On Mon, Feb 27, 2012 at 8:52 PM, Benjamin Herrenschmidt
<benh@xxxxxxxxxxxxxxxxxxx> wrote:
> On Mon, 2012-02-27 at 19:50 -0700, Bjorn Helgaas wrote:
>> --- a/arch/powerpc/kernel/pci_32.c
>> +++ b/arch/powerpc/kernel/pci_32.c
>> @@ -219,9 +219,9 @@ void __devinit pcibios_setup_phb_io_space(struct
>> pci_controller *hose)
>> struct resource *res = &hose->io_resource;
>>
>> /* Fixup IO space offset */
>> - io_offset = (unsigned long)hose->io_base_virt - isa_io_base;
>> - res->start = (res->start + io_offset) & 0xffffffffu;
>> - res->end = (res->end + io_offset) & 0xffffffffu;
>> + io_offset = pcibios_io_space_offset(hose);
>> + res->start += io_offset;
>> + res->end += io_offset;
>> }
>
> Well, you are losing the 0xffffffff mask... so basically, what happens
> is that your offset, if negative, will be 0 extended by
> pcibios_io_space_offset() instead of sign-extended, and thus the
> resource will have crap stuff in the top 32-bit.
I don't think it will have random crap. Here's what I think will
happen. Assume typical 64KB I/O port spaces, PCI bus addresses in the
range 0x0-0xffff, with isa_io_base = 0x10000000. If we compute the
offset as I did in the patch:
(unsigned long) hose->io_base_virt - _IO_BASE
the subtraction result is 32-bits and is zero-extended to 64 bits.
Alternatively, we could do this:
(unsigned long long) (unsigned long) hose->io_base_virt - _IO_BASE
Here the conversion to 64 bits happens before the subtraction, so the
upper 32 bits of the result are likely all ones. Examples:
hose->io_base_virt offset resource
0x10000000 0x00000000 [io 0x0000-0xffff]
0x11000000 0x01000000 [io 0x01000000-0x0100ffff]
0x0f000000 0xff000000 [io 0xff000000-0xff00ffff]
0x10000000 0x00000000 [io 0x0000-0xffff]
0x11000000 0x01000000 [io 0x01000000-0x0100ffff]
0x0f000000 0xffffffffff000000 [io
0xffffffffff000000-0xffffffffff00ffff]
Either way will probably work since you only care about the low 32
bits, which are the same either way. I chose the first because
casting the 32-bit pointer to a 64-bit value seemed to require two
ugly casts and the resulting resource values look unwieldy when
printed.
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
