Re: pci-3.14 resource alloc

[Steven Newbury <steve@snewbury.org.uk>]

On Tue, 2014-02-18 at 07:37 -0700, Bjorn Helgaas wrote:
> On Tue, Feb 18, 2014 at 6:59 AM, Bjorn Helgaas <bhelgaas@google.com> wrote:
> > On Tue, Feb 18, 2014 at 3:08 AM, Steven Newbury <steve@snewbury.org.uk> wrote:
> 
> >> No, the only unusual option is I use pci=nocrs since the bios is
> >> completely ignorant of support for 64bit BARs.
> >
> > Hi Steven, this is a tangent, but can you collect the complete dmesg
> > log with "pci=nocrs" and let me know what happens when you *don't* use
> > "pci=nocrs" (a complete dmesg log there would be useful too, but I
> > don't know if you can boot without it)?  I think Linux *should* be
> > able to handle 64bit BARs even if the BIOS doesn't.
> >
> > Please post these as a separate thread so we don't muddle this
> > conversation with Yinghai.
> 
> Oh, never mind.  I think I remember now.  I think the problem is not
> so much that the BIOS doesn't handle 64bit *BARs*, but that your BIOS
> doesn't report 64bit host bridge *apertures*, and you want to use
> space above 4G for a graphics card or something.
> 
Yes, exactly. I should have stated it better.  I need to push the
on-board i965 up above 4G to make room for a radeon which is accessed
through a 32 bit only bridge.  The BIOS doesn't provide any
functionality like this.  The intel devs seemed a little surprised that
the i965 actually worked stably in this configuration given known
errata, but I've actually had no problems with this side of things. 

> That's not really anything we can fix in PCI because PCI doesn't
> discover those apertures.  Host bridge drivers like the generic ACPI
> pci_root.c or the native amd_bus.c discover them.  If those drivers
> don't discover them, PCI can't just make them up.

32 bit only bridges must be getting pretty rare by now, and I presume
any new hardware out there with a dockable PCI bus would provide 64 bit
resource allocation for host bridge apertures when booting to a 64 bit
OS?  Maybe not?
> 
> It might be nice if we had cleaner kernel options to say "assume a
> host bridge window here" and "put this device here," so you didn't
> have to use the big hammer of "pci=nocrs".  It's sort of a pain to
> figure out how to attach info like that to the correct device (there
> might be several host bridges, so how do you specify which one you
> care about?), but maybe it could be done.
> 
That would be much better, even if it's only some way to say "ignore crs
for *this* host bridge (and parents)", or "for host bridges to which
*this* device is attached".