[U-Boot] Displays on i.MX (was Re: [PATCH 1/2])

[Eric Nelson]

Hi Stefano,

I changed the subject line because we've veered off-topic.

On 09/23/2012 10:27 AM, Stefano Babic wrote:
> On 23/09/2012 18:46, Eric Nelson wrote:
>> Hi Stefano,
>>
>> On 09/23/2012 08:56 AM, Stefano Babic wrote:
>>> On 22/09/2012 16:37, Fabio Estevam wrote:
>>>> On Sat, Sep 22, 2012 at 10:42 AM, Otavio Salvador
>>>> <otavio at ossystems.com.br>   wrote:
>>>>> Hello Eric,
>>>>>
>>>>> On Fri, Sep 21, 2012 at 5:36 PM, Eric Nelson
>>>>> <eric.nelson at boundarydevices.com>   wrote:
>>>>>> Signed-off-by: Eric Nelson<eric.nelson at boundarydevices.com>
>>>>>
>>>>> Did you test it in mx5 too? We seem to need to handle it in mx5 too as
>>>>> we had hungs in FSL kernel when using framebuffer in U-Boot. We're
>>>>> using a patch in kernel for workaround it but it seems your fix does
>>>>> what is need.
>>>>
>>>> I have just tested Eric's series on a mx53loco and it does fix the
>>>> kernel hang issue.
>>>>
>>>> I made some comments on this series and hopefully Eric's v2 can get
>>>> into 2012.10, since this is a bug fix.
>>>
>>> Ok, I am waiting for V2 and I will push it.
>>>
>>
>> I'll forward this later today.
>
> Ok
>
>>
>>> Anyway, a question about the issue. It seems to me that it is not
>>> possible with IPUV3 (I have not tested myself, so my question) to get
>>> the u-boot splashscreen displayed on the LCD until the kernel has
>>> finished to boot. This could be possible (and it is possible on other
>>> SOC) if the IPU is loaded as module instead of statically linked to the
>>> kernel, and if the kernel does not touch the IPU setup. This means also
>>> that it should not disable the clocks used by U-Boot for the IPU.
>>>
>>
>> I'm not sure I understand. The splash screen comes up as soon as
>> the call to ipuv3_fb_init() is made (in board_video_skip() in my
>> implementation for SABRE Lite).
>>
>> As it stands, if we leave the IPU running, we'll see garbage on
>> the display as the kernel re-purposes the RAM used by U-Boot's
>> frame buffer.
>
> Right, if the kernel reuse the same memory. I am aware that it is not
> implemented, I am asking if there some reason to make it impossible.
>
> Some customers want to have a picture shown on the LCD until their
> application is running. This makes sense, because the application can
> take a lot of time before displaying something on the LCD.
>
> If we reserve some memory for this scope, that is not used by the kernel
> later (passing the mem= parameter, for example, or using a .reserve
> entry in the board initialization code), we can reach the goal. In
> U-Boot we have a single display with 16bit, that means that the memory
> consumption is not very high.
>
>>
>>> But I understand from your patch that this way is not possible on
>>> iMX53/MX6, and IPU must be always disabled. Is this correct ?
>>>
>>
>> Sascha responded to a note about this on AKML that the hand-over of
>> a live FB isn't a supported kernel use case and it's definitely
>> tricky.
>
> And I agree with him, the handover is tricky and not easy, I mean, it
> can work with a SOC (it remains tricky..) but not with another one. What
> I am saying is not this, but what happens if the IPU is not touched
> until the IPU modules are loaded.
>
>>
>> I don't know about the policy, but from a practical matter, the
>> IPU frame buffer implementation in U-Boot isn't currently up to
>> that task, since:
>>      it only supports a single display (i.MX6 can handle 4)
>>      it only supports 16bpp
>>
>> Additional functionality would be helpful here.
>
> Agree with you. And this is the reason I am not supposing that the
> kernel takes the IPU setup made by U-Boot. It could be a nightmare.
>
>>
>> I would like to see a handoff of display settings from U-Boot to
>> the kernel, but that's also a tricky thing as long as we're supporting
>> different mechanisms (DT in main-line and kernel parameters in
>> older kernels).
>
> Yes, and a lot of other things. I know Anatolji implemented this
> behavior for a PPC5121, but we cannot generalize. I agree that the
> handoff is difficult and not maintainable. My question is different: if
> the IPU drivers in kernel are compiled as modules, and I will load them
> only after booting, and the framebuffer's memory is reserved so that the
> kernel does not touch it, is there still a known reason because the IPU
> should not run when we boot the kernel? I know this issue with USB,
> maybe we have now the same with the IPU.
i
> Note: this has nothing to do with this patch ;-). I will merge it into
> the current release when you push V2.
>

I agree with all of your goals but have a more mundane set for U-Boot
and i.MX6:

-- HDMI support
-- Better strategy for storing display setup. The implementation
of a compiled-in list of supported displays doesn't scale.
-- Hand-off of display setup to DT-enabled kernel (Sascha's drivers)
-- Multiple displays
-- Hand-off of display setup to older kernels

My thought is to take things in more or less this order, although
a couple of things can make the last one trivial to do in a boot
script and it's a pain point for me. (My blog post on setting up
bootargs doesn't seem to be doing the trick)

My thinking is that:
	HDMI support can and should support EDID and flexible choice
	of resolution, and

	We'll need to store that choice somewhere, and

	The manner of display should be enough for the kernel to
	configure at least the screen resolution.

Some of this really belongs in the boot loader where we're dealing
with the board-specifics. It seems entirely appropriate for a
board to tell the kernel that it wants:

	HDMI on IPU0/DI0 at 1920x1080
	LVDS on IPU0/DI1 at 1024x768 using SPWG standard, and
	PRGB on IPU1/DI0 at 800x480 with inverted pixel clock
		and 18 bits of output

Handing this off gracefully will help keep board-specific fixups
out of the kernel and a lot of thrashing out of DT.

Once we have this general display configuration and hand-off right,
handing off a frame buffer address (or two or four) will likely
be easier on the kernel side.

I'm certainly over-simplifying, and the kernel and userspace will
need more information, especially for multiple displays:

	who's on /dev/fb0?
	what bit depth is wanted?
	which of multiple touch screens is associated with which panel?
	which backlight is connected to which panel?

Sorry for the long e-mail, but this gave me the a push to get
some thoughts in writing.

Regards,


Eric