Crap, here's the cover letter I meant to include. [RFC] PCI, PNP, resources: effectively reserve ACPI device resources 2.6.37-rc includes some patches of mine to change the way we allocate address space for devices, and these caused some regressions. Most device resource allocation is done by the BIOS, but Linux does allocate space if the BIOS doesn't do it (e.g., for CardBus or other hotplug devices) or if the BIOS did it wrong (e.g., it put a PCI device outside the host bridge windows). My original 2.6.37-rc patches make Linux allocate space from the top down instead of from the bottom up. We did this because that's what Windows does, and doing it the same way helps us avoid some BIOS defects, such as https://bugzilla.kernel.org/show_bug.cgi?id=16228 . The regressions caused by these changes include: https://bugzilla.kernel.org/show_bug.cgi?id=23332 Here, we allocated a 64MB bridge window for a subtractive-decode bridge leading to a CardBus controller, and the window conflicts with devices not reported by the BIOS. Windows doesn't see this problem because it relies on the fact that the bridge is in subtractive-decode mode and doesn't program the window at all. https://bugzilla.kernel.org/show_bug.cgi?id=23542 https://bugzilla.kernel.org/show_bug.cgi?id=23802 In both cases, Linux allocated space that conflicts with an ACPI device because it ignores resource usage reported by ACPI. Windows avoids this problem because it pays attention to the ACPI _CRS methods that report device resource usage. I think the best way to address these problems would be to put ACPI device resources in the resource maps, just like we do for PCI devices. We tried this a couple years ago (http://lkml.org/lkml/2007/11/1/214) and tripped over some subtle issues. We should try again, but not at this stage of the -rc. Instead, these patches take the approach of adding a filter in the allocate_resource() path so each arch can prevent allocation of any address ranges it cares about. For x86, that means looking through all the PNP devices and protecting the address space they use. In addition, for 23332, I added a quirk that builds a fake PNP device to protect the unreported devices. I'm a little nervous about this because I've been told that the WHQL tests look for unreported devices, and the HP nx6325 should have passwd. But it's the best I can think of for now. Bjorn -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
