i965: Add batch_flush_notify hook to create new vertex-buffer bo

This avoids mapping a buffer object which is being referenced
by a batch that has already been flushed, (which is a terribly
expensive operation).

On my machine this brings the performance of x11perf -aa10text
from 85k back to 150k, (where it was before a recent kernel
upgrade). Also, before this patch, when I used my X server
actively performance would drop as low as 15k---hopefully that
bug is gone with this change.

src/i830.h

@@ -538,6 +538,8 @@ typedef struct _I830Rec {
 #endif
    CloseScreenProcPtr CloseScreen;
 
+   void (*batch_flush_notify)(ScrnInfoPtr pScrn);
+
 #ifdef I830_USE_EXA
    ExaDriverPtr	EXADriverPtr;
 #endif
@@ -944,6 +946,9 @@ Bool i965_prepare_composite(int op, PicturePtr pSrc, PicturePtr pMask,
 void i965_composite(PixmapPtr pDst, int srcX, int srcY,
 		    int maskX, int maskY, int dstX, int dstY, int w, int h);
 
+void
+i965_batch_flush_notify(ScrnInfoPtr pScrn);
+
 Bool
 i830_get_transformed_coordinates(int x, int y, PictTransformPtr transform,
 				 float *x_out, float *y_out);

src/i830_batchbuffer.c

@@ -201,4 +201,7 @@ intel_batch_flush(ScrnInfoPtr pScrn, Bool flushed)
      */
     if (pI830->memory_manager != NULL)
 	pI830->need_mi_flush = TRUE;
+
+    if (pI830->batch_flush_notify)
+	pI830->batch_flush_notify (pScrn);
 }

src/i830_driver.c

@@ -3364,6 +3364,11 @@ I830ScreenInit(int scrnIndex, ScreenPtr pScreen, int argc, char **argv)
       }
    }
 
+   if (IS_I965G(pI830))
+       pI830->batch_flush_notify = i965_batch_flush_notify;
+   else
+       pI830->batch_flush_notify = NULL;
+
    miInitializeBackingStore(pScreen);
    xf86SetBackingStore(pScreen);
    xf86SetSilkenMouse(pScreen);

src/i965_render.c

@@ -1000,6 +1000,7 @@ _emit_batch_header_for_composite_internal (ScrnInfoPtr pScrn, Bool check_twice)
 	render_state->vertex_buffer_bo = dri_bo_alloc (pI830->bufmgr, "vb",
 						       sizeof (gen4_vertex_buffer),
 						       4096);
+	render_state->vb_offset = 0;
     }
 
     bo_table[0] = pI830->batch_bo;
@@ -1480,15 +1481,22 @@ i965_composite(PixmapPtr pDst, int srcX, int srcY, int maskX, int maskY,
 	}
     }
 
+    /* We're about to do a BEGIN_BATCH(12) for the vertex setup. And
+     * we first need to ensure that that's not going to cause a flush
+     * since we need to not flush between setting up our vertices in
+     * the VB and emitting them into the batch. */
+    intel_batch_require_space(pScrn, pI830, 12 * 4);
+
     /* If the vertex buffer is too full, then we flush and re-emit all
      * necessary state into the batch for the composite operation. */
     if (render_state->vb_offset + VERTEX_FLOATS_PER_COMPOSITE > VERTEX_BUFFER_SIZE) {
 	dri_bo_unreference (render_state->vertex_buffer_bo);
 	render_state->vertex_buffer_bo = NULL;
-	render_state->vb_offset = 0;
-	_emit_batch_header_for_composite (pScrn);
     }
 
+    if (render_state->vertex_buffer_bo == NULL)
+	_emit_batch_header_for_composite (pScrn);
+
     /* Map the vertex_buffer buffer object so we can write to it. */
     dri_bo_map (render_state->vertex_buffer_bo, 1);
     vb = render_state->vertex_buffer_bo->virtual;
@@ -1571,6 +1579,19 @@ i965_composite(PixmapPtr pDst, int srcX, int srcY, int maskX, int maskY,
 #endif
 }
 
+void
+i965_batch_flush_notify(ScrnInfoPtr pScrn)
+{
+    I830Ptr pI830 = I830PTR(pScrn);
+    struct gen4_render_state *render_state = pI830->gen4_render_state;
+
+    /* Once a batch is emitted, we never want to map again any buffer
+     * object being referenced by that batch, (which would be very
+     * expensive). */
+    dri_bo_unreference (render_state->vertex_buffer_bo);
+    render_state->vertex_buffer_bo = NULL;
+}
+
 /**
  * Called at EnterVT so we can set up our offsets into the state buffer.
  */