Re: [PATCH v7 03/60] sparc/PCI: Unify pci_register_region()

[Bjorn Helgaas <helgaas@kernel.org>]

On Thu, Oct 22, 2015 at 02:47:42PM +1100, Benjamin Herrenschmidt wrote:
> On Wed, 2015-10-21 at 18:27 -0700, David Miller wrote:
> > From: Yinghai Lu <yinghai@kernel.org>
> > Date: Wed, 21 Oct 2015 11:16:53 -0700
> > 
> > > otherwise we need to compare res with pbm->mem_space or pbm
> > ->mem64_space
> > > to get direct parent for request_resource_conflict() calling in
> > > pci_register_legacy_regions().
> > 
> > Right, this is the issue.
> > 
> > On sparc64, the cpu physical base address used for accessing 64-bit
> > and non-64-bit memory spaces is different.
> > 
> > And that's why the resource values will be different.
> 
> This is the same on power btw, we have separate PowerBus windows to the
> PHB that get mapped respectively to 32-bit PCI MMIO and 64-bit PCI
> MMIO. The former gets "remapped" to generates 0-based PCI cycles, while
> the latter is 1:1. So the offset between CPU and PCI changes depending
> on which window you hit.

Lots of architectures do this, and the PCI core supports it by keeping
a list of all the host bridge windows.  Each entry has a struct
resource that contains the CPU address, as well as the offset from the
CPU address to the PCI address.  The arch or host bridge driver uses
pci_add_resource_offset() to tell the core about the window and the
offset.

Bjorn