Re: [Intel-gfx] GPU hang with kernel 4.10rc3

[Juergen Gross <jgross@suse.com>]

On 12/01/17 10:21, Chris Wilson wrote:
> On Thu, Jan 12, 2017 at 07:03:25AM +0100, Juergen Gross wrote:
>> On 11/01/17 18:08, Chris Wilson wrote:
>>> On Wed, Jan 11, 2017 at 05:33:34PM +0100, Juergen Gross wrote:
>>>> With kernel 4.10rc3 running as Xen dm0 I get at each boot:
>>>>
>>>> [   49.213697] [drm] GPU HANG: ecode 7:0:0x3d1d3d3d, in gnome-shell
>>>> [1431], reason: Hang on render ring, action: reset
>>>> [   49.213699] [drm] GPU hangs can indicate a bug anywhere in the entire
>>>> gfx stack, including userspace.
>>>> [   49.213700] [drm] Please file a _new_ bug report on
>>>> bugs.freedesktop.org against DRI -> DRM/Intel
>>>> [   49.213700] [drm] drm/i915 developers can then reassign to the right
>>>> component if it's not a kernel issue.
>>>> [   49.213700] [drm] The gpu crash dump is required to analyze gpu
>>>> hangs, so please always attach it.
>>>> [   49.213701] [drm] GPU crash dump saved to /sys/class/drm/card0/error
>>>> [   49.213755] drm/i915: Resetting chip after gpu hang
>>>> [   60.213769] drm/i915: Resetting chip after gpu hang
>>>> [   71.189737] drm/i915: Resetting chip after gpu hang
>>>> [   82.165747] drm/i915: Resetting chip after gpu hang
>>>> [   93.205727] drm/i915: Resetting chip after gpu hang
>>>>
>>>> The dump is attached.
>>>
>>> That's a nasty one. The first couple of pages of the batchbuffer appear
>>> to be overwritten. (Full of 0xc2c2c2c2, i.e. probably pixel data.) That
>>> may be a concurrent write by either the GPU or CPU, or we may have
>>> incorrected mapped a set of pages. That it doesn't recovered suggests
>>> that the corruption occurs frequently, probably on every request/batch.
>>
>> I hoped someone would have an idea already.
> 
> Sorry, first report of something like this in a long time (that I can
> remember at least). And the problem is that it can be anything from a
> coherency to a concurrency issue, so no one patch springs to mind.
> Thankfully it appears to be kernel related.
> -Chris
> 

Bisecting took longer than I thought, but I had to cherry pick some
patches and rebase one of them multiple times...

Finally I found the commit to blame: 920cf4194954ec ("drm/i915:
Introduce an internal allocator for disposable private objects")

In case you need me to produce some more data or test a patch
feel free to reach out.


Juergen