Re: [RFC PATCH 5/6] xen/arm: Intercept vPCI config accesses and forward them to emulator

[Stefano Stabellini <sstabellini@xxxxxxxxxx>]

On Wed, 15 Nov 2023, Julien Grall wrote:
> Hi,
> 
> On 15/11/2023 11:26, Sergiy Kibrik wrote:
> > From: Oleksandr Tyshchenko <oleksandr_tyshchenko@xxxxxxxx>
> > 
> > This is needed for supporting virtio-pci.
> > 
> > In case when the PCI Host bridge is emulated outside of Xen
> > (IOREQ server), we need some mechanism to intercept config space
> > accesses on Xen on Arm, and forward them to the emulator
> > (for example, virtio backend) via IOREQ request.
> > 
> > Unlike x86, on Arm these accesses are MMIO, there is no CFC/CF8 method
> > to detect which PCI device is targeted.
> > 
> > In order to not mix PCI passthrough with virtio-pci features we add
> > one more region to cover the total configuration space for all possible
> > host bridges which can serve virtio-pci devices for that guest.
> > We expose one PCI host bridge per virtio backend domain.
> I am a little confused. If you expose one PCI host bridge per virtio backend,
> then why can't the backend simply register the MMIO region and do the
> translation itself when it receives the read/write?
> 
> To me, it only makes sense for Xen to emulate the host bridge access if you
> plan to have one host bridge shared between multiple IOREQ domains or mix with
> PCI pasthrough.

+1

This is exactly what Stewart, Vikram and I were thinking


> From my perspective, I don't expect we would have that many virtio PCI
> devices. So imposing a host bridge per device emulator will mean extra
> resource in the guest as well (they need to keep track of all the hostbridge).

+1


> So in the longer run, I think we want to allow mixing PCI passthrough and
> virtio-PCI (or really any emulated PCI because nothing here is virtio
> specific).

+1


> For now, your approach would be OK to enable virtio PCI on Xen. But I don't
> think there are any changes necessary in Xen other than reserving some MMIO
> regions/IRQ.

I don't mean to slow down EPAM, but I think we can jump in and do the
necessary changes to vPCI for everyone's benefits and with a timeframe
that works for AMD, EPAM and the Xen community.