Re: Regression: drm/radeon: brightness control hard system lockup

[Eldad Zack]

On Mon, 7 Jan 2013, Alex Deucher wrote:
> On Mon, Jan 7, 2013 at 4:33 PM, Eldad Zack <eldad@xxxxxxxxxxxxxxx> wrote:
> >
> > On Mon, 7 Jan 2013, Alex Deucher wrote:
> >> On Sun, Jan 6, 2013 at 7:59 AM, Eldad Zack <eldad@xxxxxxxxxxxxxxx> wrote:
> >> >
> >> > Hi Alex,
> >> >
> >> > Commit 0ecebb9e0d14e9948e0b1529883a776758117d6f "drm/radeon: switch to a
> >> > finer grained reset for evergreen" introduced a hard system lockup to my
> >> > setup. I found it after bisecting, and confirmed it by reverting it on
> >> > the latest mainline ( 5f243b9 ).
> >> >
> >> > This:
> >> >
> >> >    echo 7 > /sys/devices/pci0000:00/0000:00:01.0/0000:01:00.0/backlight/acpi_video0/brightness
> >> >
> >> > Causes a hard lock-up hard, i.e. immediate freeze, without any logs.
> >> >
> >> > See lspci output and kernel .config below.
> >> > If there's any more info I can provide, please let me know.
> >>
> >> Do you normally see GPU resets when changing the backlight?  Please
> >> attach your dmesg output when changing the backlight with the patch
> >> reverted.
> >
> > I see nothing. Just to make sure, I cleared the buffer, cycled through
> > 0-7 a couple of hunderd times (until the flicker annoyed), but I see no
> > messages at all.
> > Is there any debug config I should turn on?
> 
> Can you try adding a printk() in evergreen_asic_reset() and see if it
> is somehow getting called when you change the brightness?  When you
> use the apci backlight control, the radeon driver is not involved at
> all.  They only way the driver would get involved is if the acpi
> backlight control somehow caused the GPU to hang and then the driver
> detected the hang and attempted to reset the GPU.  I don't see any
> evidence of a GPU reset in your kernel log however.  Note that the
> driver also registers native backlight contol.  Does that work any
> better than acpi?

The native backlight controls work very well. Thanks for that, I didn't
even noticed that. It has finer control over brightness too.

I worked out a fix for the problem, but I think it's not a proper one.
What I noticed is that evergreen_gpu_soft_reset() is only ever called
once on my system, at boot.
Then I realized from the dmesg that neither 
evergreen_gpu_soft_reset_dma() nor evergreen_gpu_soft_reset_gfx() actually
do anything. Both return on the first if statements there.

So as far as I can tell, the difference your patch introduced is calling 
evergreen_mc_stop() and evergreen_mc_resume(), which somehow puts my 
system in some state that ACPI brightness control leads to a lock up.

BTW, I don't see a GPU reset happening at all - I also tried suspend/resume
and starting an OpenGL application. Do you know how I can trigger it?

To fix this, I moved the GUI_ACTIVE test before evergreen_mc_stop(), but
I think it just masks the issue.
Patch below (against latest HEAD) just in case.

Cheers,
Eldad