Re: Regressions in resume on intel graphics

[Dave Airlie <airlied@redhat.com>]

On Fri, 2010-11-19 at 22:35 +0100, Zdenek Kabelac wrote:
> 2010/11/19 Jesse Barnes <jbarnes@virtuousgeek.org>:
> > On Fri, 12 Nov 2010 10:17:07 +0100
> > Zdenek Kabelac <zdenek.kabelac@gmail.com> wrote:
> >
> >> 2010/11/12 Zdenek Kabelac <zdenek.kabelac@gmail.com>:
> >> > Hi
> >> >
> >> > I've noticed that after resume  with 2.6.36 kernel I need to switch
> >> > between console and back to Xorg to get usable Xsession again.
> >> > (my hw - T61, gma965, 4GB)
> >> >
> >> > I've played bisect game -  and this is the first broken kernel:
> >> >
> >> > ---
> >> > commit 8fd4bd22350784d5b2fe9274f6790ba353976415
> >> > Author: Jesse Barnes <jbarnes@virtuousgeek.org>
> >> > Date:   Wed Jun 23 12:56:12 2010 -0700
> >> >
> >> >    vt/console: try harder to print output when panicing
> >> > ---
> >>
> >> I've been able to boot and test with 2.6.37-rc1-00170-gf6614b7
> >> - and this problem seems to be fixed in this version (vt.c file seems
> >> to be gone?).
> >
> > This is really weird; there was one issue Dave tracked down related to
> > lockdep and the new oops code, but I think it's been fixed.  And it
> > should manifest as something other than a GPU hang...
> >
> 
> Well - unsure - how this worked ok - for 2.6.37-rc1-00170-gf6614b7 (I
> could try few more times - though this kernel seems to be crashing for
> various other reasons - so I've had quite few testruns before I've
> been able to test this in some very limited X environment - so maybe
> it was just some lucky case - but I still keep it on my disk)
> 
> Anyway - as an update - I'm now regularly using 2.6.37-rc2 - and it
> has exactly same problems - as in my original report -  so the problem
> has not magically disappeared and it is still there.  Is there
> anything I should try ?
> 
> Basically every time after resume from suspend I have to switch from X
> to console and back to get usable screen back. Without this I seen
> only black screen with mouse over it - is there some way for reverting
> of this patch - or maybe just disabling some part of it ?
> 
> 
> To have reliable resume I still have to keep drm_kms_help thread
> disable - otherwise I observe GPU errors.
> Switch between consoles is 'relatively' easy to overcome.
> 
> And another thing which might help -  during suspend/resume  I could
> usually see weird switch to console with some special text on the
> whole console screen -
> 
> [ ###.... ]
> [ ###.... ]
> 
> where  '###' is some changing number - and it seems to be different
> between resumes.
> Also I should add I'm using  'no-console-suspend'  kernel boot option.

Can you see if e0fdace10e75dac67d906213b780ff1b1a4cc360 reverts cleanly?
and fixes it?

The problem is I think you are getting a lockdep splat or rcu issue
before suspending, which sets oops_in_progress and never unsets it,
which means on resume the fb resume path kicks in to show you the oops
that is happening when there isn't actually anything to show. I've
gotten acks to have this reverted I just need to send the patch.

Dave