[Intel-gfx] [PATCH i-g-t] tests/kms_pipe_crc_basic: Reset GPU before running subtests

[Mika Kahola]

On Wed, 2017-06-07 at 11:24 +0100, Chris Wilson wrote:
> Quoting Mika Kahola (2017-06-07 09:00:10)
> > 
> > On Tue, 2017-06-06 at 14:23 +0100, Chris Wilson wrote:
> > > 
> > > Quoting Mika Kahola (2017-06-06 13:33:14)
> > > > 
> > > > 
> > > > On Tue, 2017-06-06 at 15:27 +0300, Ville Syrjälä wrote:
> > > > > 
> > > > > 
> > > > > On Tue, Jun 06, 2017 at 03:20:46PM +0300, Mika Kahola wrote:
> > > > > > 
> > > > > > 
> > > > > > 
> > > > > > It has been noticed by our CI BAT testing that in some 1%-
> > > > > > 3%
> > > > > > probability
> > > > > > kms_pipe_crc_basic subtest read-crc-pipe-x-frame-sequence
> > > > > > fails
> > > > > > when
> > > > > > comparing gathered CRC frames. However, running
> > > > > > kms_pipe_crc_basic
> > > > > > subtests alone i.e. outside BAT I was unable to replicate
> > > > > > the
> > > > > > issue.
> > > > > > 
> > > > > > The patch proposes a GPU reset before running the subtests.
> > > > > > This
> > > > > > way we
> > > > > > can ensure that GPU register settings are reinitialized if
> > > > > > they
> > > > > > have been
> > > > > > altered by the tests executed earlier in BAT.
> > > > > What does GPU reset have to do with display CRCs?
> > > > It's unclear but when you do it I was unable to replicate this
> > > > in
> > > > couple hundred BAT runs. Maybe the error probability just hits
> > > > below 1%
> > > > mark with this.
> > > So trace through what side-effect the reset may have and find the
> > > bug
> > > in
> > > the kernel.
> > I'm a bit puzzled with this bug. By running the test alone doesn't
> > trigger the issue but this occurs only when running fast-feedback.
> > testlist. Maybe tearing down the previous tests cause some sort of
> > instability which is then triggered when running this test? Any
> > hints
> > or pointers where to look at?
> Try forcing a mode change at the start (i.e. -> 1024x768 then back to
> preferred) and see if that clears up the residual failures. Either
> way
> (modeset or gpu reset), dump a snapshot of the registers at the start
> of
> a test and after a gpu hang and look at the differences.
> -Chris
Thanks Chris! I'll try those tricks.

-- 
Mika Kahola - Intel OTC