Graphics driver on i.MX53 Android issues

[Bernhard Rosenkraenzer]

Booting Android on i.MX53 QSB using your bsp/freescale/linux-linaro-android.git kernel, with a display at a resolution with >= 1024 pixels width, results in endless loops of

[...]
<7>[ 120.044524] DMFC high resolution has set, will not change
<7>[ 120.144532] DMFC high resolution has set, will not change
<7>[ 120.224553] DMFC high resolution has set, will not change
[...]

This error message comes from drivers/mxc/ipu3/ipu_disp.c, _ipu_dmfc_init.

_ipu_dmfc_init is called from _ipu_dmfc_set_wait4eot, ipu_probe and ipu_resume.
_ipu_dmfc_set_wait4eot is called by ipu_init_channel_buffer (exported
through the IPU_INIT_CHANNEL_BUFFER ioctl).

There doesn't seem to be a reasonable code path that could hit those
10+ times a second, but I haven't put a lot of time into debugging this
yet.

In my current builds, I've speeded it up by hiding the warning and avoiding the logging activity. We can now see the right picture, but any screen update (e.g. moving the mouse cursor, or automatic updates when e.g. running 0xbench) causes the screen to blank for a while (and then come back up showing what it should be showing, behaves pretty much like an e-ink display). My current guess is that any screen update causes the driver to re-initialize, causing both the message and the blanking.

[Zach Pfeffer (pfefferz)]

This is a critical bug. What's going on with it?

[Fathi Boudra (fboudra)]

Here's the relevant thread on linaro-dev mailing list:

1. http://lists.linaro.org/pipermail/linaro-dev/2011-August/006962.html

There could be mis-use of the framebuffer's FBIOPUT_VSCREENINFO as a
way to update the screen instead of using FBIOPAN_DISPLAY. The former
ioctl could take significant amount of time re-calculating the timing
assuming resolution/color depth changes, and re-initializing the controller
for a new mode, and that could result in screen blinking. Please check
your Android code against this theory.

2. http://lists.linaro.org/pipermail/linaro-dev/2011-August/006965.html
> There could be mis-use of the framebuffer's FBIOPUT_VSCREENINFO as a
> way to update the screen instead of using FBIOPAN_DISPLAY.

That's exactly what Android does in a couple of places, e.g.
system/core/init/logo.c, function fb_update() [that one actually calls
FBIOPUT_VSCREENINFO _twice_ to update the screen]
hardware/libhardware/modules/gralloc/framebuffer.cpp, function fb_post

This should be fairly easy to fix, but googling it, I found an
explanation from an upstream engineer on why they're doing it that way
-- apparently on some graphics drivers, FBIOPAN_DISPLAY does smooth
panning, which obviously isn't what you want for a simple screen
refresh.

I'm definitely going to change it to see whether or not this is the
problem (and also to probably get a bit of a performance boost on
Panda) - but I wonder what hardware we're going to break with that.

Given other pieces of code (e.g. SDL) use FBIOPAN_DISPLAY for screen
refreshes, I think using that ioctl should be mostly safe.

> The former
> ioctl could take significant amount of time re-calculating the timing
> assuming resolution/color depth changes, and re-initializing the controller
> for a new mode, and that could result in screen blinking.

That is consistent with the messages I'm seeing in dmesg output.
Maybe (in addition to fixing userland) the driver should be a bit
smarter and not reinitialize the controller if it realizes the old and
new modes are actually the same, given Android is probably not the
only piece of code that uses FBIOPUT_VSCREENINFO, given apparently
some fb drivers interpret FBIOPAN_DISPLAY as something different...

3. http://lists.linaro.org/pipermail/linaro-dev/2011-August/006971.html

On 27 August 2011 21:45, Bernhard Rosenkranzer
<bernhard.rosenkranzer at linaro.org> wrote:
> That's exactly what Android does in a couple of places, e.g.
> system/core/init/logo.c, function fb_update() [that one actually calls
> FBIOPUT_VSCREENINFO _twice_ to update the screen]
> hardware/libhardware/modules/gralloc/framebuffer.cpp, function fb_post

I just changed it, and it works! We now have a really working display
on iMX53 Android. Thanks to everyone who helped!

[Bernhard Rosenkraenzer (berolinux)]

Fix committed on Android userland side.
Leaving the bug open on Freescale LT side because you may want to fix the driver anyway (probably Android isn't the only application out there that misuses FBIOPUT_VSCREENINFO).

[Ying-Chun Liu (paulliu)]

In today's meeting, we might want to keep this as is in FSL's kernel. If the bug is in the user space, we should fix all the bugs in user space. I'll mark this bug as wishlist in FSL landing team for tracking the issues.

[Zach Pfeffer (pfefferz)]

See https://android-build.linaro.org/builds/~linaro-android/stage-imx53-11.09-release/