[PATCH] media: add V4L2 DT binding documentation

Stephen Warren swarren at wwwdotorg.org
Wed Sep 12 18:59:59 EDT 2012


On 09/12/2012 03:17 PM, Guennadi Liakhovetski wrote:
> On Wed, 12 Sep 2012, Stephen Warren wrote:
> 
>> On 09/12/2012 01:28 PM, Guennadi Liakhovetski wrote:
>>> Hi Stephen
>>>
>>> On Wed, 12 Sep 2012, Stephen Warren wrote:
>>>
>>>> On 09/11/2012 09:51 AM, Guennadi Liakhovetski wrote:
>>>>> This patch adds a document, describing common V4L2 device tree bindings.
>>>>>
>>>>> Co-authored-by: Sylwester Nawrocki <s.nawrocki at samsung.com>
>>>>> Signed-off-by: Guennadi Liakhovetski <g.liakhovetski at gmx.de>
>>>>
>>>> Overall, I think this looks pretty reasonable, so:
>> ...
>>>>
>>>>> +			clock-frequency = <50000000>;	/* max clock frequency */
>>>>> +			clock-output-names = "mclk";
>>>>> +		};
>>>>> +
>>>>> +		port {
>>>> ...
>>>>> +			ceu0_0: link at 0 {
>>>>> +				reg = <0>;
>>>>> +				remote = <&csi2_2>;
>>>>> +				immutable;
>>>>
>>>> Did we decide "immutable" was actually needed? Presumably the driver for
>>>> the HW in question knows the HW isn't configurable, and would simply not
>>>> attempt to apply any configuration even if the .dts author erroneously
>>>> provided some?
>>>
>>> Well, I've been thinking about this today. I think, bridge drivers will 
>>
>> Sorry, I'm not sure what a "bridge" driver is; is it any v4l2 device?
> 
> Right, sorry for not explaining. We call a bridge a device, that's sitting 
> on a bus like USB, PCI or - as in the SoC case - on an internal one and 
> interfacing to, say, a video sensor or a TV decoder or encoder. In the SoC 
> case most primitive bridges just take data from video sensors attached 
> over, say, an 8-bit parallel bus and DMA it into RAM buffers. Some bridges 
> can also perform some data processing.
> 
>>> call centrally provided helper functions to enumerate links inside ports. 
>>
>> Presumably a given driver would only parse the ports/links of its own DT
>> node, and hence would be able to provide a parameter to any helper
>> function that indicated the same information that "immutable" would?
> 
> Well, let's drop "immutable" for now, since we don't have a good idea 
> whether we need it or not. If we do need it, we'll add it later, removing 
> a part of a standart is more difficult.
> 
>>> While doing that they will want to differentiate between links to external 
>>> devices with explicit configuration, and links to internal devices, whose 
>>> configuration drivers might be able to figure out themselves. How should a 
>>> driver find out what device this link is pointing to? Should it follow the 
>>> "remote" phandle and then check the "compatible" property? The word 
>>> "immutable" is a hint, that this is a link to an internal device, but it 
>>> might either be unneeded or be transformed into something more 
>>> informative.
>>
>> I would imagine that a given driver would only ever parse its own DT
>> node; the far end of any link is purely the domain of the other driver.
>> I thought that each link node would contain whatever hsync-active,
>> data-lanes, ... properties that were needed to configure the local device?
> 
> I think, information needed to configure a bridge device to connect to a 
> camera sensor, like sync polarities etc., might not be sufficient to 
> configure it to interface to an SoC internal unit, like a MIPI CSI-2 
> interface. I can see 2 possibilities to recognise, that this link is going 
> to an internal device: (1) follow the remote phandle, (2) use a 
> vendor-specific property to specify the remote device type. I guess, you'd 
> prefer the latter?

I guess I'm still not exactly clear on what/why the bridge device needs
to know; a concrete example might help.

Either way though, (2) sounds like it's probably best; all information
required to configure a given device is within that device's node.

Being pedantic, I'm not sure precisely what you mean by vendor-specific.
If you mean the DT property name would include a vendor prefix, then
yes, I imagine that's quite possible, although it'd be good to try and
create standardized properties that multiple vendors and devices can use
if it makes sense. If you mean something more than that, then I'd argue
that those properties aren't just vendor-specific, but rather specific
to the individual binding for the individual device (which does imply
the vendor prefix, but not the general applicability of the property to
all of a vendor's devices).



More information about the linux-arm-kernel mailing list