From: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
Userspace waits coming via the drm_syncobj route have so far been
bypassing the waitboost mechanism.
Use the previously added dma-fence wait tracking API and apply the
same waitboosting logic which applies to other entry points.
This should fix the perfomance regressions experience by clvk and
similar userspace which relies on drm_syncobj.
At the same time, but for documentation purposes only, use the new
dma-fence API from i915_request_wait too.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
---
drivers/gpu/drm/i915/i915_request.c | 13 ++++++++++---
1 file changed, 10 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c
index 7503dcb9043b..e24fac5c1567 100644
--- a/drivers/gpu/drm/i915/i915_request.c
+++ b/drivers/gpu/drm/i915/i915_request.c
@@ -94,7 +94,12 @@ static bool i915_fence_signaled(struct dma_fence *fence)
static bool i915_fence_enable_signaling(struct dma_fence *fence)
{
- return i915_request_enable_breadcrumb(to_request(fence));
+ struct i915_request *rq = to_request(fence);
+
+ if (dma_fence_wait_count(&rq->fence) && !i915_request_started(rq))
+ intel_rps_boost(rq);
+
+ return i915_request_enable_breadcrumb(rq);
}
static signed long i915_fence_wait(struct dma_fence *fence,
@@ -2037,11 +2042,13 @@ long i915_request_wait_timeout(struct i915_request *rq,
* but at a cost of spending more power processing the workload
* (bad for battery).
*/
- if (flags & I915_WAIT_PRIORITY && !i915_request_started(rq))
+ if (((flags & I915_WAIT_PRIORITY) || dma_fence_wait_count(&rq->fence))
+ && !i915_request_started(rq))
intel_rps_boost(rq);
wait.tsk = current;
- if (dma_fence_add_callback(&rq->fence, &wait.cb, request_wait_wake))
+ if (dma_fence_add_wait_callback(&rq->fence, &wait.cb,
+ request_wait_wake))
goto out;
/*
--
2.34.1
