[Intel-gfx] [PATCH v2] tests/gem_error_capture: Initial testcase for error state capture/dump
Mateo Lozano, Oscar
oscar.mateo at intel.com
Fri May 9 14:07:04 CEST 2014
On Tue, 2014-04-15 at 21:38 +0200, Daniel Vetter wrote:
> On Mon, Apr 14, 2014 at 01:03:58PM +0000, Mateo Lozano, Oscar wrote:
> > > I would add a little more smarts to both the kernel and error-decode.
> > > In the kernel, we can print the guilty request, which you can then use to
> > > confirm that it is yours. That seems to me to be a stronger validation of
> > > gem_error_capture, and a useful bit of information from hangstats that we do
> > > not expose currently.
> >
> > That sounds good. I have to add a number of other things to
> > i915_gpu_error as part of the Execlists code, so I´ll add a "--- guilty
> > request" as well and resubmit this test together with the series.
>
> If we want this much smarts then we need a properly hanging batch, e.g.
> like the looping batch used in gem_reset_stats.
>
> The problem with that is that this will kill the gpu if reset doesn't work
> (i.e. gen2/3) so we need to skip this test there. Or maybe split things
> into 2 subtests and use the properly hanging batch only when we do the
> extended guilty testing under discussion here.
>
> But in any case just checking that the batch is somewhere in the ring
> (properly masking of lower bits 0-11 ofc) and checking whether the batch
> is correctl dumped (with the magic value) would catch a lot of the
> past&present execbuf bugs - we've had issues with dumping fancy values of
> 0 a lot.
>
> For the guilty stuff we have an extensive set of tests in gem_reset_stat
> using the reset stat ioctl already. And for the occasional "the hang
> detection logic is busted bug" I think nothing short of a human brain
> locking at the batch really helps. At least if we want to be somewhat
> platform agnostic ...
>
> So imo the current level of checking loosk Good Enough. But I'm certainly
> not going to stop you ;-)
Ok, for the moment I'm happy that I can unblock Execlists. I don't mind
looking at the "--- guilty requests" in the future, but first I need to
get Execlists out of the way...
Cheers,
Oscar
More information about the Intel-gfx
mailing list