[Intel-gfx] [PATCH 3/3] drm/i915/execlists: Suppress redundant preemption

[Chris Wilson]

Quoting Chris Wilson (2019-01-23 14:14:05)
> Quoting Chris Wilson (2019-01-23 13:47:29)
> > Quoting Chris Wilson (2019-01-23 12:36:02)
> > > On unwinding the active request we give it a small (limited to internal
> > > priority levels) boost to prevent it from being gazumped a second time.
> > > However, this means that it can be promoted to above the request that
> > > triggered the preemption request, causing a preempt-to-idle cycle for no
> > > change. We can avoid this if we take the boost into account when
> > > checking if the preemption request is valid.
> > > 
> > > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > > ---
> > >  drivers/gpu/drm/i915/intel_lrc.c | 29 +++++++++++++++++++++++++----
> > >  1 file changed, 25 insertions(+), 4 deletions(-)
> > > 
> > > diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
> > > index d9d744f6ab2c..74726f647e47 100644
> > > --- a/drivers/gpu/drm/i915/intel_lrc.c
> > > +++ b/drivers/gpu/drm/i915/intel_lrc.c
> > > @@ -163,6 +163,8 @@
> > >  #define WA_TAIL_DWORDS 2
> > >  #define WA_TAIL_BYTES (sizeof(u32) * WA_TAIL_DWORDS)
> > >  
> > > +#define ACTIVE_PRIORITY (I915_PRIORITY_NEWCLIENT)
> > > +
> > >  static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
> > >                                             struct intel_engine_cs *engine,
> > >                                             struct intel_context *ce);
> > > @@ -181,13 +183,31 @@ static inline int rq_prio(const struct i915_request *rq)
> > >         return rq->sched.attr.priority;
> > >  }
> > >  
> > > +static inline int active_prio(const struct i915_request *rq)
> > > +{
> > > +       int prio = rq_prio(rq);
> > > +
> > > +       /*
> > > +        * On unwinding the active request, we give it a priority bump
> > > +        * equivalent to a freshly submitted request. This protects it from
> > > +        * being gazumped again, but it would be preferrable if we didn't
> > > +        * let it be gazumped in the first place!
> > > +        *
> > > +        * See __unwind_incomplete_requests()
> > > +        */
> > > +       if (i915_request_started(rq))
> > > +               prio |= ACTIVE_PRIORITY;
> > 
> > Hmm, actually we are put to the tail of that priolist so we don't get
> > rerun ahead of the preemptee if we end up at the same priority.
> > ACTIVE_PRIORITY - 1 would seem to be the right compromise.
> 
> gem_sync/switch-default says ACTIVE_PRIORITY though. Hmm.

The answer is don't be lazy.

-       if (i915_request_started(rq))
+       if ((prio & ACTIVE_PRIORITY) != ACTIVE_PRIORITY &&
+           i915_request_started(rq)) {
                prio |= ACTIVE_PRIORITY;
+               prio--;
+       }

That doesn't break switch-default while providing a more accurate
estimate of prio after preemption.
-Chris