drm/i915/gvt: Optimize ring siwtch 2x faster by removing unnecessary POSTING_READ
authorChangbin Du <changbin.du@intel.com>
Fri, 23 Jun 2017 07:45:31 +0000 (15:45 +0800)
committerZhenyu Wang <zhenyuw@linux.intel.com>
Thu, 10 Aug 2017 02:26:05 +0000 (10:26 +0800)
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>
drivers/gpu/drm/i915/gvt/render.c

index 504e57c3bc23c824036dff1bf51cd1168f5b15f3..5a08bcd436a1de06e8798a563761929dec9573c5 100644 (file)
@@ -209,7 +209,6 @@ static void load_mocs(struct intel_vgpu *vgpu, int ring_id)
        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 *vgpu, int ring_id)
                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_vgpu *vgpu, int ring_id)
        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_vgpu *vgpu, int ring_id)
                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 intel_vgpu *vgpu, int ring_id)
        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 intel_vgpu *vgpu, int ring_id)
                        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 intel_vgpu *vgpu, int ring_id)
 {
        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 intel_vgpu *vgpu, int ring_id)
                        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);
 }
 
 /**