drm/i915: Scale ring, rather than ia, frequency on Haswell

Haswell introduces a separate frequency domain for the ring (uncore). So
where we used to increase the CPU (IA) clock with GPU busyness, we now
need to scale the ring frequency directly instead. As the ring limits
our memory bandwidth, it is vital for performance that when the GPU is
busy, we increase the frequency of the ring to increase the available
memory bandwidth.

v2: Fix the algorithm to actually use the scaled gpu frequency for the ring.
v3: s/max_ring_freq/min_ring_freq/ as that is what it is

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Jesse Barnes <jbarnes@virtuousgeek.org>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
[danvet: Add space checkpatch complained about.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

1 files changed, 29 insertions, 14 deletions

diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
index 4dc06a1bd43..5e1cdd54ba1 100644
--- a/drivers/gpu/drm/i915/intel_pm.c
+++ b/drivers/gpu/drm/i915/intel_pm.c
@@ -2684,8 +2684,8 @@ static void gen6_update_ring_freq(struct drm_device *dev)
 {
 	struct drm_i915_private *dev_priv = dev->dev_private;
 	int min_freq = 15;
-	int gpu_freq;
-	unsigned int ia_freq, max_ia_freq;
+	unsigned int gpu_freq;
+	unsigned int max_ia_freq, min_ring_freq;
 	int scaling_factor = 180;
 
 	WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
@@ -2701,6 +2701,10 @@ static void gen6_update_ring_freq(struct drm_device *dev)
 	/* Convert from kHz to MHz */
 	max_ia_freq /= 1000;
 
+	min_ring_freq = I915_READ(MCHBAR_MIRROR_BASE_SNB + DCLK);
+	/* convert DDR frequency from units of 133.3MHz to bandwidth */
+	min_ring_freq = (2 * 4 * min_ring_freq + 2) / 3;
+
 	/*
 	 * For each potential GPU frequency, load a ring frequency we'd like
 	 * to use for memory access.  We do this by specifying the IA frequency
@@ -2709,21 +2713,32 @@ static void gen6_update_ring_freq(struct drm_device *dev)
 	for (gpu_freq = dev_priv->rps.max_delay; gpu_freq >= dev_priv->rps.min_delay;
 	     gpu_freq--) {
 		int diff = dev_priv->rps.max_delay - gpu_freq;
-
-		/*
-		 * For GPU frequencies less than 750MHz, just use the lowest
-		 * ring freq.
-		 */
-		if (gpu_freq < min_freq)
-			ia_freq = 800;
-		else
-			ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
-		ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
-		ia_freq <<= GEN6_PCODE_FREQ_IA_RATIO_SHIFT;
+		unsigned int ia_freq = 0, ring_freq = 0;
+
+		if (IS_HASWELL(dev)) {
+			ring_freq = (gpu_freq * 5 + 3) / 4;
+			ring_freq = max(min_ring_freq, ring_freq);
+			/* leave ia_freq as the default, chosen by cpufreq */
+		} else {
+			/* On older processors, there is no separate ring
+			 * clock domain, so in order to boost the bandwidth
+			 * of the ring, we need to upclock the CPU (ia_freq).
+			 *
+			 * For GPU frequencies less than 750MHz,
+			 * just use the lowest ring freq.
+			 */
+			if (gpu_freq < min_freq)
+				ia_freq = 800;
+			else
+				ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
+			ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
+		}
 
 		sandybridge_pcode_write(dev_priv,
 					GEN6_PCODE_WRITE_MIN_FREQ_TABLE,
-					ia_freq | gpu_freq);
+					ia_freq << GEN6_PCODE_FREQ_IA_RATIO_SHIFT |
+					ring_freq << GEN6_PCODE_FREQ_RING_RATIO_SHIFT |
+					gpu_freq);
 	}
 }