Re: [PATCH v5 2/7] drm/ingenic: Add support for JZ4780 and HDMI output

[H. Nikolaus Schaller]

Hi Paul,

> Am 18.01.2022 um 17:58 schrieb Paul Cercueil <paul@xxxxxxxxxxxxxxx>:
> 
> Hi Nikolaus,
> 
> Le mar., janv. 18 2022 at 15:50:01 +0100, H. Nikolaus Schaller <hns@xxxxxxxxxxxxx> a écrit :
>> Hi Paul,
>> any insights on the JZ_REG_LCD_OSDC issue below?
> 
> Sorry, I missed your previous email. I blame the holidays ;)

No problem... We all had deserved them.

> 
>> There is a second, unrelated issue with the introduction of
>> "drm/bridge: display-connector: implement bus fmts callbacks"
>> which breaks bus format negotiations.
>> These are the two last unsolved issues to submit a fully working driver.
>>> Am 22.12.2021 um 15:03 schrieb H. Nikolaus Schaller <hns@xxxxxxxxxxxxx>:
>>>> Am 08.11.2021 um 10:37 schrieb Paul Cercueil <paul@xxxxxxxxxxxxxxx>:
>>>> Hi Nikolaus,
>>>> Le dim., nov. 7 2021 at 21:25:38 +0100, H. Nikolaus Schaller <hns@xxxxxxxxxxxxx> a écrit :
>>>>> Hi Paul,
>>>>>>>>> @@ -1274,7 +1319,7 @@ static int ingenic_drm_bind(struct device *dev, bool has_components)
>>>>>>>>> 	/* Enable OSD if available */
>>>>>>>>> 	if (soc_info->has_osd)
>>>>>>>>> -		regmap_write(priv->map, JZ_REG_LCD_OSDC, JZ_LCD_OSDC_OSDEN);
>>>>>>>>> +		regmap_set_bits(priv->map, JZ_REG_LCD_OSDC, JZ_LCD_OSDC_OSDEN);
>>>>>>>> This change is unrelated to this patch, and I'm not even sure it's a valid change. The driver shouldn't rely on previous register values.
>>>>>>> I think I already commented that I think the driver should also not reset
>>>>>>> previous register values to zero.
>>>>>> You did comment this, yes, but I don't agree. The driver *should* reset the registers to zero. It should *not* have to rely on whatever was configured before.
>>>>>> Even if I did agree, this is a functional change unrelated to JZ4780 support, so it would have to be splitted to its own patch.
>>>>> Well it is in preparation of setting more bits that are only available for the jz4780.
>>>>> But it will be splitted into its own patch for other reasons - if we ever make osd working...
>>>>>>> If I counted correctly this register has 18 bits which seem to include
>>>>>>> some interrupt masks (which could be initialized somewhere else) and we
>>>>>>> write a constant 0x1.
>>>>>>> Of course most other bits are clearly OSD related (alpha blending),
>>>>>>> i.e. they can have any value (incl. 0) if OSD is disabled. But here we
>>>>>>> enable it. I think there may be missing some setting for the other bits.
>>>>>>> So are you sure, that we can unconditionally reset *all* bits
>>>>>>> except JZ_LCD_OSDC_OSDEN for the jz4780?
>>>>>>> Well I have no experience with OSD being enabled at all. I.e. I have no
>>>>>>> test scenario.
>>> It turns out that the line
>>> 		ret = clk_prepare_enable(priv->lcd_clk);
>>> initializes JZ_REG_LCD_OSDC to 0x00010005 (i.e. printk tells 0x0 before).
>> more detailled test shows that it is the underlying
>> 	clk_enable(priv->lcd_clk)
>> (i.e. not the prepare).
>>> and writing
>>> 		regmap_write(priv->map, JZ_REG_LCD_OSDC, JZ_LCD_OSDC_OSDEN);
>>> overwrites it with 0x00000001.
>>> This makes the HDMI monitor go/stay black until I write manually 0x5 to
>>> JZ_REG_LCD_OSDC.
>>> This means that JZ_LCD_OSDC_ALPHAEN must be enabled on jz4780 as well.
>>> Hence overwriting just with JZ_LCD_OSDC_OSDEN breaks it.
>>> Now the questions:
>>> a) 0x5 is understandable that it sets JZ_LCD_OSDC_ALPHAEN - but why is it needed?
>>>   Is this a not well documented difference between jz4740/60/70/80?
>>> b) how can clk_prepare_enable() write 0x00010005 to JZ_REG_LCD_OSDC? Bug or feature?
>>>   Something in cgu driver going wrong?
>> I now suspect that it is an undocumented SoC feature.
> 
> Not at all. If the clock is disabled, the LCD controller is disabled, so all the registers read zero, this makes sense. You can only read the registers when the clock is enabled. On some SoCs, reading disabled registers can even cause a complete lockup.

This is the right answer to the wrong question :)
The question is why the register becomes 0x10005 as soon as the clock is enabled. Without writing to it.  And contrary to the documented reset state.
Or are you aware that u-boot initialized the lcdc and clocks get disabled when/during starting the kernel (I am using the good old v2013.10)?
That could be an explanation that we read 0 before the clock is enabled and 0x10005 after.

> Why is this JZ_LCD_OSDC_ALPHAEN bit needed now? I remember it working fine last time I tried, and now I indeed get a black screen unless this bit is set. The PM doesn't make it obvious that the bit is required,

exactly.

> but that wouldn't be surprising.
> 
> Anyway, feel free to send a patch to fix it (with a Fixes: tag). Ideally something like this:
> 
> u32 osdc = 0;
> ...
> if (soc_info->has_osd)
>   osdc |= JZ_LCD_OSDC_OSDEN;
> if (soc_info->has_alpha)
>   osdc |= JZ_LCD_OSDC_ALPHAEN;
> regmap_write(priv->map, JZ_REG_LCD_OSDC, osdc);

Looks good to me. I'll give it a try.

> 
> This also ensures that the OSDC register is properly initialized in the !has_osd case. The driver shouldn't rely on pre-boot values in the registers.

Ok. Not all SoC have osd.

BR and thanks,
Nikolaus