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>

drivers/gpu/drm/i915/i915_debugfs.c

@@ -1357,7 +1357,7 @@ static int i915_ring_freq_table(struct seq_file *m, void *unused)
 	if (ret)
 		return ret;
 
-	seq_printf(m, "GPU freq (MHz)\tEffective CPU freq (MHz)\n");
+	seq_printf(m, "GPU freq (MHz)\tEffective CPU freq (MHz)\tEffective Ring freq (MHz)\n");
 
 	for (gpu_freq = dev_priv->rps.min_delay;
 	     gpu_freq <= dev_priv->rps.max_delay;
@@ -1366,7 +1366,10 @@ static int i915_ring_freq_table(struct seq_file *m, void *unused)
 		sandybridge_pcode_read(dev_priv,
 				       GEN6_PCODE_READ_MIN_FREQ_TABLE,
 				       &ia_freq);
-		seq_printf(m, "%d\t\t%d\n", gpu_freq * GT_FREQUENCY_MULTIPLIER, ia_freq * 100);
+		seq_printf(m, "%d\t\t%d\t\t\t\t%d\n",
+			   gpu_freq * GT_FREQUENCY_MULTIPLIER,
+			   ((ia_freq >> 0) & 0xff) * 100,
+			   ((ia_freq >> 8) & 0xff) * 100);
 	}
 
 	mutex_unlock(&dev_priv->rps.hw_lock);

drivers/gpu/drm/i915/i915_reg.h

@@ -1210,6 +1210,9 @@
 
 #define MCHBAR_MIRROR_BASE_SNB	0x140000
 
+/* Memory controller frequency in MCHBAR for Haswell (possible SNB+) */
+#define DCLK 0x5e04
+
 /** 915-945 and GM965 MCH register controlling DRAM channel access */
 #define DCC			0x10200
 #define DCC_ADDRESSING_MODE_SINGLE_CHANNEL		(0 << 0)
@@ -4390,6 +4393,7 @@
 #define   GEN6_DECODE_RC6_VID(vids)		(((vids) * 5) + 245)
 #define GEN6_PCODE_DATA				0x138128
 #define   GEN6_PCODE_FREQ_IA_RATIO_SHIFT	8
+#define   GEN6_PCODE_FREQ_RING_RATIO_SHIFT	16
 
 #define VLV_IOSF_DOORBELL_REQ			0x182100
 #define   IOSF_DEVFN_SHIFT			24

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;
+		unsigned int ia_freq = 0, ring_freq = 0;
 
-		/*
-		 * For GPU frequencies less than 750MHz, just use the lowest
-		 * ring freq.
+		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);
-		ia_freq <<= GEN6_PCODE_FREQ_IA_RATIO_SHIFT;
+		}
 
 		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);
 	}
 }