Re: [PATCH 1/3] xen/pci: do not register devices outside of PCI segment scope

[Roger Pau Monné <roger.pau@xxxxxxxxxx>]

On Fri, Jan 10, 2025 at 04:21:29PM -0600, Bjorn Helgaas wrote:
> On Fri, Jan 10, 2025 at 03:01:48PM +0100, Roger Pau Monne wrote:
> > The PCI segment value is limited to 16 bits, however there are buses like 
> > VMD
> > that fake being part of the PCI topology by adding segment with a number
> > outside the scope of the PCI firmware specification range (>= 0x10000). The
> > MCFG ACPI Table "PCI Segment Group Number" field is defined as having a 16 
> > bit
> > width.
> >
> > Attempting to register or manage those devices with Xen would result in 
> > errors
> > at best, or overlaps with existing devices living on the truncated 
> > equivalent
> > segment values.
> 
> The ACPI _SEG method (ACPI r6.5, sec 6.5.6) and the corresponding
> value in the MCFG table (PCI Firmware r3.3, sec 4.1.2) are clearly
> 16-bit values.
> 
> But otherwise, the segment value is pretty much an arbitrary software
> value, and the kernel works fine with the larger domain values from
> vmd_find_free_domain(), so this isn't quite enough to explain what the
> issue with Xen is.
> 
> Does Xen truncate the domain to 16 bits or use it to look up something
> in ACPI?

In the interface between Xen and Linux the segment field is 16 bit
width, so with the current interface is not possible to reference
devices that are past the 0xffff segment.

I also wonder whether Xen and Linux (or guest OSes in general) would
agree on how to reference such devices.  AFAICT VMD segment numbers
are purely a software construct, but not something enforced by the
specification.  Could for example FreeBSD assign a different segment
to VMD devices?

If so we would need some kind of specification about how VMD segment
values are assigned so that OSes have a coherent way of referencing
VMD devices without ambiguity.

Thanks, Roger.