[Intel-gfx] [PATCH 08/14] drm/i915: Only reschedule the submission tasklet if preemption is possible
Chris Wilson
chris at chris-wilson.co.uk
Fri May 3 10:58:06 UTC 2019
Quoting Tvrtko Ursulin (2019-05-03 11:53:31)
>
> On 01/05/2019 12:45, Chris Wilson wrote:
> > If we couple the scheduler more tightly with the execlists policy, we
> > can apply the preemption policy to the question of whether we need to
> > kick the tasklet at all for this priority bump.
> >
> > v2: Rephrase it as a core i915 policy and not an execlists foible.
> >
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> > ---
> > drivers/gpu/drm/i915/gt/intel_engine.h | 18 ------------------
> > drivers/gpu/drm/i915/gt/intel_lrc.c | 4 ++--
> > drivers/gpu/drm/i915/gt/selftest_lrc.c | 7 ++++++-
> > drivers/gpu/drm/i915/i915_request.c | 2 --
> > drivers/gpu/drm/i915/i915_scheduler.c | 18 +++++++++++-------
> > drivers/gpu/drm/i915/i915_scheduler.h | 18 ++++++++++++++++++
> > drivers/gpu/drm/i915/intel_guc_submission.c | 3 ++-
> > 7 files changed, 39 insertions(+), 31 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/i915/gt/intel_engine.h b/drivers/gpu/drm/i915/gt/intel_engine.h
> > index f5b0f27cecb6..06d785533502 100644
> > --- a/drivers/gpu/drm/i915/gt/intel_engine.h
> > +++ b/drivers/gpu/drm/i915/gt/intel_engine.h
> > @@ -106,24 +106,6 @@ hangcheck_action_to_str(const enum intel_engine_hangcheck_action a)
> >
> > void intel_engines_set_scheduler_caps(struct drm_i915_private *i915);
> >
> > -static inline bool __execlists_need_preempt(int prio, int last)
> > -{
> > - /*
> > - * Allow preemption of low -> normal -> high, but we do
> > - * not allow low priority tasks to preempt other low priority
> > - * tasks under the impression that latency for low priority
> > - * tasks does not matter (as much as background throughput),
> > - * so kiss.
> > - *
> > - * More naturally we would write
> > - * prio >= max(0, last);
> > - * except that we wish to prevent triggering preemption at the same
> > - * priority level: the task that is running should remain running
> > - * to preserve FIFO ordering of dependencies.
> > - */
> > - return prio > max(I915_PRIORITY_NORMAL - 1, last);
> > -}
> > -
> > static inline void
> > execlists_set_active(struct intel_engine_execlists *execlists,
> > unsigned int bit)
> > diff --git a/drivers/gpu/drm/i915/gt/intel_lrc.c b/drivers/gpu/drm/i915/gt/intel_lrc.c
> > index 7be54b868d8e..35aae7b5c6b9 100644
> > --- a/drivers/gpu/drm/i915/gt/intel_lrc.c
> > +++ b/drivers/gpu/drm/i915/gt/intel_lrc.c
> > @@ -251,8 +251,8 @@ static inline bool need_preempt(const struct intel_engine_cs *engine,
> > * ourselves, ignore the request.
> > */
> > last_prio = effective_prio(rq);
> > - if (!__execlists_need_preempt(engine->execlists.queue_priority_hint,
> > - last_prio))
> > + if (!i915_scheduler_need_preempt(engine->execlists.queue_priority_hint,
> > + last_prio))
> > return false;
> >
> > /*
> > diff --git a/drivers/gpu/drm/i915/gt/selftest_lrc.c b/drivers/gpu/drm/i915/gt/selftest_lrc.c
> > index 84538f69185b..4b042893dc0e 100644
> > --- a/drivers/gpu/drm/i915/gt/selftest_lrc.c
> > +++ b/drivers/gpu/drm/i915/gt/selftest_lrc.c
> > @@ -638,14 +638,19 @@ static struct i915_request *dummy_request(struct intel_engine_cs *engine)
> > GEM_BUG_ON(i915_request_completed(rq));
> >
> > i915_sw_fence_init(&rq->submit, dummy_notify);
> > - i915_sw_fence_commit(&rq->submit);
> > + set_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags);
> >
> > return rq;
> > }
> >
> > static void dummy_request_free(struct i915_request *dummy)
> > {
> > + /* We have to fake the CS interrupt to kick the next request */
> > + i915_sw_fence_commit(&dummy->submit);
> > +
> > i915_request_mark_complete(dummy);
> > + dma_fence_signal(&dummy->fence);
> > +
> > i915_sched_node_fini(&dummy->sched);
> > i915_sw_fence_fini(&dummy->submit);
> >
> > diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c
> > index af8c9fa5e066..2e22da66a56c 100644
> > --- a/drivers/gpu/drm/i915/i915_request.c
> > +++ b/drivers/gpu/drm/i915/i915_request.c
> > @@ -1358,9 +1358,7 @@ long i915_request_wait(struct i915_request *rq,
> > if (flags & I915_WAIT_PRIORITY) {
> > if (!i915_request_started(rq) && INTEL_GEN(rq->i915) >= 6)
> > gen6_rps_boost(rq);
> > - local_bh_disable(); /* suspend tasklets for reprioritisation */
> > i915_schedule_bump_priority(rq, I915_PRIORITY_WAIT);
> > - local_bh_enable(); /* kick tasklets en masse */
> > }
> >
> > wait.tsk = current;
> > diff --git a/drivers/gpu/drm/i915/i915_scheduler.c b/drivers/gpu/drm/i915/i915_scheduler.c
> > index 39bc4f54e272..88d18600f5db 100644
> > --- a/drivers/gpu/drm/i915/i915_scheduler.c
> > +++ b/drivers/gpu/drm/i915/i915_scheduler.c
> > @@ -261,16 +261,20 @@ sched_lock_engine(const struct i915_sched_node *node,
> > return engine;
> > }
> >
> > -static bool inflight(const struct i915_request *rq,
> > - const struct intel_engine_cs *engine)
> > +static inline int rq_prio(const struct i915_request *rq)
> > {
> > - const struct i915_request *active;
> > + return rq->sched.attr.priority | __NO_PREEMPTION;
> > +}
> > +
> > +static bool kick_tasklet(const struct intel_engine_cs *engine, int prio)
> > +{
> > + const struct i915_request *inflight =
> > + port_request(engine->execlists.port);
> >
> > - if (!i915_request_is_active(rq))
> > + if (!inflight)
> > return false;
> >
> > - active = port_request(engine->execlists.port);
> > - return active->hw_context == rq->hw_context;
> > + return i915_scheduler_need_preempt(prio, rq_prio(inflight));
> > }
> >
> > static void __i915_schedule(struct i915_request *rq,
> > @@ -400,7 +404,7 @@ static void __i915_schedule(struct i915_request *rq,
> > * If we are already the currently executing context, don't
> > * bother evaluating if we should preempt ourselves.
> > */
> > - if (inflight(node_to_request(node), engine))
> > + if (!kick_tasklet(engine, prio))
> > continue;
>
> I don't see other callers for kick_tasklet in the series so I am
> thinking why not make the function called kick_tasklet actually kick the
> tasklet? ;)
You don't see the ghostly presence of set_preemption_timeout() here? Ok,
that could be pushed into the function as well.
> Could call it maybe_kick_tasklet and just end the loop with it.
>
> We could even abstract to the engine like engine->signal_preemption() or
> something, to hide the tasklet from the scheduler. But probably not
> worth it right now.
Haven't quite figured out the best approach. I've done both so far for
preemption-timeout, at the moment I've pulled the hrtimer into
i915_scheduler.c (frankly because it was easier to on the old rebasing);
but the dust is nowhere close to settled.
-Chris
More information about the Intel-gfx
mailing list