On Tue, May 2, 2017 at 7:15 PM, Lorenzo Pieralisi
[off-list ref] wrote:

On Fri, Apr 28, 2017 at 02:28:38PM +0200, Arnd Bergmann wrote:

quoted

On Wed, Apr 26, 2017 at 1:17 PM, Lorenzo Pieralisi
[off-list ref] wrote:

Good idea, yes. To avoid growing the number of interfaces too
much, should we change the existing users of pci_register_host_bridge
in host drivers over to this entry point, and make the other one
local to probe.c then?

Yes, the problem is that there are drivers (ie pcie-iproc.c) that
require the struct pci_bus (created by pci_register_host_bridge())
to fiddle with it to check link status and THEN scan the bus (so
the pci_register_host_bridge() call can't be embedded in the scan
interface - the driver requires struct pci_bus for pci_ops to work
before scanning the bus itself).

I will see how I can accommodate this change because you definitely
have a point.

The obvious answer for that particular problem would be a link_check()
callback in the bridge operations. I think that would also fit in well with
the dw_pcie driver that has some private infrastructure for it. I don't
know if that callback is sufficient to solve all related problems though.

       Arnd