[Intel-gfx] [PATCH i-g-t] i915/gem_exec_balance: Check chain of submit dependencies

[Chris Wilson]

Quoting Tvrtko Ursulin (2019-11-27 15:08:25)
> 
> On 27/11/2019 14:50, Chris Wilson wrote:
> > Check that the submit fence couples into the dependency chain so that
> > the bonded pair cannot over take any of the earlier requests.
> > 
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> > ---
> >   tests/i915/gem_exec_balancer.c | 88 ++++++++++++++++++++++++++++++++++
> >   1 file changed, 88 insertions(+)
> > 
> > diff --git a/tests/i915/gem_exec_balancer.c b/tests/i915/gem_exec_balancer.c
> > index 86028cfdd..ff5bd0134 100644
> > --- a/tests/i915/gem_exec_balancer.c
> > +++ b/tests/i915/gem_exec_balancer.c
> > @@ -855,6 +855,91 @@ static void bonded_slice(int i915)
> >       munmap(stop, 4096);
> >   }
> >   
> > +static void bonded_chain(int i915)
> > +{
> > +     struct drm_i915_gem_exec_object2 batch = {
> > +             .handle = batch_create(i915),
> > +     };
> > +     uint32_t ctx;
> > +
> > +     /*
> > +      * Given batches A, B and B', where B and B' are a bonded pair, with
> > +      * B' depending on B with a submit fence and B depending on A as
> > +      * an ordinary fence; prove B' cannot complete before A.
> > +      */
> > +
> > +     ctx = gem_context_create(i915);
> > +
> > +     for (int class = 0; class < 32; class++) {
> > +             struct i915_engine_class_instance *siblings;
> > +             struct drm_i915_gem_execbuffer2 execbuf = {
> > +                     .buffers_ptr = to_user_pointer(&batch),
> > +                     .buffer_count = 1,
> > +                     .rsvd1 = ctx,
> > +             };
> > +             unsigned int count;
> > +             igt_spin_t *spin;
> > +
> > +             siblings = list_engines(i915, 1u << class, &count);
> > +             if (!siblings)
> > +                     continue;
> > +
> > +             if (count < 2) {
> > +                     free(siblings);
> > +                     continue;
> > +             }
> > +
> > +             /* A: spin forever on engine 1 */
> > +             set_load_balancer(i915, ctx, siblings, count, NULL);
> > +             spin = igt_spin_new(i915,
> > +                                 .ctx = ctx,
> > +                                 .engine = 1,
> > +                                 .flags = (IGT_SPIN_POLL_RUN |
> > +                                           IGT_SPIN_FENCE_OUT));
> > +             igt_spin_busywait_until_started(spin);
> > +
> > +             /* B: waits for A on engine 2 */
> > +             set_load_balancer(i915, ctx, siblings, count, NULL);
> > +             execbuf.rsvd2 = spin->out_fence;
> > +             execbuf.flags = I915_EXEC_FENCE_IN | I915_EXEC_FENCE_OUT;
> > +             execbuf.flags |= 2; /* opposite engine to spinner */
> 
> Why or?

execbuf.flags = 2 | FENCE_IN | FENCE_OUT

> 
> > +             gem_execbuf_wr(i915, &execbuf);
> > +
> > +             /* B': run in parallel with B on engine 1, i.e. not before A! */
> > +             set_load_balancer(i915, ctx, siblings, count, NULL);
> 
> Why are you repeating the same set_load_balancer calls?

Because we need a new timeline wrt the spinner.
(Well B' does. I originally tried using the virtual engine here, but
semaphores are not guaranteed so it works better by forcing the engine
selection to the known bad pattern.)

> > +             execbuf.flags = I915_EXEC_FENCE_SUBMIT | I915_EXEC_FENCE_OUT;
> > +             execbuf.flags |= 1; /* same engine as spinner */
> 
> Engine 3? But there are only two engines in the map, plus the virtual. 
> So 0, 1, 2 - how does this work for you?

execbuf.flags = 1 | FENCE_SUBMIT | FENCE_OUT;

> > +             execbuf.rsvd2 >>= 32;
> > +             gem_execbuf_wr(i915, &execbuf);
> > +
> > +             /* Wait for any magic timeslicing or preemptions... */
> > +             igt_assert_eq(sync_fence_wait(execbuf.rsvd2 >> 32, 1000),
> > +                           -ETIME);
> > +
> > +             igt_debugfs_dump(i915, "i915_engine_info");
> > +
> > +             /*
> > +              * ... which should not have happened, so everything is still
> > +              * waiting on the spinner
> > +              */
> > +             igt_assert_eq(sync_fence_status(spin->out_fence), 0);
> > +             igt_assert_eq(sync_fence_status(execbuf.rsvd2 & 0xffffffff), 0);
> > +             igt_assert_eq(sync_fence_status(execbuf.rsvd2 >> 32), 0);
> > +
> > +             igt_spin_free(i915, spin);
> > +             gem_sync(i915, batch.handle);
> > +
> > +             igt_assert_eq(sync_fence_status(execbuf.rsvd2 & 0xffffffff), 1);
> > +             igt_assert_eq(sync_fence_status(execbuf.rsvd2 >> 32), 1);
> > +
> > +             close(execbuf.rsvd2);
> > +             close(execbuf.rsvd2 >> 32);
> > +     }
> 
> How evil to lower ctx prio for A only?

A draw. You swap the timeslicing catch for the explicit preemption
check.
-Chris