[Intel-gfx] [PATCH i-g-t] igt/gem_ctx_exec: Exercise I915_CONTEXT_PARAM_RECOVERABLE

[Chris Wilson]

Quoting Mika Kuoppala (2018-10-04 14:26:16)
> Chris Wilson <chris at chris-wilson.co.uk> writes:
> 
> > When RECOVERABLE is set, the kernel will attempt to automatically recover
> > a context after a hang. But if it is unset, the kernel will ban the
> > guilty context on a hang, preventing subsequent execution.
> >
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > ---
> >  tests/gem_ctx_exec.c | 38 ++++++++++++++++++++++++++++++++++++++
> >  1 file changed, 38 insertions(+)
> >
> > diff --git a/tests/gem_ctx_exec.c b/tests/gem_ctx_exec.c
> > index 908b59af..1cd3b20e 100644
> > --- a/tests/gem_ctx_exec.c
> > +++ b/tests/gem_ctx_exec.c
> > @@ -147,6 +147,41 @@ static void invalid_context(int fd, unsigned ring, uint32_t handle)
> >       igt_assert_eq(__gem_execbuf(fd, &execbuf), -ENOENT);
> >  }
> >  
> > +static void norecovery(int i915)
> > +{
> > +     for (int pass = 1; pass >= 0; pass--) {
> > +             struct drm_i915_gem_context_param param = {
> > +                     .ctx_id = gem_context_create(i915),
> > +                     .param = 0x7,
> > +                     .value = pass,
> > +             };
> > +             int expect = pass == 0 ? -EIO : 0;
> > +             igt_spin_t *spin;
> > +
> > +             if (pass)
> > +                     igt_require(__gem_context_set_param(i915, &param) == 0);
> > +             else
> > +                     gem_context_set_param(i915, &param);
> 
> You are trying in here to make the asserts unique?

Nah, since pass=1 was meant to be standard procedure, this was just a
"has param check". So I was being lazy and not separating it out into
its own little helper.

> > +             gem_context_get_param(i915, &param);
> > +             igt_assert_eq(param.value, pass);
> > +
> > +             spin = __igt_spin_batch_new(i915,
> > +                                         .ctx = param.ctx_id,
> > +                                         .flags = IGT_SPIN_POLL_RUN);
> > +             igt_assert(spin->running);
> > +
> > +             while (!READ_ONCE(*spin->running))
> > +                     ;
> 
> wait that it is running...
> 
> > +             igt_force_gpu_reset(i915);
> > +
> 
> then reset.

Aye, that's our current MO for gpu resets.

> > +             igt_spin_batch_end(spin);
> 
> > +             igt_assert_eq(__gem_execbuf(i915, &spin->execbuf), expect);
> 
> Hmm where do we sync with hang resolution. Is it the handoff in
> the wedged debugfs entry?

Yes. The force-gpu reset is synchronous, you should be able to remember
the fun we had with that little race.
-Chris