Re: [PATCH] x86/PVH: PHYSDEVOP_pci_mmcfg_reserved should not blindly register a region

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

On Fri, May 08, 2020 at 03:49:35PM +0200, Jan Beulich wrote:
> [CAUTION - EXTERNAL EMAIL] DO NOT reply, click links, or open attachments 
> unless you have verified the sender and know the content is safe.
> 
> On 08.05.2020 14:54, Andrew Cooper wrote:
> > On 08/05/2020 13:43, Jan Beulich wrote:
> >> The op has a register/unregister flag, and hence registration shouldn't
> >> happen unilaterally. Introduce unregister_vpci_mmcfg_handler() to handle
> >> this case.
> >>
> >> Fixes: eb3dd90e4089 ("x86/physdev: enable PHYSDEVOP_pci_mmcfg_reserved for 
> >> PVH Dom0")
> >> Signed-off-by: Jan Beulich <jbeulich@xxxxxxxx>
> > 
> > I agree in principle that registration shouldn't be unilateral, but why
> > on earth does the API behave like that to begin with?
> > 
> > There is no provision to move or update MMCFG regions in any spec I'm
> > aware of, and hardware cannot in practice update memory routing like this.
> > 
> > Under what circumstances should we tolerate an unregister in the first
> > place?
> 
> Hot unplug of an entire segment, for example.

An OS could also rebalance the resources of a host bridge, according
to the PCI firmware spec, in which case _CBA should be re-evaluated.

I'm not sure whether rebalancing would work anyway, or if there even
are systems that support this and OSes that would attempt to do it,
but since we have the interface for this let's try to do something
sensible.

The other options is simply returning -EOPNOTSUPP. Iff the domain
doesn't try to access devices that would reside on the segment
hot-unplugged it shouldn't make much of a difference, rebalancing is
the case were Xen must support add/remove in order to re-place the
position of the ECAM areas.

Thanks, Roger.