On Thu, Feb 9, 2012 at 8:03 PM, Yinghai Lu <yinghai@xxxxxxxxxx> wrote: > On Thu, Feb 9, 2012 at 6:36 PM, Bjorn Helgaas <bhelgaas@xxxxxxxxxx> wrote: >> There's a lot of PCI-related code under arch/, but much of it is not actually >> architecture-specific. This series removes some of that code by moving most >> of the bus-to-resource conversions into the core. >> >> We currently read PCI bus addresses from BARs in the core (pci_setup_device()). >> Then every arch is responsible for converting those bus addresses to CPU >> resources, usually in pcibios_fixup_bus(). >> >> We already have a way for architectures to tell the core what the windows >> through a host bridge are: >> >> LIST_HEAD(resources); >> pci_add_resource(&resources, io_space); >> pci_add_resource(&resources, mem_space); >> pci_scan_root_bus(parent, bus, ops, sysdata, &resources); >> >> This series extends that so the arch can also tell the core about address >> translation performed by the host bridge: >> >> LIST_HEAD(resources); >> pci_add_resource_offset(&resources, io_space, io_offset); >> pci_add_resource_offset(&resources, mem_space, mem_offset); >> pci_scan_root_bus(parent, bus, ops, sysdata, &resources); >> >> Given that offset (the difference between bus address and CPU address for >> each aperture), the core can do the bus-to-resource conversion immediately >> when it reads the BARs. >> >> This removes an opportunity for bugs (some PCI fixups currently see bus >> addresses in struct pci_dev resources when they're expecting CPU addresses), >> but the main reason to do this is to make our PCI resource handling simpler >> and more uniform. >> >> These patches are also available in this git repo: >> git://github.com/bjorn-helgaas/linux.git pci-offset-v2-d15af52258dd >> >> Or you can browse them here: >> https://github.com/bjorn-helgaas/linux/compare/master...pci-offset-v2-d15af52258dd >> >> I'd like to get these into linux-next soon to be ready for the 3.4 merge >> window, so please let me know if you see any issues. >> >> Changes since v1: >> - mips: remove Cobalt legacy IDE fixup >> - show bus address range, not offset, e.g., this: >> pci_bus 0000:00: root bus resource [mem 0xf0000000000-0xf007edfffff] (bus address [0x80000000-0xfedfffff]) >> instead of this: >> pci_bus 0000:00: root bus resource [mem 0xf0000000000-0xf007edfffff] (bus offset 0xeff80000000) >> > > still think make pci_sysdata to generic and add one offset field to > that would be more simple. Maybe I'm misunderstanding you, because this doesn't make sense to me at all. *One* offset field is insufficient. A host bridge can have an arbitrary number of apertures, and each aperture can have a different offset. So we must have a list or other variable-size structure. pci_scan_bus() takes a "void *sysdata". Every arch defines its own structure (struct pci_sysdata for x86, struct pci_controller for most others) to supply here. Are you suggesting that we make a common structure like this: struct pci_sysdata { struct list_head windows; void *sysdata; /* arch-specific stuff */ }; That's possible, but I don't see the advantage over what I've done. It doesn't seem simpler, because all the arch code that looks at bus->sysdata would have to change to use bus->generic_sysdata->sysdata. If you can show an example of what you mean, maybe it will help me understand. 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
