Hi Stefan,

On 1 September 2015 at 09:23, Stefan Roese sr@denx.de wrote:

Hi Simon,

On 01.09.2015 08:25, Stefan Roese wrote:

...

I'm currently enabling DM support for the Marvell MVEBU SoC's I've been working lately on (Armada XP and 38x right now). A problem I'm facing here is the bus translation, as this is quite complex for these SoC's. With multiple levels of translation ranges (multiple simple-bus nodes to walk through).

The current implementation in dev_get_addr() does not work. This has 2 reasons:

a) It only translates 1 simple-bus node - we need to walk the complete DT to get the correct address.

b) The "ranges" property can have different sizes for "address" and "size" in all these busses / nodes. And also multiple ranges tuples are allowed and used in these dts files.

Of course this can be solved in the current implementation (device.c / simple_bus.c). But especially b) is not that trivial to solve in a generic way.

So my main question is, why don't you use fdt_translate_address() from fdt_support.c instead of implementing your own translation function simple_bus_translate()? Is this a size question because this may be used in SPL as well?

The attached small patch demonstrates the usage of fdt_translate_address() and enables bus translation on the MVEBU platforms.

Thanks for digging into this.

One concern I have is fdt_parent_offset(). That function is very slow - we should use dev->parent->of_offset instead with driver model.

Other than that I think this is fine. But yes it should be enabled by CONFIG because the code size gets a lot larger and many platforms don't need it. Perhaps we should have a new CONFIG_SPL_OF_TRANSLATE?

Regards, Simon