[igt-dev] [PATCH i-g-t] igt/gem_eio: Preserve batch between reset-stress iterations

[Chris Wilson]

Quoting Tvrtko Ursulin (2018-08-08 13:38:53)
> 
> On 08/08/2018 12:31, Chris Wilson wrote:
> > We can keep the original batch around and avoid recreating it between
> > reset iterations to focus on the impact of resets.
> > 
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > Cc: Tvrtko Ursulin <tvrtko.ursulin at linux.intel.com>
> > ---
> >   tests/gem_eio.c | 30 +++++++++++++++++-------------
> >   1 file changed, 17 insertions(+), 13 deletions(-)
> > 
> > diff --git a/tests/gem_eio.c b/tests/gem_eio.c
> > index de161332d..5250a414c 100644
> > --- a/tests/gem_eio.c
> > +++ b/tests/gem_eio.c
> > @@ -650,35 +650,38 @@ static void reset_stress(int fd,
> >                        uint32_t ctx0, unsigned int engine,
> >                        unsigned int flags)
> >   {
> > +     const uint32_t bbe = MI_BATCH_BUFFER_END;
> > +     struct drm_i915_gem_exec_object2 obj = {
> > +             .handle = gem_create(fd, 4096)
> > +     };
> > +     struct drm_i915_gem_execbuffer2 execbuf = {
> > +             .buffers_ptr = to_user_pointer(&obj),
> > +             .buffer_count = 1,
> > +             .flags = engine,
> > +     };
> > +     gem_write(fd, obj.handle, 0, &bbe, sizeof(bbe));
> > +
> >       igt_until_timeout(5) {
> > -             struct drm_i915_gem_execbuffer2 execbuf = { };
> > -             struct drm_i915_gem_exec_object2 obj = { };
> > -             uint32_t bbe = MI_BATCH_BUFFER_END;
> > +             uint32_t ctx = context_create_safe(fd);
> 
> There is still this per loop...

I thought that was intentional :)

It felt like the spirit of the test to try and mix up the contexts as
much as possible; one constant, one fresh.

> >               igt_spin_t *hang;
> >               unsigned int i;
> > -             uint32_t ctx;
> >   
> >               gem_quiescent_gpu(fd);
> >   
> >               igt_require(i915_reset_control(flags & TEST_WEDGE ?
> >                                              false : true));
> >   
> > -             ctx = context_create_safe(fd);
> > -
> >               /*
> >                * Start executing a spin batch with some queued batches
> >                * against a different context after it.
> >                */
> >               hang = spin_sync(fd, ctx0, engine);
> 
> ... and a ton of operations in this one, so I wonder why bother with one 
> batch?

Because I don't have spin_sync() in my pattern recognition matrix yet.
One excuse is that it doesn't have any create verb in its name, so easy
to forget its hidden costs.
-Chris