[Intel-gfx] [PATCH 1/2] drm/i915/selftests: Calculate maximum ring size for preemption chain

[Chris Wilson]

Quoting Caz Yokoyama (2019-03-21 18:41:10)
> inline
> -caz
> On Thu, 2019-03-21 at 07:37 +0000, Chris Wilson wrote:
> > 32 is too many for the likes of kbl, and in order to insert that many
> Not only kbl. ring_size is 25 on my cfl.
> 
> > requests into the ring requires us to declare the first few hung --
> The hung is not caused by 32. It is caused by accumulation of requests
> for all prime numbers.

Sure, but the design of the test is that we don't care for more than
ring_size.
> 
> > understandably a slow and unexpected process. Instead, measure the
> > size
> > of a singe requests and use that to estimate the upper bound on the
> > chain length we can use for our test, remembering to flush the
> > previous
> > chain between tests for safety.
> > 
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> > ---
> >  drivers/gpu/drm/i915/selftests/intel_lrc.c | 40
> > ++++++++++++++++++++--
> >  1 file changed, 37 insertions(+), 3 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/i915/selftests/intel_lrc.c
> > b/drivers/gpu/drm/i915/selftests/intel_lrc.c
> > index d61520ea03c1..42068ed5eec0 100644
> > --- a/drivers/gpu/drm/i915/selftests/intel_lrc.c
> > +++ b/drivers/gpu/drm/i915/selftests/intel_lrc.c
> > @@ -615,14 +615,33 @@ static int live_chain_preempt(void *arg)
> >               struct i915_sched_attr attr = {
> >                       .priority =
> > I915_USER_PRIORITY(I915_PRIORITY_MAX),
> >               };
> > -             int count, i;
> > +             struct i915_request *rq;
> > +             int ring_size, count, i;
> >  
> >               if (!intel_engine_has_preemption(engine))
> >                       continue;
> >  
> > -             for_each_prime_number_from(count, 1, 32) { /* must fit
> > ring! */
> > -                     struct i915_request *rq;
> > +             rq = igt_spinner_create_request(&lo.spin,
> > +                                             lo.ctx, engine,
> > +                                             MI_ARB_CHECK);
> > +             if (IS_ERR(rq))
> > +                     goto err_wedged;
> > +             i915_request_add(rq);
> > +
> > +             ring_size = rq->wa_tail - rq->head;
> > +             if (ring_size < 0)
> > +                     ring_size += rq->ring->size;
> > +             ring_size = rq->ring->size / ring_size;
> > +             pr_debug("%s(%s): Using maximum of %d requests\n",
> > +                      __func__, engine->name, ring_size);
> >  
> > +             igt_spinner_end(&lo.spin);
> > +             if (i915_request_wait(rq, I915_WAIT_LOCKED, HZ / 2) <
> > 0) {
> > +                     pr_err("Timed out waiting to flush %s\n",
> > engine->name);
> > +                     goto err_wedged;
> > +             }
> > +
> > +             for_each_prime_number_from(count, 1, ring_size) {
> >                       rq = igt_spinner_create_request(&hi.spin,
> >                                                       hi.ctx, engine,
> >                                                       MI_ARB_CHECK);
> > @@ -664,6 +683,21 @@ static int live_chain_preempt(void *arg)
> >                               goto err_wedged;
> >                       }
> >                       igt_spinner_end(&lo.spin);
> > +
> > +                     rq = i915_request_alloc(engine, lo.ctx);
> > +                     if (IS_ERR(rq))
> > +                             goto err_wedged;
> > +                     i915_request_add(rq);
> This request add is redundant. Wait for the last rq for lo.

Or not. And use a simple request as a flush and test.
-Chris