Arnd, thanks for the comments.

On 07/29/2014 03:00 PM, Arnd Bergmann wrote:

On Tuesday 29 July 2014 14:42:31 Stanimir Varbanov wrote:

quoted

        taddr = of_translate_address(dev, addrp);
-       if (taddr == OF_BAD_ADDR)
-               return -EINVAL;
+       /*
+        * if the address is non-translatable to cpu physical address
+        * fallback to a IORESOURCE_REG resource.
+        */
+       if (taddr == OF_BAD_ADDR) {
+               memset(r, 0, sizeof(*r));
+               taddr = of_read_number(addrp, 1);
+               if (taddr == OF_BAD_ADDR)
+                       return -EINVAL;
+               r->start = taddr;
+               r->end = taddr + size - 1;
+               r->flags = IORESOURCE_REG;
+               r->name = name ? name : dev->full_name;
+               return 0;
+       }
+

I don't think that everything returning OF_BAD_ADDR makes sense
to turn into IORESOURCE_REG. It could be an e.g. invalid DT
representation, a node with #size-cells=<0>, or it could be
something that gets translated one or more nodes up in the
tree before it reaches a bus without a ranges property.

Also, you should not rely on #address-cells being hardcoded
to <1> above.

This was just an example. Of course it has many issues and probaly it is
wrong:) The main goal was to understand does IORESOURCE_REG resource
type and parsing the *reg* properties for non-translatable addresses are
feasible. And also does it acceptable by community and OF platform
maintainers.

How about modifying of_get_address() rather than
__of_address_to_resource() instead? You could introduce
a new of_bus entry for each bus you expect to return
an IORESOURCE_REG, or you could change of_bus_default_get_flags
to return IORESOURCE_REG if the parent node has no ranges property
and is not the root node.

IMO the clearer solution is to introduce a new of_bus bus. In that case
one possible problem will be how to distinguish the non-translatable and
the other buses. Also the *device_type* property is deprecated for non
PCI devices.

The second option will need to change the prototype of .get_flags method
to receive device_node structure.

Thoughts?

-- 
regards,
Stan