Re: [PATCH] xen/mcfg: Call PHYSDEVOP_pci_mmcfg_reserved before PCI enumeration

[Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>]

On Tue, Sep 22, 2015 at 08:11:41AM -0600, Jan Beulich wrote:
> >>> On 22.09.15 at 16:03, <konrad.wilk@oracle.com> wrote:
> > On Tue, Sep 22, 2015 at 07:52:19AM -0600, Jan Beulich wrote:
> >> >>> On 22.09.15 at 15:39, <konrad.wilk@oracle.com> wrote:
> >> > On Tue, Sep 22, 2015 at 06:26:11AM -0700, Ed Swierk wrote:
> >> >> Any other ideas?
> >> > 
> >> > I like it - as it will update it right away. However we would need some
> >> > extra smarts in Xen to reconfigure its view of the PCI device now that the
> >> > extended configuration space has become available.
> >> 
> >> What parts are you thinking of that would need updating (and
> >> aren't getting updated already)?
> > 
> > The VF data. As before this call Xen might have not been able to
> > get to the extended configuration.
> 
> I still don't understand: Afaics pci_add_device() updates everything
> that may need updating already. And things appear to be working
> fine with our kernel (where the MMCFG notification comes right out
> of the x86 code earlier referred to), despite this meaning updates
> to the data collected during early boot. I guess you'll need to be
> more specific on what you see missing...

This is all ancient recollection of what I had seen a year or so ago
so take it a with a cup of salt.

I have one of those Intel machines where the MMCFG is not marked
in the E820 but is in the ACPI and I remember getting frequent
WARN_ON from Xen in the msi.c code when doing passthrough on the VF.
I don't have the logs but my vague recollection was that Xen could
not validate the VF's MSI-X because at the time it gathered the
VF PCI device information the extended configurations were not
available. This was prior to the XSA120 discovery so ancient
code.

With Ed's solution (along with your optimizations you suggested)
this should work. I can certainly try it out on my rig.


> 
> Jan
>