From f846c8de64ced9965e04cc9ae1922036175e395b Mon Sep 17 00:00:00 2001
From: Changbin Du <changbin.du@intel.com>
Date: Fri, 23 Jun 2017 15:45:31 +0800
Subject: [PATCH] drm/i915/gvt: Optimize ring siwtch 2x faster by removing unnecessary POSTING_READ
Git-commit: f846c8de64ced9965e04cc9ae1922036175e395b
Patch-mainline: v4.14-rc1
References: FATE#322643 bsc#1055900

There are lots of POSTING_READ alongside each mmio write Op. While
actually this is not necessary. It just bring too much latency since
PCIe read Op is very slow which is of non-posted transaction.

For PCIe device, the mem transaction for strong ordering rules are:
  o PCIe mmio write sequence is FIFO. Posted request cannot
    pass previous posted request.
  o PCIe mmio read will not go ahead of previous write.

Intel graphics doesn't support RO, so we can apply above rules. In
our case, we only need one POSTING_READ at last. This can remove
half of mmio read Op and then the average ring switch performance
is nearly doubled.
         Before       After
cycles  ~970000      ~550000

Signed-off-by: Changbin Du <changbin.du@intel.com>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Acked-by: Takashi Iwai <tiwai@suse.de>

---
 drivers/gpu/drm/i915/gvt/render.c |   19 +++++++++++++------
 1 file changed, 13 insertions(+), 6 deletions(-)

--- a/drivers/gpu/drm/i915/gvt/render.c
+++ b/drivers/gpu/drm/i915/gvt/render.c
@@ -209,7 +209,6 @@ static void load_mocs(struct intel_vgpu
 	for (i = 0; i < 64; i++) {
 		gen9_render_mocs[ring_id][i] = I915_READ(offset);
 		I915_WRITE(offset, vgpu_vreg(vgpu, offset));
-		POSTING_READ(offset);
 		offset.reg += 4;
 	}
 
@@ -218,7 +217,6 @@ static void load_mocs(struct intel_vgpu
 		for (i = 0; i < 32; i++) {
 			gen9_render_mocs_L3[i] = I915_READ(l3_offset);
 			I915_WRITE(l3_offset, vgpu_vreg(vgpu, l3_offset));
-			POSTING_READ(l3_offset);
 			l3_offset.reg += 4;
 		}
 	}
@@ -244,7 +242,6 @@ static void restore_mocs(struct intel_vg
 	for (i = 0; i < 64; i++) {
 		vgpu_vreg(vgpu, offset) = I915_READ(offset);
 		I915_WRITE(offset, gen9_render_mocs[ring_id][i]);
-		POSTING_READ(offset);
 		offset.reg += 4;
 	}
 
@@ -253,7 +250,6 @@ static void restore_mocs(struct intel_vg
 		for (i = 0; i < 32; i++) {
 			vgpu_vreg(vgpu, l3_offset) = I915_READ(l3_offset);
 			I915_WRITE(l3_offset, gen9_render_mocs_L3[i]);
-			POSTING_READ(l3_offset);
 			l3_offset.reg += 4;
 		}
 	}
@@ -272,6 +268,7 @@ static void switch_mmio_to_vgpu(struct i
 	u32 ctx_ctrl = reg_state[CTX_CONTEXT_CONTROL_VAL];
 	u32 inhibit_mask =
 		_MASKED_BIT_ENABLE(CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT);
+	i915_reg_t last_reg = _MMIO(0);
 
 	if (IS_SKYLAKE(vgpu->gvt->dev_priv)
 		|| IS_KABYLAKE(vgpu->gvt->dev_priv)) {
@@ -305,12 +302,17 @@ static void switch_mmio_to_vgpu(struct i
 			v = vgpu_vreg(vgpu, mmio->reg);
 
 		I915_WRITE(mmio->reg, v);
-		POSTING_READ(mmio->reg);
+		last_reg = mmio->reg;
 
 		trace_render_mmio(vgpu->id, "load",
 				  i915_mmio_reg_offset(mmio->reg),
 				  mmio->value, v);
 	}
+
+	/* Make sure the swiched MMIOs has taken effect. */
+	if (likely(INTEL_GVT_MMIO_OFFSET(last_reg)))
+		POSTING_READ(last_reg);
+
 	handle_tlb_pending_event(vgpu, ring_id);
 }
 
@@ -319,6 +321,7 @@ static void switch_mmio_to_host(struct i
 {
 	struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
 	struct render_mmio *mmio;
+	i915_reg_t last_reg = _MMIO(0);
 	u32 v;
 	int i, array_size;
 
@@ -347,12 +350,16 @@ static void switch_mmio_to_host(struct i
 			continue;
 
 		I915_WRITE(mmio->reg, v);
-		POSTING_READ(mmio->reg);
+		last_reg = mmio->reg;
 
 		trace_render_mmio(vgpu->id, "restore",
 				  i915_mmio_reg_offset(mmio->reg),
 				  mmio->value, v);
 	}
+
+	/* Make sure the swiched MMIOs has taken effect. */
+	if (likely(INTEL_GVT_MMIO_OFFSET(last_reg)))
+		POSTING_READ(last_reg);
 }
 
 /**