On Monday 19 January 2015 11:12:50 Yijing Wang wrote:

On 2015/1/16 17:18, Arnd Bergmann wrote:

quoted

On Friday 16 January 2015 09:44:08 Yijing Wang wrote:

quoted

@@ -2066,11 +2064,11 @@ struct pci_bus *pci_scan_root_bus(struct device *parent, u32 db,
 {
        struct pci_host_bridge *host;
 
-       host = pci_create_host_bridge(parent, db, resources);
+       host = pci_create_host_bridge(parent, db, resources, sysdata);
        if (!host)
                return NULL;
 
-       return __pci_scan_root_bus(host, ops, sysdata);
+       return __pci_scan_root_bus(host, ops);
 }
 EXPORT_SYMBOL(pci_scan_root_bus);

How about keeping the sysdata out of the pci_create_host_bridge interface, and
refactoring it so that the call sequence becomes

      host = pci_create_host_bridge(parent, db, resources);
      host->sysdata = sysdata;
      __pci_scan_root_bus(host, ops);

This way, we can make sysdata completely option. I assume that more of the
fields we have in sysdata today can get moved into pci_host_bridge
over time, so a host bridge driver can just assign those members individually
between pci_create_host_bridge and __pci_scan_root_bus.

I put the sysdata in pci_host_bridge, because some platforms x86/ia64
need to call pcibios_root_bridge_prepare() before pci_host_bridge registration
in pci_create_host_bridge(), and pcibios_root_bridge_prepare() need the companion
acpi device pointer which is stored in sysdata.

I like you idea, but I haven't find a better way yet.

Ok, fair enough. We can always try to find ways for getting rid of
pcibios_root_bridge_prepare() later, though I suspect as soon as we
have the callback, it will get used for other things as well ;-)

	Arnd