Re: 48GB NUMA-Q boots, with major IO-APIC hassles

[Anton Blanchard]

> 	-- I've already tried and failed on PCI segments.
> 	-- I would be much obliged for advice on this one, especially
> 	-- since the workaround cripples the machine.

I was wondering if we could use pci_scan_bus_parented as a start.
Assuming we fixed pci_bus_exists, this should allow us to create
overlapping busses.

What else needs fixing? To start with:

* pci config read/writes
	Should be arch specific, can hide things in struct pci_bus
	which is passed in. (eg on ppc64 we have to pass an identifier
	into the low level config read or write)

* pci IO/memory reads/writes
	No problems on ppc64, it should just work. Are there problems
	on x86? 

* /proc/bus/pci/
	As davem pointed out we have a backwards compatibility bit
	where all devices on domain 0 appear as they always have.
	Additionally we create another level of directories to represent
	the domain.

* /proc/pci
	Need to print the domain

* device printing
	We need a macro that drivers can use to print their PCI location
	Then we make that print the domain.

* sysfs
	I havent looked into this yet

Does anything else spring to mind?

Anton