drm/i915: Fix eDP link training when switching pipes on VLV/CHV

When switching from one pipe to another, the power sequencer of the new
pipe seems to need a bit of kicking to lock into the port. Even the vdd
force bit doesn't work before the power sequencer has been sufficiently
kicked, so this must be done before any AUX transactions are attempted.

After extensive experimentation I've determined that it's sufficient
to first write the port register with the correct values except the
port must remain disabled, then we can do a second write to enable the
port, after which the power sequencer is operational and allows the port
to start up properly.

Contrary to my earlier theories we don't need to enable the port with
the idle pattern, so let's just use training pattern 1 as that's what
other platforms use here.

v2: Refine the kick procedure
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drivers/gpu/drm/i915/intel_dp.c

@@ -2517,14 +2517,23 @@ static void intel_dp_enable_port(struct intel_dp *intel_dp)
 	struct drm_device *dev = intel_dp_to_dev(intel_dp);
 	struct drm_i915_private *dev_priv = dev->dev_private;
 
-	intel_dp->DP |= DP_PORT_EN;
-
 	/* enable with pattern 1 (as per spec) */
 	_intel_dp_set_link_train(intel_dp, &intel_dp->DP,
 				 DP_TRAINING_PATTERN_1);
 
 	I915_WRITE(intel_dp->output_reg, intel_dp->DP);
 	POSTING_READ(intel_dp->output_reg);
+
+	/*
+	 * Magic for VLV/CHV. We _must_ first set up the register
+	 * without actually enabling the port, and then do another
+	 * write to enable the port. Otherwise link training will
+	 * fail when the power sequencer is freshly used for this port.
+	 */
+	intel_dp->DP |= DP_PORT_EN;
+
+	I915_WRITE(intel_dp->output_reg, intel_dp->DP);
+	POSTING_READ(intel_dp->output_reg);
 }
 
 static void intel_enable_dp(struct intel_encoder *encoder)