[Intel-gfx] [PATCH] drm/i915/gt: Flush submission tasklet before waiting/retiring

[Chris Wilson]

Quoting Summers, Stuart (2019-10-08 15:52:15)
> On Tue, 2019-10-08 at 11:56 +0100, Chris Wilson wrote:
> > A common bane of ours is arbitrary delays in ksoftirqd processing our
> > submission tasklet. Give the submission tasklet a kick before we wait
> > to
> > avoid those delays eating into a tight timeout.
> > 
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > ---
> >  drivers/gpu/drm/i915/gt/intel_engine.h      |  3 +-
> >  drivers/gpu/drm/i915/gt/intel_engine_cs.c   | 33 +++++++++++++----
> > ----
> >  drivers/gpu/drm/i915/gt/intel_gt_requests.c | 12 ++++++++
> >  3 files changed, 34 insertions(+), 14 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/i915/gt/intel_engine.h
> > b/drivers/gpu/drm/i915/gt/intel_engine.h
> > index c9e8c8ccbd47..d624752f2a92 100644
> > --- a/drivers/gpu/drm/i915/gt/intel_engine.h
> > +++ b/drivers/gpu/drm/i915/gt/intel_engine.h
> > @@ -407,8 +407,9 @@ static inline void __intel_engine_reset(struct
> > intel_engine_cs *engine,
> >       engine->serial++; /* contexts lost */
> >  }
> >  
> > -bool intel_engine_is_idle(struct intel_engine_cs *engine);
> >  bool intel_engines_are_idle(struct intel_gt *gt);
> > +bool intel_engine_is_idle(struct intel_engine_cs *engine);
> > +void intel_engine_flush_submission(struct intel_engine_cs *engine);
> >  
> >  void intel_engines_reset_default_submission(struct intel_gt *gt);
> >  
> > diff --git a/drivers/gpu/drm/i915/gt/intel_engine_cs.c
> > b/drivers/gpu/drm/i915/gt/intel_engine_cs.c
> > index 6220b7151bb9..7e2aa7a6bef0 100644
> > --- a/drivers/gpu/drm/i915/gt/intel_engine_cs.c
> > +++ b/drivers/gpu/drm/i915/gt/intel_engine_cs.c
> > @@ -1040,6 +1040,25 @@ static bool ring_is_idle(struct
> > intel_engine_cs *engine)
> >       return idle;
> >  }
> >  
> > +void intel_engine_flush_submission(struct intel_engine_cs *engine)
> > +{
> > +     struct tasklet_struct *t = &engine->execlists.tasklet;
> > +
> > +     if (__tasklet_is_scheduled(t)) {
> > +             local_bh_disable();
> > +             if (tasklet_trylock(t)) {
> > +                     /* Must wait for any GPU reset in progress. */
> > +                     if (__tasklet_is_enabled(t))
> > +                             t->func(t->data);
> > +                     tasklet_unlock(t);
> > +             }
> > +             local_bh_enable();
> > +     }
> > +
> > +     /* Otherwise flush the tasklet if it was running on another cpu
> > */
> > +     tasklet_unlock_wait(t);
> > +}
> > +
> >  /**
> >   * intel_engine_is_idle() - Report if the engine has finished
> > process all work
> >   * @engine: the intel_engine_cs
> > @@ -1058,21 +1077,9 @@ bool intel_engine_is_idle(struct
> > intel_engine_cs *engine)
> >  
> >       /* Waiting to drain ELSP? */
> >       if (execlists_active(&engine->execlists)) {
> > -             struct tasklet_struct *t = &engine->execlists.tasklet;
> > -
> >               synchronize_hardirq(engine->i915->drm.pdev->irq);
> >  
> > -             local_bh_disable();
> > -             if (tasklet_trylock(t)) {
> > -                     /* Must wait for any GPU reset in progress. */
> > -                     if (__tasklet_is_enabled(t))
> > -                             t->func(t->data);
> > -                     tasklet_unlock(t);
> > -             }
> > -             local_bh_enable();
> > -
> > -             /* Otherwise flush the tasklet if it was on another cpu
> > */
> > -             tasklet_unlock_wait(t);
> > +             intel_engine_flush_submission(engine);
> >  
> >               if (execlists_active(&engine->execlists))
> >                       return false;
> > diff --git a/drivers/gpu/drm/i915/gt/intel_gt_requests.c
> > b/drivers/gpu/drm/i915/gt/intel_gt_requests.c
> > index ca606b79fd5e..cbb4069b11e1 100644
> > --- a/drivers/gpu/drm/i915/gt/intel_gt_requests.c
> > +++ b/drivers/gpu/drm/i915/gt/intel_gt_requests.c
> > @@ -4,6 +4,7 @@
> >   * Copyright © 2019 Intel Corporation
> >   */
> >  
> > +#include "i915_drv.h" /* for_each_engine() */
> >  #include "i915_request.h"
> >  #include "intel_gt.h"
> >  #include "intel_gt_pm.h"
> > @@ -19,6 +20,15 @@ static void retire_requests(struct intel_timeline
> > *tl)
> >                       break;
> >  }
> >  
> > +static void flush_submission(struct intel_gt *gt)
> > +{
> > +     struct intel_engine_cs *engine;
> > +     enum intel_engine_id id;
> > +
> > +     for_each_engine(engine, gt->i915, id)
> > +             intel_engine_flush_submission(engine);
> > +}
> > +
> >  long intel_gt_retire_requests_timeout(struct intel_gt *gt, long
> > timeout)
> >  {
> >       struct intel_gt_timelines *timelines = &gt->timelines;
> > @@ -32,6 +42,8 @@ long intel_gt_retire_requests_timeout(struct
> > intel_gt *gt, long timeout)
> >       if (unlikely(timeout < 0))
> >               timeout = -timeout, interruptible = false;
> >  
> > +     flush_submission(gt); /* kick the ksoftirqd tasklets */
> 
> Won't this add a performance hit if we are doing this across all
> engines? Is there a way we can isolate this a bit more?

One of my earlier thoughts was to use request->engine just before the
wait, but that depends on actually hitting the wait and so is a bit
weaker than always flushing.
-Chris