Re: [PATCH 3/4] PowerPC: Add PCI entry to 440EPx Sequoia DTS.

[Sergei Shtylyov <sshtylyov@ru.mvista.com>]

Hello.

Valentine Barshak wrote:

>>> --- linux-2.6.orig/arch/powerpc/boot/dts/sequoia.dts    2007-12-21 
>>> 17:14:17.000000000 +0300
>>> +++ linux-2.6/arch/powerpc/boot/dts/sequoia.dts    2007-12-21 
>>> 17:18:32.000000000 +0300
>>> @@ -324,6 +324,33 @@
>>>                  has-new-stacr-staopc;
>>>              };
>>>          };
>>> +
>>> +        PCI0: pci@1ec000000 {
>>> +            device_type = "pci";
>>> +            #interrupt-cells = <1>;
>>> +            #size-cells = <2>;
>>> +            #address-cells = <3>;
>>> +            compatible = "ibm,plb440epx-pci", "ibm,plb-pci";
>>> +            primary;
>>> +            reg = <1 eec00000 8    /* Config space access */
>>> +                   1 eed00000 4    /* IACK */
>>> +                   1 eed00000 4    /* Special cycle */
>>> +                   1 ef400000 40>;    /* Internal registers */
>>> +
>>> +            /* Outbound ranges, one memory and one IO,
>>> +             * later cannot be changed. Chip supports a second
>>> +             * IO range but we don't use it for now
>>> +             */
>>> +            ranges = <02000000 0 80000000 1 80000000 0 10000000

>>    I wonder why the AMCC's Sequoia/Rainier manual has PCI memory 
>> mapped at 0x80000000-0xbfffffff? The 0x80000000-0x8fffffff mapping was 
>> assumed by arch/ppc/ code.  What/why changed here?

> The addresses in the manual are relative to bus base.

    Hm, that's hard to infer from the manual, and even from arch/ppc/ sources...

> PCI controller is 
> located on the PLB and PLB base address is 0x100000000ULL on Sequoia.

     The question is where cam one read about that. :-)

> Older PPC code has ioremap64 function that did the 64 to 32-bit trick

    Ah, seeing fixup_bigphys_addr() at last -- it has escaped me before...

> It's been abolished. The kernel has support for 64-bit physical 
> addresses on 32-bit. IMHO there's no big reason to keep doing that 
> address trick. However, there are some drivers that use unsigned long 
> for storing physical addresses. This is wrong, since 
> pci_resource_start() returns a resource_size_t value. I think it's these 
> drivers that have to be fixed instead of adding workarounds to ppc4xx code.

    Well, I'm not arguing with that. Just tried to clarify the PCI mapping 
thing for myself. :-)

>>    As we now both know, having PCI memory space mapped beyound 4 GB 
>> makes some drivers misbehave as they use 'unsigned long' to store the 
>> result of pci_resource_start() and later ioremap() this truncated 
>> value -- which is 64-bit on Sequoia due to CONFIG_RESOURCE_64BIT=y 
>> that is needed to store the beyond-4GB addresses.

    Luckily, this one concerns the memory resources, as the I/O resources are 
actually limited to 'unsigned long' anyway...

WBR, Sergei