On Wed, Oct 10, 2012 at 4:25 AM, Felix Fietkau <n...@openwrt.org> wrote:
> On 2012-10-09 3:00 AM, Tim Harvey wrote: > > Moving the PCIe hw initialization to an early init call allows PCI based > > devices to be used when built as static. An example of where this is > needed > > is when using PCI based UARTs such as the Gateworks GW2201 Octal UART. > What do you mean by 'built as static'? Why does the initialization time > matter here? > > - Felix > > Felix, I've had some time to dig deeper on this issue and I've found the following: The kernel init levels are not documented well - the only documentation I've found was an uncommitted patch here: http://thread.gmane.org/gmane.linux.kernel/313880 . The ARM kernel's pci support provides 'pci_common_init' which is meant to be called from an init level prior to device init. All ARM platforms I looked at call their PCI/PCIe controller init either via arch_initcall (level 3) or subsys_initcall (level 4) - I don't think either one is necessary more correct, they both preceed device init (level 6). The ARM kernel does not support HOTPLUG_PCI - if it did then it would be fine to init PCI/PCIe controllers following device_init (level 6). In the current case of the cns3xxx kernel (or any ARM platform), calling pci_common_init from late_init (level 7) which follows device_init causes any drivers that have already been probed by nature of being static in the kernel to fail with resource conflicts. I'm going to re-submit this patch using subsys_initcall instead of device_initcall which I feel is more correct than what I've done here. Tim
_______________________________________________ openwrt-devel mailing list openwrt-devel@lists.openwrt.org https://lists.openwrt.org/mailman/listinfo/openwrt-devel
