Re: [Xen-devel] [PATCH v3 13/24] xen/arm: Implement hypercall PHYSDEVOP_{, un}map_pirq

[Julien Grall <julien.grall@xxxxxxxxxx>]

Hi Ian,

On 23/02/2015 16:34, Ian Campbell wrote:
> On Mon, 2015-02-23 at 16:11 +0000, Julien Grall wrote:
> > On 23/02/15 16:04, Ian Campbell wrote:
> > > On Mon, 2015-02-23 at 15:53 +0000, Julien Grall wrote:
> > > > Hi Ian,
> > > >
> > > > On 23/02/15 15:28, Ian Campbell wrote:
> > > > > On Mon, 2015-02-23 at 09:33 +0000, Jan Beulich wrote:
> > > > > > > > > On 20.02.15 at 17:53, <ian.campbell@xxxxxxxxxx> wrote:
> > > > > > > Jan, do you have any feeling for how this is going to play out on x86
> > > > > > > with the vapic stuff?
> > > > > > The vapic logic shouldn't require any physdevop involvement, so if
> > > > > > I read right what you propose (not having such a requirement /
> > > > > > connection on ARM) either, I agree that a new domctl should be all
> > > > > > that's needed (if XEN_DOMCTL_{,un}bind_pt_irq can't be re-used).
> > > > > Actually, I think bind_pt_irq with a new PT_IRQ_TYPE_* would be a good
> > > > > option.
> > > > >
> > > > > An ARM SPI is a bit like an ISA IRQ, but not close enough to reuse IMHO
> > > > > (and the datatype would need widening).
> > > > We have to think about MSI and other type too...
> > > >
> > > > In any case a DOMCTL is not suitable here because a guest kernel may
> > > > need to map/unmap IRQ too (think about ACPI or device passthrough).
> > > I don't follow, setting up device passthrough is very much a toolstack
> > > operation, isn't it? Where does the guest kernel get involved?
> > Sorry I meant the DOM0 kernel.
> >
> > Not really. On platform device pass-through there is no way to know the
> > IRQ, so for now the routing is done by the toolstack.
> >
> > But we could decide to implement a driver in DOM0 which will
> > unbind/bind/reset device.
> Sure, but...
>
> >   In this case it will require to
> > assign/deassign the IRQ from DOM0.
> ...why does that follow?
Because we may decide to re-use the device to DOM0.

In the case of the DOMCTL, it's not possible to do it.

> > There is also the case of MSI.
> Handled via XEN_DOMCTL_bind_pt_irq for the toolstack configuration
> angle, the actual guest usage of them is a separate interface which
> doesn't yet concern us, at least not in this series.
That would require some rework in the toolstack to make the IRQ routing 
(xc_physdev...) per architecture.
Also what about IRQ permission? Should we just ignore it and not give 
permission to the guest domain?
> > > As for ACPI, I think dom0 propagating ACPI derived platform info to Xen
> > > should be handled differently (at the hypercall interface at least)
> > > separate from passthrough.
> > There is no difference between routing because of ACPI and/or because
> > pass-through. So this should be done the same way.
> I'm not convinced. Routing all the IRQs is only one aspect of dom0
> propagating ACPI derived platform info to Xen.
>
> I suppose we will see once I look at the ACPI series. In the meantime I
> think XEN_DOMCTL_bind_pt_irq matches your requirements in for this
> series (and is a domctl so we aren't tied to it once we have a better
> understanding of the other stuff).
ACPI is one part of the problem, MSI with PCI is another problem.

Anyway, I suspect we will have the same talk before 4.6 release so we 
just defer the problem.
Regards,

--
Julien Grall

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel