[Intel-gfx] [PATCH 1/4] drm/i915: Add more dev ops for MIPI sub encoder

[Jani Nikula]

On Wed, 23 Oct 2013, Shobhit Kumar <shobhit.kumar at intel.com> wrote:
> Hi Jani,
> On 10/22/2013 05:23 PM, Jani Nikula wrote:
>> On Tue, 22 Oct 2013, Shobhit Kumar <shobhit.kumar at intel.com> wrote:
>>> On 10/21/2013 6:57 PM, Jani Nikula wrote:
>>>>
>>>> Hi Shobhit -
>>>>
>>>> On Mon, 21 Oct 2013, Shobhit Kumar <shobhit.kumar at intel.com> wrote:
>>>>> Also add new fields in intel_dsi to have all dphy related parameters.
>>>>> These will be useful even when we go for pure generic MIPI design
>>>>
>>>> I feel like we have a different idea of what the ideal generic design
>>>> is. For me, the goal is that we change the struct intel_dsi_device to
>>>> struct drm_bridge, and those drm_bridge drivers are all about the panel,
>>>> and have as few details about i915 or our hardware as possible. Having
>>>> the bridge drivers fill in register values to be written by the core DSI
>>>> code does not fit that. Yes, I had some of those in my bridge conversion
>>>> patches too, but I did not intend we'd keep adding more.
>>>>
>>>> I'd rather we provide generic helpers the bridge driver can call.
>>>
>>> Yeah, look like our ideas are different. In your goal with drm_bridge
>>> architecture, we will still end up having multiple bridge drivers for
>>> each different panel. But my goal is to have a single driver which can
>>> work for multiple panels.
>>
>> I'm trying to look one or two steps further, and what it will mean to
>> the driver. Here's the long term goal in upstream as I see it: There
>> will be a framework in place that allows one to write a (DSI) panel
>> driver once, using generic APIs, and use that panel driver with any SoC
>> (that implements the other side of the framework).
>
> To clarify in terms of what we have currently, in my opinion intel_dsi.c 
> is one such SoC specific implementation for BYT and associated MIPI host 
> controller. And sub-encoder drivers are other end and I want to unify 
> sub-encoder side to have just one sub-encoder for any panel out there. 
> What you are talking makes perfect sense if we are going to have each 
> panel driver as a separate driver. But we can still achieve this 
> separation with common panel driver as well.
>
>>
>> We are obviously far away from that goal at the moment. But IMHO we
>> should keep that in mind as a guide to what we are doing now. Moving
>> towards a model with a clearly defined API between the DSI core and the
>> panels, where the panel specific things are abstracted away from the
>> core, or towards a model where the core and panel driver depend on the
>> implementation of each other, communicating via variables.
>
> I think we are aligned on the goal and I feel there is a need for common 
> DSI core to be separate
>
>>
>>> Since we already have enabled some panels with sub-encoder
>>> architecture for completeness I was planning to maintain generic
>>> driver as one sub-encoder.
>>
>> Nothing prevents you from doing that, as long as the separation between
>> the core and the panel drivers remains clear.
>>
>>> But actually we can do away with all sub-encoder and do not need even
>>> drm_bridge and all implementation will be in intel_dsi.c. Panel
>>> specific configurations or sequences will come from VBT which I have
>>> tried to convert as parameters.
>>
>> With this model it is all too easy to forget what is the panel driver
>> and what is the SoC driver. They *are* two separate things, and should
>> not be mixed. It will be all too easy to keep adding new parameters and
>> conditions in the code as new panel drivers appear to need them. It will
>> lead to code that is very difficult to understand and maintain.
>
> I think here you have misunderstood my proposal. I still treat SoC 
> driver and actual Panel driver as separate. And whatever parameters I 
> have tried to add are all DSI/DPHY spec related. There is not even 
> single parameter which is panel specific. If you are confusing this 
> because of use of -
>
> 	if (intel_dsi->dsi_clock_freq)
> 		dsi_clk = intel_dsi->dsi_clock_freq;
>
> I feel it is okay to have this parameter which provides provision for 
> spec parameters to be hard-coded instead of calculating if a panel needs 
> and this is *only* such parameter and I already have ways to remove it 
> as well. IMHO it is a small price to pay to get one generic panel driver.
>
> That is another thing if the DSI/DPHy specs themselves evolve and we 
> need to modify our core, but that is unavoidable I guess.
>
>>
>> A model similar to what I'm proposing has also been tried and tested,
>> with several panels: drivers/video/omap2/. It's not DRM, and the control
>> is in the panel drivers, but the separation is extremely clear (panel
>> drivers are separate kernel modules).
>
> Understand, but again my idea is to have one single driver which can 
> work for all panels and hence there is no need of multiple panel 
> drivers. This works with SoC side of the framework. I have not yet 
> described how we can achieve this, but first does it has any merit to 
> have something like this ?

My educated guess is that a single panel driver is not going to work for
*all* panels, *but* it is of course benefitial to share each panel
driver for as many panels as technically makes sense.

So yes, I agree there's merit in having a panel driver that works on a
bunch of panels based on parameters from the VBT. Apologies if it
sounded like I'm against that.

Based on my (limited, but not insignificant) experience with DSI panels,
you can't parametrize everything, and you will need panel specific code
to handle them. I presume there will also be cases where we won't have
the parameters in VBT. The DSI specs do not cover everything, and some
panels are outright out of spec. And then there are DSI->LVDS bridges
etc. needing special attention.

So I think let's keep trying to find the right abstractions to separate
the DSI core and the panel drivers, make it possible to support several
panels with one driver, and make it possible to have independent drivers
for panels that don't fit the assumptions of the generic panel driver.

Does that conflict with your goals? Are we in agreement here?

>> No doubt the clear separation between the core and the panel drivers
>> will be harder and more work in the short term, but it will pay off in
>> the long term. And it doesn't all have to happen at once, as long as we
>> work *towards* that goal, not away from it.
>
> I think we should take into account the amount of effort required to 
> develop and maintain bridge drivers for tens of MIPI panel out there Vs 
> having one panel driver to maintain and make fully spec compliant taking 
> care of open ends left by the specs in the best way we can to achieve 
> this generality.

See above. I'm not proposing you split it all out to separate
drivers. Again, I'm sorry if I haven't been clear about that.


Now, the question is how do we achieve all this. First things first,
let's try to get any straightforward changes merged (small patches!), so
we can at least narrow the gap to upstream. And I need to have a look at
your panel driver code.


BR,
Jani.


-- 
Jani Nikula, Intel Open Source Technology Center