Re: [Qemu-devel] [PATCH] MSI / MSIX injection for Xen HVM

[Wei Liu]

On Thu, 2012-03-01 at 15:56 +0000, Paolo Bonzini wrote:
> Il 01/03/2012 15:50, Stefano Stabellini ha scritto:
> >>> > > That is a good point actually: we already have lapic emulation in Xen,
> >>> > > it makes sense to have apic-msi in Xen too.
> >>> > > We would still need the changes to msi_notify and msix_notify though.
> >> > 
> >> > Why?  The stores would just go to the Xen interrupt controller MMIO area
> >> > which then does the xc_hvm_inject_msi.
> >  
> > Because msi(x)_notify is called by QEMU's emulated devices: it is not
> > possible from QEMU to cause an emulation trap in Xen on behalf of the
> > guest.
> 

I'm not a QEMU expert, so following question may be dumb. However I do
care about a cleaner implementation. So please be patient with me. :-)

> msi{x,}_notify doesn't have to go to Xen MMIO emulation, so in Wei's
> patch you don't need anymore the msi{,x}_notify parts, only apic_send_msi.
> 

I don't quite understand "you don't need anymore the msi{,x}_notify
parts". Virtio PCI invokes msi_notify directly. If I don't hook up
msi{,x}_notify, how can I deal with devices like Virtio PCI?


Wei.

> But you could take it further and create a separate subclass of
> apic-common that does absolutely nothing except register an MMIO memory
> region and use it to trap MSI writes.  Basically an even more stripped
> down version than hw/kvm/apic.c, but using memory_region_init_io rather
> than memory_region_init_reservation.  You would also get support for the
> monitor command inject-nmi (there is another hypercall for that in Xen
> IIRC).
> 
> Paolo