Like I said, the device-tree PCI nodes know what they are doing; they
are implicitly bound to the host controller.

RTAS just accepts bus, device, function as a configuration address. It
has no idea which one to look at without some internal hackery. It has
no clue about seperate host controllers.

Our firmware knows the difference, I have no clue about anyone else, but
in the end - Apple are the biggest user of OF besides Sun, and neither
of them implement RTAS *growl*. Use RTAS, fall back to config-[lwb][@!]
if you don't have RTAS, and it will work. Expect funny behaviour if
RTAS is broken somehow. But Linux uses RTAS on the Efika - ostensibly
to work around the AGP/PCI distinction in firmware at some point.

On the Discovery it just uses native PCI bus access.

-- 
Matt Sealey <matt@genesi-usa.com>
Genesi, Manager, Developer Relations

Frank Wille wrote:
> Matt Sealey wrote:
> 
>>> Currently I'm not smart enough to see a reason why the "PCI domain" 
>>> can be a problem, but the rest sounds conclusive.
>>
>> For instance if you send a configuration space request for bus 0, 
>> device 1,
>> function 8.. how does it know which host controller? :)
> 
> When using config-l@ with the OFW-handle of the first PCI-node,
> I expect it to read from the config registers of the first
> host controller.
> 
> There are separate registers for the two PCI host controllers,
> or not? At least it seems so from the OpenBSD source.
>