[Intel-gfx] [PATCH] drm/i915/gt: Treat idling as a RPS downclock event

[Chris Wilson]

If we park/unpark faster than we can respond to RPS events, we never
will process a downlock event after expiring a waitboost, and thus we
will forever restart the GPU at max clocks even if the workload switches
and doesn't justify full power.

Closes: https://gitlab.freedesktop.org/drm/intel/issues/1500
Fixes: 3e7abf814193 ("drm/i915: Extract GT render power state management")
Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
Cc: Andi Shyti <andi.shyti at intel.com>
Cc: Lyude Paul <lyude at redhat.com>
---
 drivers/gpu/drm/i915/gt/intel_rps.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/drivers/gpu/drm/i915/gt/intel_rps.c b/drivers/gpu/drm/i915/gt/intel_rps.c
index 7bf631ca560b..50884aeac49c 100644
--- a/drivers/gpu/drm/i915/gt/intel_rps.c
+++ b/drivers/gpu/drm/i915/gt/intel_rps.c
@@ -771,6 +771,19 @@ void intel_rps_park(struct intel_rps *rps)
 	intel_uncore_forcewake_get(rps_to_uncore(rps), FORCEWAKE_MEDIA);
 	rps_set(rps, rps->idle_freq, false);
 	intel_uncore_forcewake_put(rps_to_uncore(rps), FORCEWAKE_MEDIA);
+
+	/*
+	 * Since we will try and restart from the previously requested
+	 * frequency on unparking, treat this idle point as a downlock
+	 * interrupt and reduce the frequency for resume. If we park/unpark
+	 * more frequently than the rps worker can run, we will not respond
+	 * to any EI and never see a change in frequency.
+	 *
+	 * (Note the we accommodate Cherryview's limitation of only using
+	 * an even bin by applying it to all.)
+	 */
+	rps->cur_freq =
+		max_t(u8, round_down(rps->cur_freq - 1, 2), rps->min_freq);
 }
 
 void intel_rps_boost(struct i915_request *rq)
-- 
2.20.1