[PATCH v4 02/10] s5p-fimc: Add device tree support for FIMC devices

[Stephen Warren]

On 02/09/2013 03:29 PM, Sylwester Nawrocki wrote:
> On 02/09/2013 01:32 AM, Stephen Warren wrote:
>> On 02/08/2013 05:05 PM, Sylwester Nawrocki wrote:
>>> On 02/09/2013 12:21 AM, Stephen Warren wrote:
>>>> On 02/08/2013 04:16 PM, Sylwester Nawrocki wrote:
>>>>> On 02/07/2013 12:40 AM, Stephen Warren wrote:
>>>>>>> diff --git
>>>>>>> a/Documentation/devicetree/bindings/media/soc/samsung-fimc.txt
>>>>>>> b/Documentation/devicetree/bindings/media/soc/samsung-fimc.txt
>>>>>>
>>>>>>> +Samsung S5P/EXYNOS SoC Camera Subsystem (FIMC)
>>>>>>> +----------------------------------------------
>>>> ...
>>>>>>> +For every fimc node a numbered alias should be present in the
>>>>>>> aliases node.
>>>>>>> +Aliases are of the form fimc<n>, where<n>    is an integer (0...N)
>>>>>>> specifying
>>>>>>> +the IP's instance index.
...
>>> Different compatible values might not work, when for example there
>>> are 3 IPs out of 4 of one type and the fourth one of another type.
>>> It wouldn't even by really different types, just quirks/little
>>> differences between them, e.g. no data path routed to one of other IPs.
>>
>> I was thinking of using feature-/quirk-oriented properties. For example,
>> if there's a port on 3 of the 4 devices to connect to some other IP
>> block, simply include a boolean property to indicate whether that port
>> is present. It would be in 3 of the nodes but not the 4th.
> 
> Yes, I could add several properties corresponding to all members of this
> [3] data structure. But still it is needed to clearly identify the IP
> block in a set of the hardware instances.

Why? What decisions will be made based on the identify of the IP block
instance that wouldn't be covered by DT properties that describe which
features/bugs/... the IP block instance has?

>>> Then to connect e.g. MIPI-CSIS.0 to FIMC.2 at run time an index of the
>>> MIPI-CSIS needs to be written to the FIMC.2 data input control register.
>>> Even though MIPI-CSIS.N are same in terms of hardware structure they
>>> still
>>> need to be distinguished as separate instances.
>>
>> Oh, so you're using the alias ID as the value to write into the FIMC.2
>> register for that. I'm not 100% familiar with aliases, but they seem
>> like a more user-oriented naming thing to me, whereas values for hooking
>> up intra-SoC routing are an unrelated namespace semantically, even if
>> the values happen to line up right now. Perhaps rather than a Boolean
>> property I mentioned above, use a custom property to indicate the ID
>> that the FIMC.2 object knows the MIPI-CSIS.0 object as? While this seems
> 
> That could be 'reg' property in the MIPI-CSIS.0 'port' subnode that
> links it to the image sensor node ([4], line 165). Because MIPI-CSIS IP
> blocks are immutably connected to the SoC camera physical MIPI CSI-2
> interfaces, and the physical camera ports have fixed assignment to the
> MIPI-CSIS devices..  This way we could drop alias ID for the MIPI-CSIS
> nodes. And their instance index that is required for the top level
> driver which exposes topology and the routing capabilities to user space
> could be restored from the reg property value by subtracting a fixed
> offset.

I suppose that would work. It feels a little indirect, and I think means
that the driver needs to go find some child node defining its end of
some link, then find the node representing the other end of the link,
then read properties out of that other node to find the value. That
seems a little unusual, but I guess it would work. I'm not sure of the
long-term implications of doing that kind of thing. You'd want to run
the idea past some DT maintainers/experts.

...
> I can see aliases used in bindings of multiple devices: uart, spi, sound
> interfaces, gpio, ... And all bindings seem to impose some rules on how
> their aliases are created.

Do you have specific examples? I really don't think the bindings should
be dictating the alias values.

>> After all, what happens in some later SoC where you have two different
>> types of module that feed into the common module, such that type A
>> sources have IDs 0..3 in the common module, and type B sources have IDs
>> 4..7 in the common module - you wouldn't want to require alias ISs 4..7
>> for the type B DT nodes.
> 
> There is no need to write alias ID directly into registers, and actually
> it doesn't really happen. But we need to know that, for example camera A
> is connected to physical MIPI CSI-2 channel 0 and to capture video with
> DMA engine of FIMC.2 we need to set FIMC.2 input register to link it to
> MIPI-CSIS 0.

OK, so the IDs are selecting which register to write, or which mux
settings to access. That's pretty much semantically the same thing.