[PATCH 2/2] drm/i915/gvt: Optimize ring siwtch 2x faster again by light weight mmio access wrapper

[changbin.du at intel.com]

From: Changbin Du <changbin.du at intel.com>

The I915_READ/WRITE is not only a mmio read/write, it also contains
debug checking and Forcewake domain lookup. This is too heavy for
GVT ring switch case which access batch of mmio registers on ring
switch. We can handle Forcewake manually and use the raw
i915_read/write instead. The benefit from this is 2x faster mmio
switch performance.
         Before       After
cycles  ~550000      ~250000

Signed-off-by: Changbin Du <changbin.du at intel.com>
---
 drivers/gpu/drm/i915/gvt/render.c | 42 +++++++++++++++++++++++++++------------
 1 file changed, 29 insertions(+), 13 deletions(-)

diff --git a/drivers/gpu/drm/i915/gvt/render.c b/drivers/gpu/drm/i915/gvt/render.c
index 5a08bcd..609b0f9 100644
--- a/drivers/gpu/drm/i915/gvt/render.c
+++ b/drivers/gpu/drm/i915/gvt/render.c
@@ -37,6 +37,9 @@
 #include "gvt.h"
 #include "trace.h"
 
+#define	I915_RAW_WRITE(reg, val)	__raw_i915_write32(dev_priv, reg, val)
+#define	I915_RAW_READ(reg)		__raw_i915_read32(dev_priv, reg)
+
 struct render_mmio {
 	int ring_id;
 	i915_reg_t reg;
@@ -207,7 +210,7 @@ static void load_mocs(struct intel_vgpu *vgpu, int ring_id)
 
 	offset.reg = regs[ring_id];
 	for (i = 0; i < 64; i++) {
-		gen9_render_mocs[ring_id][i] = I915_READ(offset);
+		gen9_render_mocs[ring_id][i] = I915_RAW_READ(offset);
 		I915_WRITE(offset, vgpu_vreg(vgpu, offset));
 		offset.reg += 4;
 	}
@@ -215,8 +218,8 @@ static void load_mocs(struct intel_vgpu *vgpu, int ring_id)
 	if (ring_id == RCS) {
 		l3_offset.reg = 0xb020;
 		for (i = 0; i < 32; i++) {
-			gen9_render_mocs_L3[i] = I915_READ(l3_offset);
-			I915_WRITE(l3_offset, vgpu_vreg(vgpu, l3_offset));
+			gen9_render_mocs_L3[i] = I915_RAW_READ(l3_offset);
+			I915_RAW_WRITE(l3_offset, vgpu_vreg(vgpu, l3_offset));
 			l3_offset.reg += 4;
 		}
 	}
@@ -240,16 +243,16 @@ static void restore_mocs(struct intel_vgpu *vgpu, int ring_id)
 
 	offset.reg = regs[ring_id];
 	for (i = 0; i < 64; i++) {
-		vgpu_vreg(vgpu, offset) = I915_READ(offset);
-		I915_WRITE(offset, gen9_render_mocs[ring_id][i]);
+		vgpu_vreg(vgpu, offset) = I915_RAW_READ(offset);
+		I915_RAW_WRITE(offset, gen9_render_mocs[ring_id][i]);
 		offset.reg += 4;
 	}
 
 	if (ring_id == RCS) {
 		l3_offset.reg = 0xb020;
 		for (i = 0; i < 32; i++) {
-			vgpu_vreg(vgpu, l3_offset) = I915_READ(l3_offset);
-			I915_WRITE(l3_offset, gen9_render_mocs_L3[i]);
+			vgpu_vreg(vgpu, l3_offset) = I915_RAW_READ(l3_offset);
+			I915_RAW_WRITE(l3_offset, gen9_render_mocs_L3[i]);
 			l3_offset.reg += 4;
 		}
 	}
@@ -284,7 +287,7 @@ static void switch_mmio_to_vgpu(struct intel_vgpu *vgpu, int ring_id)
 		if (mmio->ring_id != ring_id)
 			continue;
 
-		mmio->value = I915_READ(mmio->reg);
+		mmio->value = I915_RAW_READ(mmio->reg);
 
 		/*
 		 * if it is an inhibit context, load in_context mmio
@@ -301,7 +304,7 @@ static void switch_mmio_to_vgpu(struct intel_vgpu *vgpu, int ring_id)
 		else
 			v = vgpu_vreg(vgpu, mmio->reg);
 
-		I915_WRITE(mmio->reg, v);
+		I915_RAW_WRITE(mmio->reg, v);
 		last_reg = mmio->reg;
 
 		trace_render_mmio(vgpu->id, "load",
@@ -311,7 +314,7 @@ static void switch_mmio_to_vgpu(struct intel_vgpu *vgpu, int ring_id)
 
 	/* Make sure the swiched MMIOs has taken effect. */
 	if (likely(INTEL_GVT_MMIO_OFFSET(last_reg)))
-		POSTING_READ(last_reg);
+		I915_RAW_READ(last_reg);
 
 	handle_tlb_pending_event(vgpu, ring_id);
 }
@@ -338,7 +341,7 @@ static void switch_mmio_to_host(struct intel_vgpu *vgpu, int ring_id)
 		if (mmio->ring_id != ring_id)
 			continue;
 
-		vgpu_vreg(vgpu, mmio->reg) = I915_READ(mmio->reg);
+		vgpu_vreg(vgpu, mmio->reg) = I915_RAW_READ(mmio->reg);
 
 		if (mmio->mask) {
 			vgpu_vreg(vgpu, mmio->reg) &= ~(mmio->mask << 16);
@@ -349,7 +352,7 @@ static void switch_mmio_to_host(struct intel_vgpu *vgpu, int ring_id)
 		if (mmio->in_context)
 			continue;
 
-		I915_WRITE(mmio->reg, v);
+		I915_RAW_WRITE(mmio->reg, v);
 		last_reg = mmio->reg;
 
 		trace_render_mmio(vgpu->id, "restore",
@@ -359,7 +362,7 @@ static void switch_mmio_to_host(struct intel_vgpu *vgpu, int ring_id)
 
 	/* Make sure the swiched MMIOs has taken effect. */
 	if (likely(INTEL_GVT_MMIO_OFFSET(last_reg)))
-		POSTING_READ(last_reg);
+		I915_RAW_READ(last_reg);
 }
 
 /**
@@ -374,12 +377,23 @@ static void switch_mmio_to_host(struct intel_vgpu *vgpu, int ring_id)
 void intel_gvt_switch_mmio(struct intel_vgpu *pre,
 			   struct intel_vgpu *next, int ring_id)
 {
+	struct drm_i915_private *dev_priv;
+
 	if (WARN_ON(!pre && !next))
 		return;
 
 	gvt_dbg_render("switch ring %d from %s to %s\n", ring_id,
 		       pre ? "vGPU" : "host", next ? "vGPU" : "HOST");
 
+	dev_priv = pre ? pre->gvt->dev_priv : next->gvt->dev_priv;
+
+	/**
+	 * We are using raw mmio access wrapper to improve the
+	 * performace for batch mmio read/write, so we need
+	 * handle forcewake mannually.
+	 */
+	intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
+
 	/**
 	 * TODO: Optimize for vGPU to vGPU switch by merging
 	 * switch_mmio_to_host() and switch_mmio_to_vgpu().
@@ -389,4 +403,6 @@ void intel_gvt_switch_mmio(struct intel_vgpu *pre,
 
 	if (next)
 		switch_mmio_to_vgpu(next, ring_id);
+
+	intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
 }
-- 
2.7.4