The hack won't be large
(add an alias for the name), but just one hack is enough to keep the
tree alive for a long while.  Better not to add it in the first place.

-- 
error compiling committee.c: too many arguments to function