Re: [PATCH] of: support an enumerated-bus compatible value

[Mitch Bradley <wmb-D5eQfiDGL7eakBO8gow8eQ@public.gmane.org>]

On 7/2/2012 5:59 AM, Stephen Warren wrote:
> On 07/01/2012 01:36 PM, Rob Herring wrote:
>> On 06/28/2012 06:05 PM, Stephen Warren wrote:
>>> From: Stephen Warren <swarren-DDmLM1+adcrQT0dZR+AlfA@public.gmane.org>
>>>
>>> An "enumerated" bus is one that is not memory-mapped, hence hence
>>> typically has #address-cells=1, and #size-cells=0. Such buses would be
>>> used to group related non-memory-mapped nodes together, often just under
>>> the top-level of the device tree. The ability to group nodes into a non-
>>> memory-mapped subnode of the root is important, since if nodes exist to
>>> describe multiple entities of the same type, the nodes will have the
>>> same name, and hence require a unit address to differentiate them. It
>>> doesn't make sense to assign bogus unit addresses from the CPU's own
>>> address space for this purpose. An example:
>>>
>>> 	regulators {
>>> 		compatible = "enumerated-bus";
>>> 		#address-cells = <1>;
>>> 		#size-cells = <0>;
>>>
>>> 		regulator@0 {
>>> 			compatible = "regulator-fixed";
>>> 			reg = <0>;
>>> 		};
>>>
>>> 		regulator@1 {
>>> 			compatible = "regulator-fixed";
>>> 			reg = <1>;
>>> 		};
>>> 	};
>>>
>>> Finally, because such buses are not memory-mapped, we avoid creating
>>> any IO/memory resources for the device.
>>
>> This seems like a work-around to use reg instead of using cell-index
>> (which is discouraged). reg in this case is really not a hardware
>> description. Do you have an intended use or just trying to fix the error
>> messages?
>
> I'm not familiar with cell-index; can you please describe it some more.
> Looking at some existing files in arch/powerpc/boot/dts, it looks like
> something that exists alongside reg rather than replacing it, so I don't
> see how it'd solve the problem.
>
> The portion of .dts file quoted above is the use-case. In more general
> terms, I need to add a bunch of non-memory-mapped devices to DT. There
> are multiple devices of a given type. The DT node names should be named
> after the class of device not the instance, and hence all get named the
> same. Hence, I need a unit address to differentiate the node names.
> Hence I need to use the reg property in order that the unit address
> matches the reg property. Is there some other way of solving these
> requirements other than using a unit address to make the node names unique?

One of Rob's objections was that, in this case, the reg property is not 
a hardware description.  That's an interesting point.  If in fact 
numerous such devices exist, there must be some mechanism for software 
to choose which device to talk to, typically a number.  Is that the case 
for these devices?  If so, that number is a perfectly valid "reg" 
property value.  If not, how does software choose to talk to a specific 
device?

>
> On 07/01/2012 04:03 PM, Grant Likely wrote:
> ...
>> Besides; if they are enumerated, non-memory mapped devices, then is it
>> really appropriate to use platform_{device,driver}? I don't think it
>> is.


Maybe I'm missing some implication of "platformness", but I don't see 
why not being memory-mapped would disqualify a device.  Perhaps Grant, 
or someone else, could elucidate?

>
> Hmm, well /everything/ that gets instantiated from DT is a platform
> device at present, at least for the platforms and bus types we're using
> on Tegra and I believe all/most ARM platforms, except some small amounts
> of AMBA. Changing that would hugely impact a ton of working code and
> just be churn in my opinion. Do we really want to invent another device
> type internal to Linux for this? Besides, doing that would be orthogonal
> to this patch; such a change would have no impact on the DT
> representation of the devices.
>