Re: drm i915 hangs on heavy io load

[Ben Widawsky <ben@bwidawsk.net>]

On Tue, 30 Oct 2012 10:01:38 +0900
Norbert Preining <preining@logic.at> wrote:

> Hi Dave,
> 
> On Di, 30 Okt 2012, Dave Airlie wrote:
> > > Thanks, running now with SNA. Let us see what happens.
> > 
> > Please don't, we ain't going to find the bug any quicker changing
> > variables, if the only thing that changed on your system was the
> 
> Sorry, didn't know. I supposed from the email of Chris that I should
> try it "to stress different code path" ... anyway, disabling it again.
> 
> > How long does it take you to reproduce, and does it happen when in
> 
> Very hard to say, most of the times it is in a few days scale.
> Though it happened also after a few hours once.
> 
> > actual use. On my laptop I've noticed I come back to it sometimes
> > and
> 
> Concerning actual use: I had instances on several occassions. Just
> 30min ago it was while working with shotwell on my photo collection,
> tagging photos. So there should not be a big disk activity or so, but
> a lot of screen redraws etc when going through the photos.
> On other times it was locked screen without screen saver.
> 
> Concerning coming back: For me it never worked. I always have to
> reboot to get a working state again. Ok, to be more specific. GNome3
> is dead. I can close the windows normally with kbd shortcuts and some
> mouse interaction, but no new windows, no moving etc.
> 
> > gnome-shell is dead. This never happened pre 3.7-rc's. But for me
> > its a 3-4 day window so far for it to die, which makes bisecting it
> > a bit
> 
> That sounds pretty much like my case, but since I often don't use
> the laptop for 2 days or so, it might be a bit longer.
> 
> > of a major problem. and I'm just finished bisecting the last
> > Ironlake regression that took over a month.
> 
> Ouch ...
> 
> > I would suggest starting a bisect on drivers/gpu/drm/i915 from 3.6
> > final to 3.7-rc1 or maybe -rc2.
> 
> Ok, thanks. I will try.
> 
> Best wishes
> 
> Norbert

Hi Norbert. In addition to the above, if this truly appears to be
related to i/o, can we try to decrease the time to failure with some
serious i/o tests? Off the top of my head I am not sure what's
available, but surely Google should be able to find something.

-- 
Ben Widawsky, Intel Open Source Technology Center