[Intel-gfx] [PATCH 03/14] drm/i915/execlists: Flush the tasklet on parking

[Chris Wilson]

Quoting Tvrtko Ursulin (2019-05-02 15:24:16)
> 
> On 02/05/2019 15:21, Chris Wilson wrote:
> > Quoting Tvrtko Ursulin (2019-05-02 15:14:08)
> >>
> >> On 02/05/2019 14:53, Chris Wilson wrote:
> >>> Quoting Tvrtko Ursulin (2019-05-02 14:48:18)
> >>>>
> >>>> On 01/05/2019 12:45, Chris Wilson wrote:
> >>>>> Tidy up the cleanup sequence by always ensure that the tasklet is
> >>>>> flushed on parking (before we cleanup). The parking provides a
> >>>>> convenient point to ensure that the backend is truly idle.
> >>>>>
> >>>>> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> >>>>> ---
> >>>>>     drivers/gpu/drm/i915/gt/intel_lrc.c         | 7 ++++++-
> >>>>>     drivers/gpu/drm/i915/intel_guc_submission.c | 1 +
> >>>>>     2 files changed, 7 insertions(+), 1 deletion(-)
> >>>>>
> >>>>> diff --git a/drivers/gpu/drm/i915/gt/intel_lrc.c b/drivers/gpu/drm/i915/gt/intel_lrc.c
> >>>>> index 851e62ddcb87..7be54b868d8e 100644
> >>>>> --- a/drivers/gpu/drm/i915/gt/intel_lrc.c
> >>>>> +++ b/drivers/gpu/drm/i915/gt/intel_lrc.c
> >>>>> @@ -2331,6 +2331,11 @@ static int gen8_init_rcs_context(struct i915_request *rq)
> >>>>>         return i915_gem_render_state_emit(rq);
> >>>>>     }
> >>>>>     
> >>>>> +static void execlists_park(struct intel_engine_cs *engine)
> >>>>> +{
> >>>>> +     tasklet_kill(&engine->execlists.tasklet);
> >>>>
> >>>> Isn't it actually a problem if tasklet is scheduled and unstarted, or
> >>>> even in progress at the point of engine getting parked?
> >>>
> >>> That would be a broken driver. :|
> >>>
> >>> We must be quite sure that engine isn't going to send an interrupt as we
> >>> are just about to drop the wakeref we need to service that interrupt.
> >>>
> >>> tasklet_kill()
> >>> GEM_BUG_ON(engine->execlists.active);
> >>
> >> Or instead of both:
> >>
> >> /* Tasklet must not be running or scheduled at this point. */
> >> GEM_BUG_ON(engine->execlists.tasklet.state);
> > 
> > There's the dilemma that we start parking based on retirement not
> > final CS event.
> 
> But engine->park() is called once the last engine pm reference is 
> dropped. Are we dropping the last reference with a CS event pending?

Potentially we are.

i915_request_retire() -> context->exit() -> engine->park()

At no point along that chain do we actually check we have flushed the
backend. The tasklet_kill() would flush if the interrupt had already
been sent, but that's not very strict.

Oh well, you've talked me into to re-adding the wait loop here.
-Chris