Hi,

Here I have a 8541 dev board, with 2 e1000 attached to pci0 and a
homebrewed addon board inserted into pci1. I am trying to make it work
under 2.6.26-rc8 (2.6.26 broken so I am working at rc8).

The ioports allocation reads

$ cat /proc/ioports
00000000-000fffff : /pci@e0008000
  00001000-0000103f : 0000:00:0a.0
  00001040-0000107f : 0000:00:0b.0
ffefe000-ffffdfff : /pci@e0009000
  ffeff000-ffeff00f : 0001:01:0a.0
  ffeff010-ffeff01f : 0001:01:0b.0
  ffeff020-ffeff02f : 0001:01:0c.0
  ffeff030-ffeff03f : 0001:01:0d.0

The port allocation for pci1 looks ridiculous. The addon board doesn't work.

After poking around, I find pci_process_bridge_OF_ranges() in
arch/powerpc/kernel/pci-common.c

--snip--
if (primary)
        isa_io_base = (unsigned long)hose->io_base_virt;
--snip--

then fixup_resource(), _IO_BASE = isa_io_base. fix_resource will use
isa_io_base as base address to assign io port address space.

This is reasonable, but on my board, pci1's hose->io_base_virt is
smaller than pci0's. This lead to

[    0.064214] PCI:0000:00:0a.0 Resource 4
0000000000001000-000000000000103f [20101] fixup...
[    0.064224] fixup_resource() offset=00000000, io_base_virt=fdeb4000,
_IO_BASE=fdeb4000, io_base_phys=e2000000
[    0.064232] PCI:0000:00:0a.0            0000000000001000-000000000000103f

[    0.065129] PCI:0001:01:0a.0 Resource 2
0000000000001000-000000000000100f [20101] fixup...
[    0.065139] fixup_resource() offset=ffefe000, io_base_virt=fddb2000,
_IO_BASE=fdeb4000, io_base_phys=e3000000
[    0.065147] PCI:0001:01:0a.0            00000000ffeff000-00000000ffeff00f

offset = fddb2000 - fdeb4000 = ffefe000

So far, I think the workaround is to replace the code above with

--snip--
if (!isa_io_base)
        isa_io_base = (unsigned long)hose->io_base_virt;
else if ((unsigned long)hose->io_base_virt < isa_io_base)
        isa_io_base = (unsigned long)hose->io_base_virt;
--snip--

Then ioport allocation reads (But I haven't test if it works)

$ cat /proc/ioports
00000000-000fffff : /pci@e0009000
  00001000-0000100f : 0001:01:0a.0
  00001010-0000101f : 0001:01:0b.0
  00001020-0000102f : 0001:01:0c.0
  00001030-0000103f : 0001:01:0d.0
00102000-00201fff : /pci@e0008000
  00103000-0010303f : 0000:00:0a.0
  00103040-0010307f : 0000:00:0b.0

Can someone gives a better generic fix for this?

PS: I attach the kernel log (without workaround) for reference. I added
some printk code.