[Intel-gfx] [igt-dev] [PATCH i-g-t 05/24] i915/gem_exec_schedule: Verify that using HW semaphores doesn't block

[Chris Wilson]

Quoting Tvrtko Ursulin (2019-03-26 09:19:33)
> 
> 
> On 22/03/2019 09:21, Chris Wilson wrote:
> > We may use HW semaphores to schedule nearly-ready work such that they
> > are already spinning on the GPU waiting for the completion on another
> > engine. However, we don't want for that spinning task to actually block
> > any real work should it be scheduled.
> > 
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > ---
> >   tests/i915/gem_exec_schedule.c | 87 ++++++++++++++++++++++++++++++++++
> >   1 file changed, 87 insertions(+)
> > 
> > diff --git a/tests/i915/gem_exec_schedule.c b/tests/i915/gem_exec_schedule.c
> > index 4f0577b4e..ae850c4a3 100644
> > --- a/tests/i915/gem_exec_schedule.c
> > +++ b/tests/i915/gem_exec_schedule.c
> > @@ -48,6 +48,10 @@
> >   
> >   #define MAX_CONTEXTS 1024
> >   
> > +#define LOCAL_I915_EXEC_BSD_SHIFT      (13)
> > +#define LOCAL_I915_EXEC_BSD_MASK       (3 << LOCAL_I915_EXEC_BSD_SHIFT)
> > +#define ENGINE_MASK  (I915_EXEC_RING_MASK | LOCAL_I915_EXEC_BSD_MASK)
> > +
> >   IGT_TEST_DESCRIPTION("Check that we can control the order of execution");
> >   
> >   static inline
> > @@ -320,6 +324,86 @@ static void smoketest(int fd, unsigned ring, unsigned timeout)
> >       }
> >   }
> >   
> > +static uint32_t __batch_create(int i915, uint32_t offset)
> > +{
> > +     const uint32_t bbe = MI_BATCH_BUFFER_END;
> > +     uint32_t handle;
> > +
> > +     handle = gem_create(i915, ALIGN(offset + 4, 4096));
> > +     gem_write(i915, handle, offset, &bbe, sizeof(bbe));
> > +
> > +     return handle;
> > +}
> > +
> > +static uint32_t batch_create(int i915)
> > +{
> > +     return __batch_create(i915, 0);
> > +}
> > +
> > +static void semaphore_userlock(int i915)
> > +{
> > +     struct drm_i915_gem_exec_object2 obj = {
> > +             .handle = batch_create(i915),
> > +     };
> > +     igt_spin_t *spin = NULL;
> > +     unsigned int engine;
> > +     uint32_t scratch;
> > +
> > +     igt_require(gem_scheduler_has_preemption(i915));
> > +
> > +     /*
> > +      * Given the use of semaphores to govern parallel submission
> > +      * of nearly-ready work to HW, we still want to run actually
> > +      * ready work immediately. Without semaphores, the dependent
> > +      * work wouldn't be submitted so our ready work will run.
> > +      */
> > +
> > +     scratch = gem_create(i915, 4096);
> > +     for_each_physical_engine(i915, engine) {
> > +             if (!spin) {
> > +                     spin = igt_spin_batch_new(i915,
> > +                                               .dependency = scratch,
> > +                                               .engine = engine);
> > +             } else {
> > +                     typeof(spin->execbuf.flags) saved = spin->execbuf.flags;
> 
> u64 reads better and struct eb won't change anyway.

If it were only u64!

> > +                     spin->execbuf.flags &= ~ENGINE_MASK;
> > +                     spin->execbuf.flags |= engine;
> > +
> > +                     gem_execbuf(i915, &spin->execbuf);
> 
> Do you need to wait for spinner to be running before submitting these 
> ones, to make sure the logic emits a semaphore poll for them and submits 
> them straight away?

Not required, the semaphores are emitted based on completion status.

> > +                     spin->execbuf.flags = saved;
> > +             }
> > +     }
> > +     igt_require(spin);
> > +     gem_close(i915, scratch);
> > +
> > +     /*
> > +      * On all dependent engines, the request may be executing (busywaiting
> > +      * on a HW semaphore) but it should not prevent any real work from
> > +      * taking precedence.
> > +      */
> > +     scratch = gem_context_create(i915);
> > +     for_each_physical_engine(i915, engine) {
> > +             struct drm_i915_gem_execbuffer2 execbuf = {
> > +                     .buffers_ptr = to_user_pointer(&obj),
> > +                     .buffer_count = 1,
> > +                     .flags = engine,
> > +                     .rsvd1 = scratch,
> > +             };
> > +
> > +             if (engine == (spin->execbuf.flags & ENGINE_MASK))
> > +                     continue;
> 
> Ugh saving and restoring eb flags to find the spinning engine here I 
> feel will be a land mine for the upcoming for_each_physical_engine 
> conversion but what can we do.

It will be fine. Unless you plan to randomise discovery just to make
things interesting. :)

We can make reuse of engines explicit if use ctx->engines[] instead of
for_each_physical_engine().

> > +             gem_execbuf(i915, &execbuf);
> > +     }
> > +     gem_context_destroy(i915, scratch);
> > +     gem_sync(i915, obj.handle); /* to hang unless we can preempt */
> 
> I got lost - how does this work if the spinner is still keeping the 
> obj.handle busy?

obj.handle and spinner are separate, and on different contexts. So we
fill all engines with the spinner + semaphores. Submit a new batch that
has no dependencies, and our expectation is that it is able to run ahead
of the semaphore busywait. Is that a reasonable expectation for
userspace? (Note that this demonstrates a subtle change in the ABI with
the introduction of plain semaphores, as without preemption that patch
causes this test to hang. So whether or not it is a reasonable
expectation, the change in behaviour is unwanted, but may have gone
unnoticed)
-Chris