[Ksummit-2012-discuss] PCI breakout session
benh at kernel.crashing.org
Tue Jun 19 20:17:54 UTC 2012
On Tue, 2012-06-19 at 11:22 +0000, Arnd Bergmann wrote:
> Speaking of which, I would like to discuss ways to make that less
> architecture specific. We have a bunch of architectures that provide
> more than one way to represent a PCI domain or root port, typically
> one for each SoC family for the smaller ones, or different hypervisors
> on higher-end architectures.
> Right now, depending on the architecture, we use struct pci_sys_data,
> struct pci_controller, pci_channel or pci_hba_data, which are all
> similar but not identical. I would hope that we can move the common
> parts into a framework that lets you register a pci domain
> dynamically, and lets us move a lot of the simpler implementations to
> drivers/pci/host and out of the architecture code.
Most definitely. There has already been some work in that direction,
for example the ability to handle remapped windows is now in the core,
and the device-node handling is generic as well.
We will probably always need some kind of architecture specific "add-on"
though for crazy things like EEH :-)
More information about the Ksummit-2012-discuss