RFC: book3e: DCR MMIO for IBM Blue Gene/Q

[Jimi Xenidis]

On Nov 4, 2012, at 5:51 PM, Benjamin Herrenschmidt <benh at kernel.crashing.org> wrote:

> On Sun, 2012-11-04 at 10:32 -0600, Jimi Xenidis wrote:
>> I want to use the stuff in arch/platform/sysdev/dcr.c and I have a couple of questions:
>> 1) anyone have a good devtree binding for this?
> 
> Not really, I think the cell stuff was the last we looked at this.
> 
>> I'm thinking:
>>    bgq {
>>        #address-cells = <2>;
>>        #size-cells = <2>;
>>        ...
>> 
>>        dcr {
>>            reg = <0x3ff 0xe0000000 0x0 0x08000000>;
>>            compatible = "ibm,bgq-dcr";
>>            dcr-controller;
>>            dcr-access-method = "mmio";
>>            dcr-mmio-stride = <??>;
>>        };
>>    };
>> 
>> I'm not sure what stride I will use since I have yet to figure out the DCR numbering in play at the moment, which will be painful.
>> Also, some clarity to the following would be nice:
>>    /* Stride is not properly defined yet, default to 0x10 for Axon */
>>    p = of_get_property(dp, "dcr-mmio-stride", NULL);
>>    stride = (p == NULL) ? 0x10 : *p;
>> 
>>    /* XXX FIXME: Which property name is to use of the 2 following ? */
>>    p = of_get_property(dp, "dcr-mmio-range", NULL);
>>    if (p == NULL)
>>        p = of_get_property(dp, "dcr-mmio-space", NULL);
>>    if (p == NULL)
>>        goto done;
> 
> Not sure what SLOF did back then. dcr-mmio-range sounds better to me.
> 
>> I guess my "reg" property is not useful to the code?
> 
> Right, unless the DCR "bridge" has some registers of its own. It's also
> handy to have as a "standard" way of representing the memory used by the
> DCR bridge, tho that does mean duplication... but then it allows you to
> have a nice unit address.

umm, was there a conclusion here? :-)
-jx

> 
>> Is there a preference over dcr-mmio-range/space or in the absence of these to we should use "reg"?
>> 
>> 2) I need 64 bit reads and writes
>> My intention is to have dcr_{read,write}_generic_{mmio,native} take an unsigned long for value all the way down.
>> For the native case, this would make is similar to mtspr and mfspr.
>> For the MMIO case, the selected in/out size would be based on "#ifdef __powerpc64__"
> 
> As long as it doesn't break Axon...
> 
> Cheers,
> Ben.
> 
>> Thoughts?
>> 
>> -jx
> 
>