On Thu, Feb 9, 2012 at 8:50 PM, Yinghai Lu <yinghai@xxxxxxxxxx> wrote: > On Thu, Feb 9, 2012 at 8:34 PM, Bjorn Helgaas <bhelgaas@xxxxxxxxxx> wrote: >> 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. > > I mean, we can unify sysdata usuage, make them all most the same. > > and for some sysdata, already acting like host_bridge struct with > domain, numa_node, and resources. > >> If you can show an example of what you mean, maybe it will help me understand. > > just looked the sys_data of different arch, > > looks like pci_sys_data for arm already include the io offset > > arch/arm/include/asm/mach/pci.h:: > > struct pci_sys_data { > #ifdef CONFIG_PCI_DOMAINS > int domain; > #endif > struct list_head node; > int busnr; /* primary bus number > */ > u64 mem_offset; /* bus->cpu memory mapping > offset */ > unsigned long io_offset; /* bus->cpu IO mapping offset > */ > struct pci_bus *bus; /* PCI bus > */ > struct list_head resources; /* root bus resources (apertures) Let me try again. Maybe your idea is that we should create something like "struct pcibios_sysdata" and require every arch to define a struct with that name. Every arch's struct would contain some members such as "domain," "mem_offset," "io_offset," and "resources." These would have standard names & types across all arches and would be referenced by the PCI core. In addition, the struct could include arch-specific data such as "acpi_handle," "iommu," etc. That way, the core could implement things like pci_domain_nr() and pci_bus_to_resource() by using the data in struct pcibios_sysdata. The current "void *sysdata" parameters to pci_scan_bus() and friends would become "struct pcibios_sysdata *sysdata". I like the idea of unifying the sysdata struct name and changing from a "void *" to something like "struct pcibios_sysdata *". That might be worth doing. But the whole point of this patch series is to move things that are not architecture-specific out of the arch code. I don't want to end up with a dozen pcibios_sysdata structs that all contain exactly the same generic members plus a few arch-specific ones. That would mean trying to keep those dozen places consistent all the time, which is a maintenance hassle. This series does cause some duplication, e.g., struct pci_host_bridge will contain a "struct pci_bus *" and some resource information that's also in the arch-specific struct pci_controller. That information is not actually arch-specific at all, so I think it's a good thing if we gradually remove it from the arch-specific sysdata structs. In many cases, that will also remove current arch restrictions such as "all memory apertures must have the same CPU/bus address offset," and "at most one I/O aperture and three memory apertures are supported." 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
