Re: [PATCH v3 6/6] drm/ingenic: Attach bridge chain to encoders

["H. Nikolaus Schaller" <hns@goldelico.com>]

Hi Laurent,

> Am 23.09.2021 um 12:03 schrieb Laurent Pinchart <laurent.pinchart@ideasonboard.com>:
> 
> Hi Nikolaus,
> 
> On Thu, Sep 23, 2021 at 11:55:56AM +0200, H. Nikolaus Schaller wrote:
>>> Am 23.09.2021 um 11:27 schrieb Laurent Pinchart:
>>> On Thu, Sep 23, 2021 at 11:19:23AM +0200, H. Nikolaus Schaller wrote:
>>>> 
>>>>>>> +		ret = drm_bridge_attach(encoder, &ib->bridge, NULL,
>>>>>>> +					DRM_BRIDGE_ATTACH_NO_CONNECTOR);
>>>>>> 
>>>>>> DRM_BRIDGE_ATTACH_NO_CONNECTOR makes it fundamentally incompatible
>>>>>> with synopsys/dw_hdmi.c
>>>>>> That driver checks for DRM_BRIDGE_ATTACH_NO_CONNECTOR being NOT present,
>>>>>> since it wants to register its own connector through dw_hdmi_connector_create().
>>>>>> It does it for a reason: the dw-hdmi is a multi-function driver which does
>>>>>> HDMI and DDC/EDID stuff in a single driver (because I/O registers and power
>>>>>> management seem to be shared).
>>>>> 
>>>>> The IT66121 driver does all of that too, and does not need
>>>>> DRM_BRIDGE_ATTACH_NO_CONNECTOR. The drm_bridge_funcs struct has
>>>>> callbacks to handle cable detection and DDC stuff.
>>>>> 
>>>>>> Since I do not see who could split this into a separate bridge and a connector driver
>>>>>> and test it on multiple SoC platforms (there are at least 3 or 4), I think modifying
>>>>>> the fundamentals of the dw-hdmi architecture just to get CI20 HDMI working is not
>>>>>> our turf.
>>>>> 
>>>>> You could have a field in the dw-hdmi pdata structure, that would
>>>>> instruct the driver whether or not it should use the new API. Ugly,
>>>>> I know, and would probably duplicate a lot of code, but that would
>>>>> allow other drivers to be updated at a later date.
>>>> 
>>>> Yes, would be very ugly.
>>>> 
>>>> But generally who has the knowledge (and time) to do this work?
>>>> And has a working platform to test (jz4780 isn't a good development environment)?
>>>> 
>>>> The driver seems to have a turbulent history starting 2013 in staging/imx and
>>>> apparently it was generalized since then... Is Laurent currently dw-hdmi maintainer?
>>> 
>>> "Maintainer" would be an overstatement. I've worked on that driver in
>>> the past, and I still use it, but don't have time to really maintain it.
>>> I've also been told that Synopsys required all patches for that driver
>>> developed using documentation under NDA to be submitted internally to
>>> them first before being published, so I decided to stop contributing
>>> instead of agreeing with this insane process. There's public
>>> documentation about the IP in some NXP reference manuals though, so it
>>> should be possible to still move forward without abiding by this rule.
>>> 
>>>>>> Therefore the code here should be able to detect if drm_bridge_attach() already
>>>>>> creates and attaches a connector and then skip the code below.
>>>>> 
>>>>> Not that easy, unfortunately. On one side we have dw-hdmi which
>>>>> checks that DRM_BRIDGE_ATTACH_NO_CONNECTOR is not set, and on the
>>>>> other side we have other drivers like the IT66121 which will fail if
>>>>> this flag is not set.
>>>> 
>>>> Ok, I see. You have to handle contradicting cases here.
>>>> 
>>>> Would it be possible to run it with DRM_BRIDGE_ATTACH_NO_CONNECTOR first
>>>> and retry if it fails without?
>>>> 
>>>> But IMHO the return value (in error case) is not well defined. So there
>>>> must be a test if a connector has been created (I do not know how this
>>>> would work).
>>>> 
>>>> Another suggestion: can you check if there is a downstream connector defined in
>>>> device tree (dw-hdmi does not need such a definition)?
>>>> If not we call it with 0 and if there is one we call it with
>>>> DRM_BRIDGE_ATTACH_NO_CONNECTOR and create one?
>>> 
>>> I haven't followed the ful conversation, what the reason why
>>> DRM_BRIDGE_ATTACH_NO_CONNECTOR can't always be use here ?
>> 
>> The synopsys driver creates its own connector through dw_hdmi_connector_create()
>> because the IP handles DDC/EDID directly.
> 
> That doesn't require creating a connector though. The driver implements
> drm_bridge_funcs.get_edid(), which is used to get the EDID without the
> need to create a connector in the dw-hdmi driver.

Ah, ok.

But then we still have issues.

Firstly I would assume that get_edid only works properly if it is initialized
through dw_hdmi_connector_create().

Next, in the current code, passing DRM_BRIDGE_ATTACH_NO_CONNECTOR to 
dw_hdmi_bridge_attach() indeed does not call dw_hdmi_connector_create()
but returns 0.

This patch 6/6 makes drm/ingenic unconditionally require a connector
to be attached which is defined somewhere else (device tree e.g. "connector-hdmi")
unrelated to dw-hdmi. Current upstream code for drm/ingenic does not init/attach
such a connector on its own so it did work before.

I.e. I think we can't just use parts of dw-hdmi.

If drm_bridge_attach() would return some errno if DRM_BRIDGE_ATTACH_NO_CONNECTOR
is set, initialization in ingenic_drm_bind() would fail likewise with "Unable to attach bridge".

So in any case dw-hdmi is broken by this drm/ingenic patch unless someone
reworks it to make it compatible.

Another issue is that dw_hdmi_connector_create() does not only do dcd/edid
but appears to detects hot plug and does some special initialization.
So we probably loose hotplug detect if we just use drm_bridge_funcs.get_edid().

I come to the conclusion that not creating a specific connector in dw-hdmi
and relying on a generic connector does not seem to be an option with current
code proposals.

In such a situation the question is what the least invasive surgery is to
avoid complications and lenghty regression tests on unknown platforms.
IMHO it is leaving (mature) dw-hdmi untouched and make attachment of a connector
in ingenic_drm_bind() depend on some condition.

BR and thanks,
Nikolaus