On Mon, 10 Dec 2012 09:21:51 -0600, Rob Herring [off-list ref] wrote:

On 12/10/2012 09:05 AM, Michal Simek wrote:

quoted

On 12/10/2012 03:26 PM, Rob Herring wrote:

quoted

On 12/10/2012 06:20 AM, Michal Simek wrote:

quoted

Hi Grant and others,

I have a question regarding number of cells in ranges property
for pci and pcie nodes.

Linux pci/pcie powerpc DTSes contain 7 cells (xpedite5370.dts,
sequoia.dts, etc)
but also 6 cells format too (mpc832x_mds.dts)

Here is shown 6 cells ranges format and describe
http://devicetree.org/Device_Tree_Usage#PCI_Host_Bridge

And also in documentation in the linux
Documentation/devicetree/bindings/pci/83xx-512x-pci.txt

Both format uses:
#size-cells = <2>;
#address-cells = <3>;

What is valid format?

Both. 7 cells are valid when the host (parent) bus is 64-bit and 6 cells
are valid when the host bus is 32-bit. The ranges property is <<child
address> <parent address> <size>>. The parent address #address-cells is
taken from the parent node.

Ok. Got it.

Here is what we use on zynq and microblaze - both 32bit which should be
fine.

    ps7_axi_interconnect_0: axi@0 {
        #address-cells = <1>;
        #size-cells = <1>;
        axi_pcie_0: axi-pcie@50000000 {
            #address-cells = <3>;
            #size-cells = <2>;
            compatible = "xlnx,axi-pcie-1.05.a";
            ranges = < 0x02000000 0 0x60000000 0x60000000 0 0x10000000 >;
            ...
        }
    }

What I am wondering is pci_process_bridge_OF_ranges() at
arch/powerpc/kernel/pci-common.c
where there are used some hardcoded values which should be probably
loaded from device-tree.

For example:
683         int np = pna + 5;
...
702                 pci_addr = of_read_number(ranges + 1, 2);
703                 cpu_addr = of_translate_address(dev, ranges + 3);
704                 size = of_read_number(ranges + pna + 3, 2);

These would always be correct whether you have 6 or 7 cells. pna is the
parent bus address cells size. The pci address is fixed at 3 cells.

quoted

Unfortunately we have copied it to microblaze.

I look at the PCI DT code in powerpc and see a whole bunch of code that
seems like it should be common. The different per arch pci structs
complicates that. No one has really gotten to looking at PCI DT on ARM
yet except you and Thierry for Tegra. We definitely don't want to create
a 3rd copy. Starting the process of moving it to something like
drivers/pci/pci-of.c would be great.

A lot of it should be common. The microblaze code is a copy of the
powerpc version. I'll strongly nack any attempt to add a third!  :-)

drivers/pci/pci-of.c would be good. I'd also accept drivers/of/pci.c
which might actually be a good idea in the short term so that it gets
appropriate supervision while being generalized before being moved into
the pci directory.

g.