| Message ID | 20180507092527.7359-1-chris@chris-wilson.co.uk (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Chris Wilson <chris@chris-wilson.co.uk> writes: > When called from process context tasklet_schedule() defers itself to > ksoftirqd. From experience this may cause unacceptable latencies of over > 200ms in executing the submission tasklet, our goal is to reprioritise > the HW execution queue and trigger preemption immediately, so convert I would add 'execlists preemption' even tho the context points to it, as we are also quite intimately entangled here with kernel preemption too. > the rcu protection over to include a kick of the tasklet. > > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> > Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> > --- > drivers/gpu/drm/i915/i915_gem.c | 4 ++-- > 1 file changed, 2 insertions(+), 2 deletions(-) > > diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c > index 5ece6ae4bdff..03cd30001b5d 100644 > --- a/drivers/gpu/drm/i915/i915_gem.c > +++ b/drivers/gpu/drm/i915/i915_gem.c > @@ -578,10 +578,10 @@ static void __fence_set_priority(struct dma_fence *fence, > rq = to_request(fence); > engine = rq->engine; > > - rcu_read_lock(); > + local_bh_disable(); /* RCU serialisation for set-wedged protection */ > if (engine->schedule) > engine->schedule(rq, attr); > - rcu_read_unlock(); > + local_bh_enable(); /* kick the tasklets if queues were reprioritised */ Do we end up calling the tasklet->func() explicitly on the end of engine->schedule()? Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com> > } > > static void fence_set_priority(struct dma_fence *fence, > -- > 2.17.0
Quoting Mika Kuoppala (2018-05-07 12:15:58) > Chris Wilson <chris@chris-wilson.co.uk> writes: > > > When called from process context tasklet_schedule() defers itself to > > ksoftirqd. From experience this may cause unacceptable latencies of over > > 200ms in executing the submission tasklet, our goal is to reprioritise > > the HW execution queue and trigger preemption immediately, so convert > > I would add 'execlists preemption' even tho the context points to it, > as we are also quite intimately entangled here with kernel preemption > too. I'll settle for HW preemption. > > the rcu protection over to include a kick of the tasklet. > > > > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> > > Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> > > --- > > drivers/gpu/drm/i915/i915_gem.c | 4 ++-- > > 1 file changed, 2 insertions(+), 2 deletions(-) > > > > diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c > > index 5ece6ae4bdff..03cd30001b5d 100644 > > --- a/drivers/gpu/drm/i915/i915_gem.c > > +++ b/drivers/gpu/drm/i915/i915_gem.c > > @@ -578,10 +578,10 @@ static void __fence_set_priority(struct dma_fence *fence, > > rq = to_request(fence); > > engine = rq->engine; > > > > - rcu_read_lock(); > > + local_bh_disable(); /* RCU serialisation for set-wedged protection */ > > if (engine->schedule) > > engine->schedule(rq, attr); > > - rcu_read_unlock(); > > + local_bh_enable(); /* kick the tasklets if queues were reprioritised */ > > Do we end up calling the tasklet->func() explicitly on the end of > engine->schedule()? We do tasklet_hi_schedule(). Well in the future, we might either call ->func() directly or schedule. In the former case, no softirq is raised and local_bh_enable is just a slightly heavier preempt_disable(). -Chris
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c index 5ece6ae4bdff..03cd30001b5d 100644 --- a/drivers/gpu/drm/i915/i915_gem.c +++ b/drivers/gpu/drm/i915/i915_gem.c @@ -578,10 +578,10 @@ static void __fence_set_priority(struct dma_fence *fence, rq = to_request(fence); engine = rq->engine; - rcu_read_lock(); + local_bh_disable(); /* RCU serialisation for set-wedged protection */ if (engine->schedule) engine->schedule(rq, attr); - rcu_read_unlock(); + local_bh_enable(); /* kick the tasklets if queues were reprioritised */ } static void fence_set_priority(struct dma_fence *fence,
When called from process context tasklet_schedule() defers itself to ksoftirqd. From experience this may cause unacceptable latencies of over 200ms in executing the submission tasklet, our goal is to reprioritise the HW execution queue and trigger preemption immediately, so convert the rcu protection over to include a kick of the tasklet. Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> --- drivers/gpu/drm/i915/i915_gem.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)