On 6/28/2012 7:50 AM, Stephen Warren wrote:

On 06/27/2012 06:57 PM, Mitch Bradley wrote:

quoted

On 6/27/2012 11:26 AM, Stephen Warren wrote:

quoted

I believe I've seen the following construct bandied about as the correct
way of representing a bunch of nodes that have the same name (since they
represent the same type of object) within device-tree.

     regulators {
         compatible = "simple-bus";
         #address-cells = <1>;
         #size-cells = <0>;

         regulator@0 {
             compatible = "regulator-fixed";
             reg = <0>;
...
         };

         regulator@1 {
             compatible = "regulator-fixed";
             reg = <1>;
...
         };
     };

However, when the kernel parses that, it issues messages such as:

prom_parse: Bad cell count for /regulators/regulator@0
prom_parse: Bad cell count for /regulators/regulator@1

The message comes from __of_translate_address(), which has the comment:

  * Note: We consider that crossing any level with #size-cells == 0 to mean
  * that translation is impossible (that is we are not dealing with a value
  * that can be mapped to a cpu physical address). This is not really
specified
  * that way, but this is traditionally the way IBM at least do things

So it seems that the problem only occurs if something tries to translate
the regulator's "reg" address to a CPU address.  Is that
possible/meaningful
in your case?

That's quite likely.

Note that the regulators node is compatible = "simple-bus", and I'm
doing that so that the child regulator nodes are automatically recursed
into, and a platform device created for each. Part of creating those
platform devices is to convert the reg and interrupts properties to
platform device resources, which is almost certainly what's calling
__of_translate_address(). This is a compatibility thing on ARM; I guess
pure OF-style drivers call something like of_get_address()/of_iomap()
themselves in their probe() function if appropriate, rather than relying
on calling platform_get_resource(), and hence forcing the DT parsing
code to convert resources beforehand in all cases, even if not used.

Perhaps simple-bus could be enhanced to detect when size-cells==0 and
know that this means the bus is really just a container/grouping of
non-addressed objects, and hence not provide the memory resources. Or,
perhaps we should introduce a new compatible value that only triggers
child instantiation but not resource setup, or something like that?

ISTM that if you want different semantics, a different compatible value 
is the best approach.  My mental model is that "simple-bus" has 
memory-mapped children. Perhaps something like "enumerated-bus" for 
non-memory-mapped ?