[U-Boot] [PATCH] fdt: Allow fdt_translate_address() to work with buses

[Przemyslaw Marczak]

Hello,

On 01/05/2016 06:26 PM, Stephen Warren wrote:
> On 01/05/2016 08:47 AM, Przemyslaw Marczak wrote:
>> Hello,
>>
>> On 01/05/2016 02:00 AM, Simon Glass wrote:
>>> Hi Stephen,
>>>
>>> On 4 January 2016 at 13:15, Stephen Warren <swarren at wwwdotorg.org>
>>> wrote:
>>>> On 01/03/2016 04:04 PM, Simon Glass wrote:
>>>>>
>>>>> It is common for I2C and SPI buses to have a single-cell address and a
>>>>> size
>>>>> of 0. These produce a warning at present. For example on snow:
>>>>>
>>>>> __of_translate_address: Bad cell count for gpc4
>>>>> __of_translate_address: Bad cell count for gpx0
>>>>> __of_translate_address: Bad cell count for gpv2
>>>>> __of_translate_address: Bad cell count for gpv4
>>>>>
>>>>> One of the nodes in question looks like this in part:
>>>>>
>>>>>          pinctrl_2: pinctrl at 10d10000 {
>>>>>                  #address-cells = <1>;
>>>>>                  #size-cells = <0>;
>>>>>                  gpv2: gpv2 {
>>>>>                          reg = <0x060>;
>>>>>                  };
>>>>>                  gpv4: gpv4 {
>>>>>                          reg = <0xc0>;
>>>>>                  };
>>>>>          };
>>>>>
>>>>> This is clearly valid so it looks like the conversion to use
>>>>> fdt_translate_address() in dev_get_addr() is not currently a good
>>>>> move.
>>>>
>>>>
>>>> To disable that, why not simply turn off CONFIG_OF_TRANSLATE on the
>>>> affected
>>>> platforms? That's precisely why that config option was introduced
>>>> when the
>>>> call to fdt_translate_address() was added to dev_get_addr()?
>>>>
>>>> That would prevent this patch from affecting platforms that don't
>>>> trigger
>>>> this issue, this leaving the valid check in place.
>>>
>>> But since this breaks normal behaviour we don't know what platforms
>>> are affected. We have made CONFIG_OF_TRANSLATE the default. So this
>>> approach doesn't seem (in effect) any better than Przemyslaw's newer
>>> series, below.
>>>
>>>>
>>>>> Przemyslaw Marczak sent three patches to resolve this for exynos
>>>>> boards:
>>>>>
>>>>> https://patchwork.ozlabs.org/patch/557008/
>>>>> https://patchwork.ozlabs.org/patch/557010/
>>>>> https://patchwork.ozlabs.org/patch/557009/
>>>>>
>>>>> But this involves creating a new function, and everyone will need to
>>>>> know
>>>>> when to use which one. Also the problem may affect other boards.
>>>>
>>>>
>>>> I suggest adding an extra parameter to dev_get_addr() (or whatever
>>>> calls it)
>>>> that indicates the root of the address space. The check on #size-cells
>>>> should be skipped for that one node (or level of translation) but
>>>> enabled
>>>> for all other levels. This way, there would be no need for anyone to
>>>> choose
>>>> between functions; there'd only be one. Most cases (i.e. translation
>>>> of MMIO
>>>> addresses) would simply pass 0 as the extra parameter (for the root
>>>> node),
>>>> but in special cases where it's known translation is not expected to
>>>> reach
>>>> the root MMIO space (e.g. I2C, SPI controllers), the controller node
>>>> would
>>>> be passed in.
>>>
>>> How would the caller know this root? It sounds plausible, but I do
>>> want to avoid complex rules. I think you are saying that buses that
>>> use their own address mechanism (i.e. not MMIO) must do something
>>> special. The current dev_get_addr() is really simple.
>>
>> Simon, Stephen
>>
>> As a summary of the issue, please tell me your opinion about this:
>>
>> Since the __of_translate_address() is called always with the "ranges" as
>> an argument, it looks reasonable to check it at the function beginning,
>> that the "ranges" property exists in the parent node.
>>
>> If not exists - then don't check the size-cells count and this should
>> fix the problem with additional argument.
>> This is simple and correct from specification point of view - which says
>> ranges and #size-cells property's - are not required [ePAPR v1.1].
>
> It /might/ indeed be reasonable to skip the check on #size-cells if
> there is no ranges property; a missing ranges property is already
> treated as a 1:1 translation as described by the comment at the top of
> of_translate_one(). Still, there are likely some edge-cases w.r.t. the
> "length" fields in the ranges property that would need to be thought
> through in more detail before I'd say for certain that this change does
> absolutely make sense.
>
> Either way though, making that change wouldn't solve the problem at all.
>
> It's simply not possible to translate an I2C device address beyond the
> root of the I2C address space; there is no equivalent of an I2C device's
> address in MMIO address space. The root of the problem is that the code
> is attempting to perform an I2C->MMIO address translation. If we prevent
> that (by capping any such translation at the root of the I2C address
> space for example), the problem is solved at that point. Once that's
> solved, the issue of translating across a boundary with #size-cells=0
> will never appear, and hence does not need to be solved.
>
>

The problem with I2C and I2C->MMIO address translation is solved by my 
new patch. The specification is clear for such case, if reg can't be 
mapped to it's parent address space, then don't use 'ranges' and this 
can be easy solved by additional check.

Please review my new patch:

https://patchwork.ozlabs.org/patch/564246/

Best regards,
-- 
Przemyslaw Marczak
Samsung R&D Institute Poland
Samsung Electronics
p.marczak at samsung.com