[Intel-gfx] [PATCH] drm/i915: Do a quick check on whether the fence is already signaled first

[Chris Wilson]

On Wed, Apr 26, 2017 at 11:43:03AM +0100, Tvrtko Ursulin wrote:
> 
> On 26/04/2017 11:15, Chris Wilson wrote:
> >Now that we try to signal the fence from inside the interrupt handler,
> >when we reach the signaler thread, the fence is most likely already
> >signaled. Skip manipulating the bottom-half locks if this is so.
> >
> >Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> >Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> >---
> > drivers/gpu/drm/i915/intel_breadcrumbs.c | 9 ++++++---
> > 1 file changed, 6 insertions(+), 3 deletions(-)
> >
> >diff --git a/drivers/gpu/drm/i915/intel_breadcrumbs.c b/drivers/gpu/drm/i915/intel_breadcrumbs.c
> >index 8f52fd5f6102..5f79c8135b3f 100644
> >--- a/drivers/gpu/drm/i915/intel_breadcrumbs.c
> >+++ b/drivers/gpu/drm/i915/intel_breadcrumbs.c
> >@@ -616,9 +616,12 @@ static int intel_breadcrumbs_signaler(void *arg)
> > 			request = i915_gem_request_get_rcu(request);
> > 		rcu_read_unlock();
> > 		if (signal_complete(request)) {
> >-			local_bh_disable();
> >-			dma_fence_signal(&request->fence);
> >-			local_bh_enable(); /* kick start the tasklets */
> >+			if (!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
> >+				      &request->fence.flags)) {
> >+				local_bh_disable();
> >+				dma_fence_signal(&request->fence);
> >+				local_bh_enable(); /* kick start the tasklet */
> >+			}
> >
> > 			spin_lock_irq(&b->rb_lock);
> >
> >
> 
> What are you referring to by "bottom-half locks" in the commit msg?
> AFAICS it would only skip kicking the tasklets with this change.
> That may be worth it, I haven't measured, but I don't see a
> difference wrt any locks.

It's just the preempt enable/disable pair plus the function call
 
> In fact we could change this to:
> 
> 	if (!dma_fence_signal(...)) {
> 		local_bh_disable();
> 		local_bh_enable();
> 	}
> 
> If we wanted to avoid touching the flags directly, but then would
> have a function call..

Yeah, plus that looks ridiculous ;)

The biggest cost in the signaler thread is going to sleep.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre