[Intel-gfx] [PATCH 1/2] drm/i915/guc: Support engine busy stats

[Chris Wilson]

Quoting Tvrtko Ursulin (2018-02-22 08:26:58)
> 
> On 22/02/2018 07:51, Chris Wilson wrote:
> > Quoting Tvrtko Ursulin (2018-02-22 06:07:32)
> >> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> >>
> >> Place context in/out hooks into the GuC backend, when contexts are
> >> assigned to ports, and removed from them, in order to be able to
> >> provide engine busy stats in GuC mode.
> >>
> >> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> >> Testcase: igt/perf_pmu/busy-accuracy-*-*
> >> ---
> >>   drivers/gpu/drm/i915/intel_guc_submission.c | 13 ++++++++++---
> >>   1 file changed, 10 insertions(+), 3 deletions(-)
> >>
> >> diff --git a/drivers/gpu/drm/i915/intel_guc_submission.c b/drivers/gpu/drm/i915/intel_guc_submission.c
> >> index 649113c7a3c2..8e99f8fd6da2 100644
> >> --- a/drivers/gpu/drm/i915/intel_guc_submission.c
> >> +++ b/drivers/gpu/drm/i915/intel_guc_submission.c
> >> @@ -673,6 +673,7 @@ static void guc_dequeue(struct intel_engine_cs *engine)
> >>          struct intel_engine_execlists * const execlists = &engine->execlists;
> >>          struct execlist_port *port = execlists->port;
> >>          struct i915_request *last = NULL;
> >> +       struct i915_gem_context *last_ctx = NULL;
> >>          const struct execlist_port * const last_port =
> >>                  &execlists->port[execlists->port_mask];
> >>          bool submit = false;
> >> @@ -718,8 +719,13 @@ static void guc_dequeue(struct intel_engine_cs *engine)
> >>                                          goto done;
> >>                                  }
> >>   
> >> -                               if (submit)
> >> +                               if (submit) {
> >>                                          port_assign(port, last);
> >> +                                       if (last->ctx != last_ctx) {
> >> +                                               intel_engine_context_in(last->engine);
> >> +                                               last_ctx = last->ctx;
> >> +                                       }
> >> +                               }
> >>                                  port++;
> >>                          }
> >>   
> >> @@ -741,6 +747,8 @@ static void guc_dequeue(struct intel_engine_cs *engine)
> >>          execlists->first = rb;
> >>          if (submit) {
> >>                  port_assign(port, last);
> >> +               if (last->ctx != last_ctx)
> >> +                       intel_engine_context_in(last->engine);
> >>                  execlists_set_active(execlists, EXECLISTS_ACTIVE_USER);
> >>                  guc_submit(engine);
> >>          }
> >> @@ -763,6 +771,7 @@ static void guc_submission_tasklet(unsigned long data)
> >>   
> >>          rq = port_request(&port[0]);
> >>          while (rq && i915_request_completed(rq)) {
> >> +               intel_engine_context_out(rq->engine);
> > 
> > If we only emit context_in once for 2 consecutive rq with the same
> > context, we would need to do the same for context_out.
> > 
> > Will be interesting to see if this explodes, or we may need yet another
> > test :)
> 
> 
> I thought two consecutive requests with the same context is one port, so 
> one context_in and one context_out. But maybe I'm wrong. Lets wait and 
> see as you say. :)

The difference is we don't do lite-restore in the guc. I think you want
to move context_in to guc_submit().
-Chris