diff --git a/src/mesa/drivers/dri/i915/i915_vtbl.c b/src/mesa/drivers/dri/i915/i915_vtbl.c
index 2fee0ca502a..3d914fb3fbb 100644
--- a/src/mesa/drivers/dri/i915/i915_vtbl.c
+++ b/src/mesa/drivers/dri/i915/i915_vtbl.c
@@ -45,7 +45,6 @@
 
 static void i915_render_start( struct intel_context *intel )
 {
-   GLcontext *ctx = &intel->ctx;
    struct i915_context *i915 = i915_context(&intel->ctx);
 
    i915ValidateFragmentProgram( i915 );
diff --git a/src/mesa/drivers/dri/i915/intel_blit.c b/src/mesa/drivers/dri/i915/intel_blit.c
index e55475d3583..5cbbea1005e 100644
--- a/src/mesa/drivers/dri/i915/intel_blit.c
+++ b/src/mesa/drivers/dri/i915/intel_blit.c
@@ -438,6 +438,10 @@ void intelClearWithBlit(GLcontext *ctx, GLbitfield mask, GLboolean all,
                const struct intel_renderbuffer *irb
                   = intel_renderbuffer(ctx->DrawBuffer->
                                        Attachment[buf].Renderbuffer);
+	       struct buffer *write_buffer = 
+		  intel_region_buffer(intel, irb->region,
+				      all ? INTEL_WRITE_FULL : INTEL_WRITE_PART);
+
                GLuint clearVal;
                GLint pitch, cpp;
                GLuint BR13, CMD;
@@ -448,7 +452,7 @@ void intelClearWithBlit(GLcontext *ctx, GLbitfield mask, GLboolean all,
                pitch = irb->region->pitch;
                cpp = irb->region->cpp;
 
-	       DBG("%s dst:buf(%d)/%d+%d %d,%d sz:%dx%d\n",
+	       DBG("%s dst:buf(%p)/%d+%d %d,%d sz:%dx%d\n",
 		   __FUNCTION__,
 		   irb->region->buffer, (pitch * cpp), 
 		   irb->region->draw_offset, 
@@ -495,7 +499,7 @@ void intelClearWithBlit(GLcontext *ctx, GLbitfield mask, GLboolean all,
                OUT_BATCH( BR13 );
                OUT_BATCH( (b.y1 << 16) | b.x1 );
                OUT_BATCH( (b.y2 << 16) | b.x2 );
-               OUT_RELOC( irb->region->buffer, DRM_MM_TT|DRM_MM_WRITE,
+               OUT_RELOC( write_buffer, DRM_MM_TT|DRM_MM_WRITE,
                           irb->region->draw_offset );
                OUT_BATCH( clearVal );
                ADVANCE_BATCH();
diff --git a/src/mesa/drivers/dri/i915/intel_buffer_objects.c b/src/mesa/drivers/dri/i915/intel_buffer_objects.c
index 9bd7d8b4169..6cb2a65035c 100644
--- a/src/mesa/drivers/dri/i915/intel_buffer_objects.c
+++ b/src/mesa/drivers/dri/i915/intel_buffer_objects.c
@@ -32,6 +32,7 @@
 
 #include "intel_context.h"
 #include "intel_buffer_objects.h"
+#include "intel_regions.h"
 #include "intel_bufmgr.h"
 
 
@@ -46,18 +47,44 @@ static struct gl_buffer_object *intel_bufferobj_alloc( GLcontext *ctx,
 						       GLenum target )
 {
    struct intel_context *intel = intel_context(ctx);
-   struct intel_buffer_object *obj = MALLOC_STRUCT(intel_buffer_object);
+   struct intel_buffer_object *obj = CALLOC_STRUCT(intel_buffer_object);
 
    _mesa_initialize_buffer_object(&obj->Base, name, target);
 
-   /* XXX:  We generate our own handle, which is different to 'name' above.
-    */
    bmGenBuffers(intel, "bufferobj", 1, &obj->buffer, 0);
 
    return &obj->Base;
 }
 
 
+/* Break the COW tie to the region.  The region gets to keep the data.
+ */
+void intel_bufferobj_release_region( struct intel_context *intel,
+				     struct intel_buffer_object *intel_obj )
+{
+   assert(intel_obj->region->buffer == intel_obj->buffer);
+   intel_obj->region->pbo = NULL;
+   intel_obj->region = NULL;
+   intel_obj->buffer = NULL; /* refcount? */
+
+   /* This leads to a large number of buffer deletion/creation events.
+    * Currently the drm doesn't like that:
+    */
+   bmGenBuffers(intel, "buffer object", 1, &intel_obj->buffer, 0);
+   bmBufferData(intel, intel_obj->buffer, intel_obj->Base.Size, NULL, 0);
+}
+
+/* Break the COW tie to the region.  Both the pbo and the region end
+ * up with a copy of the data.
+ */
+void intel_bufferobj_cow( struct intel_context *intel,
+			  struct intel_buffer_object *intel_obj )
+{
+   assert(intel_obj->region);
+   intel_region_cow( intel, intel_obj->region );
+}
+
+
 /**
  * Deallocate/free a vertex/pixel buffer object.
  * Called via glDeleteBuffersARB().
@@ -70,8 +97,12 @@ static void intel_bufferobj_free( GLcontext *ctx,
 
    assert(intel_obj);
 
-   if (intel_obj->buffer) 
+   if (intel_obj->region) {
+      intel_bufferobj_release_region(intel, intel_obj);
+   }
+   else if (intel_obj->buffer) {
       bmDeleteBuffers( intel, 1, &intel_obj->buffer );
+   }
   
    _mesa_free(intel_obj);
 }
@@ -94,13 +125,11 @@ static void intel_bufferobj_data( GLcontext *ctx,
    struct intel_context *intel = intel_context(ctx);
    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 
-   /* XXX: do something useful with 'usage' (eg. populate flags
-    * argument below)
-    */
-   assert(intel_obj);
+   intel_obj->Base.Size = size;
+   intel_obj->Base.Usage = usage;
 
-   obj->Size = size;
-   obj->Usage = usage;
+   if (intel_obj->region)
+      intel_bufferobj_release_region(intel, intel_obj);
 
    bmBufferData(intel, intel_obj->buffer, size, data, 0);
 }
@@ -123,6 +152,10 @@ static void intel_bufferobj_subdata( GLcontext *ctx,
    struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
 
    assert(intel_obj);
+
+   if (intel_obj->region)
+      intel_bufferobj_cow(intel, intel_obj);
+
    bmBufferSubData(intel, intel_obj->buffer, offset, size, data);
 }
 
@@ -160,6 +193,10 @@ static void *intel_bufferobj_map( GLcontext *ctx,
    /* XXX: Translate access to flags arg below:
     */
    assert(intel_obj);
+
+   if (intel_obj->region)
+      intel_bufferobj_cow(intel, intel_obj);
+
    obj->Pointer = bmMapBuffer(intel, intel_obj->buffer, 0);
    return obj->Pointer;
 }
@@ -182,8 +219,17 @@ static GLboolean intel_bufferobj_unmap( GLcontext *ctx,
    return GL_TRUE;
 }
 
-struct buffer *intel_bufferobj_buffer( const struct intel_buffer_object *intel_obj )
+struct buffer *intel_bufferobj_buffer( struct intel_context *intel,
+				       struct intel_buffer_object *intel_obj,
+				       GLuint flag )
 {
+   if (intel_obj->region) {
+      if (flag == INTEL_WRITE_PART)
+	 intel_bufferobj_cow(intel, intel_obj);
+      else if (flag == INTEL_WRITE_FULL)
+	 intel_bufferobj_release_region(intel, intel_obj);
+   }
+
    return intel_obj->buffer;
 }  
 
diff --git a/src/mesa/drivers/dri/i915/intel_buffer_objects.h b/src/mesa/drivers/dri/i915/intel_buffer_objects.h
index 445ddf4273b..93dfbbc3c92 100644
--- a/src/mesa/drivers/dri/i915/intel_buffer_objects.h
+++ b/src/mesa/drivers/dri/i915/intel_buffer_objects.h
@@ -31,6 +31,7 @@
 #include "mtypes.h"
 
 struct intel_context;
+struct intel_region;
 struct gl_buffer_object;
 
 
@@ -40,12 +41,18 @@ struct gl_buffer_object;
 struct intel_buffer_object {
    struct gl_buffer_object Base;
    struct buffer *buffer;   /* the low-level buffer manager's buffer handle */
+   
+   struct intel_region *region; /* Is there a zero-copy texture
+				   associated with this (pixel)
+				   buffer object? */
 };
 
 
 /* Get the bm buffer associated with a GL bufferobject:
  */
-struct buffer *intel_bufferobj_buffer( const struct intel_buffer_object *obj );
+struct buffer *intel_bufferobj_buffer( struct intel_context *intel,
+				       struct intel_buffer_object *obj,
+				       GLuint flag );
 
 /* Hook the bufferobject implementation into mesa: 
  */
@@ -67,4 +74,12 @@ intel_buffer_object( struct gl_buffer_object *obj )
       return NULL;
 }
 
+/* Helpers for zerocopy image uploads.  See also intel_regions.h:
+ */
+void intel_bufferobj_cow( struct intel_context *intel,
+			  struct intel_buffer_object *intel_obj );
+void intel_bufferobj_release_region( struct intel_context *intel,
+				     struct intel_buffer_object *intel_obj );
+
+
 #endif
diff --git a/src/mesa/drivers/dri/i915/intel_context.h b/src/mesa/drivers/dri/i915/intel_context.h
index c7a4855aee2..150baffc4ea 100644
--- a/src/mesa/drivers/dri/i915/intel_context.h
+++ b/src/mesa/drivers/dri/i915/intel_context.h
@@ -68,6 +68,9 @@ extern void intelFallback( struct intel_context *intel, GLuint bit, GLboolean mo
 #define FALLBACK( intel, bit, mode ) intelFallback( intel, bit, mode )
 
 
+#define INTEL_WRITE_PART  0x1
+#define INTEL_WRITE_FULL  0x2
+#define INTEL_READ        0x4
 
 struct intel_texture_object
 {
diff --git a/src/mesa/drivers/dri/i915/intel_pixel_draw.c b/src/mesa/drivers/dri/i915/intel_pixel_draw.c
index 52d6856184f..b35ab635658 100644
--- a/src/mesa/drivers/dri/i915/intel_pixel_draw.c
+++ b/src/mesa/drivers/dri/i915/intel_pixel_draw.c
@@ -299,6 +299,7 @@ static GLboolean do_blit_drawpixels( GLcontext *ctx,
       drm_clip_rect_t *box = dPriv->pClipRects;
       drm_clip_rect_t rect;
       drm_clip_rect_t dest_rect;
+      struct buffer *src_buffer = intel_bufferobj_buffer(intel, src, INTEL_READ);
       int i;
       
       dest_rect.x1 = dPriv->x + x;
@@ -314,7 +315,7 @@ static GLboolean do_blit_drawpixels( GLcontext *ctx,
 	 intelEmitCopyBlit( intel,
 			    dest->cpp,
 			    rowLength, 
-			    intel_bufferobj_buffer(src), src_offset,
+			    src_buffer, src_offset,
 			    dest->pitch, 
 			    dest->buffer, 0,
 			    rect.x1 - dest_rect.x1, 
diff --git a/src/mesa/drivers/dri/i915/intel_pixel_read.c b/src/mesa/drivers/dri/i915/intel_pixel_read.c
index 7a5740fae82..13890550d11 100644
--- a/src/mesa/drivers/dri/i915/intel_pixel_read.c
+++ b/src/mesa/drivers/dri/i915/intel_pixel_read.c
@@ -250,6 +250,13 @@ static GLboolean do_blit_readpixels( GLcontext *ctx,
 
    if (intel->driDrawable->numClipRects)
    {
+      GLboolean all = (width * height * src->cpp == dst->Base.Size &&
+		       x == 0 &&
+		       dst_offset == 0);
+		       
+      struct buffer *dst_buffer = intel_bufferobj_buffer(intel, dst, 
+							 all ? INTEL_WRITE_FULL : 
+							 INTEL_WRITE_PART);
        __DRIdrawablePrivate *dPriv = intel->driDrawable;
       int nbox = dPriv->numClipRects;
       drm_clip_rect_t *box = dPriv->pClipRects;
@@ -273,7 +280,7 @@ static GLboolean do_blit_readpixels( GLcontext *ctx,
 			    src->cpp,
 			    src->pitch, src->buffer, 0,
 			    rowLength, 
-			    intel_bufferobj_buffer(dst), dst_offset,
+			    dst_buffer, dst_offset,
 			    rect.x1, 
 			    rect.y1,
 			    rect.x1 - src_rect.x1, 
diff --git a/src/mesa/drivers/dri/i915/intel_regions.c b/src/mesa/drivers/dri/i915/intel_regions.c
index 5c874625e75..2cde3df9822 100644
--- a/src/mesa/drivers/dri/i915/intel_regions.c
+++ b/src/mesa/drivers/dri/i915/intel_regions.c
@@ -43,6 +43,7 @@
 #include "intel_regions.h"
 #include "intel_blit.h"
 #include "intel_bufmgr.h"
+#include "intel_buffer_objects.h"
 
 #define FILE_DEBUG_FLAG DEBUG_BUFMGR
 
@@ -53,6 +54,9 @@ GLubyte *intel_region_map(struct intel_context *intel, struct intel_region *regi
 {
    DBG("%s\n", __FUNCTION__);
    if (!region->map_refcount++) {
+      if (region->pbo)
+	 intel_region_cow(intel, region);
+
       region->map = bmMapBuffer(intel, region->buffer, 0);
    }
 
@@ -112,7 +116,12 @@ void intel_region_release( struct intel_context *intel,
 
    if ((*region)->refcount == 0) {
       assert((*region)->map_refcount == 0);
-      bmDeleteBuffers(intel, 1, &(*region)->buffer);
+
+      if ((*region)->pbo) 
+	 intel_region_release_pbo( intel, *region );
+      else
+	 bmDeleteBuffers(intel, 1, &(*region)->buffer);
+
       free(*region);
    }
    *region = NULL;
@@ -203,6 +212,17 @@ void intel_region_data(struct intel_context *intel,
 {
    DBG("%s\n", __FUNCTION__);
 
+   if (dst->pbo) {
+      if (dstx == 0 && 
+	  dsty == 0 && 
+	  width == dst->pitch && 
+	  height == dst->height)
+	 intel_region_release_pbo(intel, dst);
+      else
+	 intel_region_cow(intel, dst);
+   }
+	 
+
    LOCK_HARDWARE(intel);
    
    _mesa_copy_rect(intel_region_map(intel, dst) + dst_offset,
@@ -234,6 +254,16 @@ void intel_region_copy( struct intel_context *intel,
 {
    DBG("%s\n", __FUNCTION__);
 
+   if (dst->pbo) {
+      if (dstx == 0 && 
+	  dsty == 0 && 
+	  width == dst->pitch && 
+	  height == dst->height)
+	 intel_region_release_pbo(intel, dst);
+      else
+	 intel_region_cow(intel, dst);
+   }
+
    assert(src->cpp == dst->cpp);
 
    intelEmitCopyBlit(intel,
@@ -257,6 +287,16 @@ void intel_region_fill( struct intel_context *intel,
 {
    DBG("%s\n", __FUNCTION__);
    
+   if (dst->pbo) {
+      if (dstx == 0 && 
+	  dsty == 0 && 
+	  width == dst->pitch && 
+	  height == dst->height)
+	 intel_region_release_pbo(intel, dst);
+      else
+	 intel_region_cow(intel, dst);
+   }
+
    intelEmitFillBlit(intel,
 		     dst->cpp,
 		     dst->pitch, dst->buffer, dst_offset, 
@@ -265,3 +305,96 @@ void intel_region_fill( struct intel_context *intel,
 		     color );
 }
 
+/* Attach to a pbo, discarding our data.  Effectively zero-copy upload
+ * the pbo's data.
+ */
+void intel_region_attach_pbo( struct intel_context *intel,
+			       struct intel_region *region,
+			       struct intel_buffer_object *pbo )
+{
+   if (region->pbo == pbo)
+      return;
+
+   /* If there is already a pbo attached, break the cow tie now.
+    * Don't call intel_region_release_pbo() as that would
+    * unnecessarily allocate a new buffer we would have to immediately
+    * discard.
+    */
+   if (region->pbo) {
+      region->pbo->region = NULL;
+      region->pbo = NULL;
+      region->buffer = NULL;	/* refcount? */
+   }
+
+   if (region->buffer) {
+      bmDeleteBuffers(intel, 1, region->buffer);
+   }
+
+   region->pbo = pbo;
+   region->pbo->region = region;
+   region->buffer = pbo->buffer;	/* refcount? */
+
+   _mesa_printf("%s attach buffer %p from pbo\n", region->buffer);
+}
+
+
+/* Break the COW tie to the pbo.  The pbo gets to keep the data.
+ */
+void intel_region_release_pbo( struct intel_context *intel,
+			       struct intel_region *region )
+{
+   assert(region->buffer == region->pbo->buffer);
+   region->pbo->region = NULL;   
+   region->pbo = NULL;
+   region->buffer = NULL;	/* refcount? */
+
+   bmGenBuffers(intel, "region", 1, &region->buffer, 0);
+   bmBufferData(intel, region->buffer, 
+		region->cpp * region->pitch * region->height, NULL, 0);
+}
+
+/* Break the COW tie to the pbo.  Both the pbo and the region end up
+ * with a copy of the data.
+ */
+void intel_region_cow( struct intel_context *intel,
+		       struct intel_region *region )
+{
+   struct intel_buffer_object *pbo = region->pbo;
+
+   intel_region_release_pbo(intel, region);
+
+   assert(region->cpp * 
+	  region->pitch * 
+	  region->height == pbo->Base.Size);
+
+   _mesa_printf("%s (%d bytes)\n", __FUNCTION__, pbo->Base.Size);
+
+   /* Now blit from the texture buffer to the new buffer: 
+    */
+
+   /* LOCKING??? */
+   intelEmitCopyBlit( intel,
+		      region->cpp,
+		      region->pitch, 
+		      region->buffer, 0,
+		      region->pitch, 
+		      pbo->buffer, 0,
+		      0,0,
+		      0,0,
+		      region->pitch,
+		      region->height );
+}
+
+struct buffer *intel_region_buffer( struct intel_context *intel,
+				    struct intel_region *region,
+				    GLuint flag )
+{
+   if (region->pbo) {
+      if (flag == INTEL_WRITE_PART)
+	 intel_region_cow(intel, region);
+      else if (flag == INTEL_WRITE_FULL)
+	 intel_region_release_pbo(intel, region);
+   }
+
+   return region->buffer;
+}  
diff --git a/src/mesa/drivers/dri/i915/intel_regions.h b/src/mesa/drivers/dri/i915/intel_regions.h
index eb744dfda55..a54de926d35 100644
--- a/src/mesa/drivers/dri/i915/intel_regions.h
+++ b/src/mesa/drivers/dri/i915/intel_regions.h
@@ -31,6 +31,7 @@
 #include "mtypes.h"
 #include "intel_bufmgr.h"		/* for DBG! */
 struct intel_context;
+struct intel_buffer_object;
 
 /**
  * A layer on top of the bufmgr buffers that adds a few useful things:
@@ -50,6 +51,8 @@ struct intel_region {
    GLuint map_refcount;  /**< Reference count for mapping */
 
    GLuint draw_offset; /**< Offset of drawing address within the region */
+
+   struct intel_buffer_object *pbo; /* zero-copy uploads */
 };
 
 
@@ -115,5 +118,18 @@ void intel_region_fill( struct intel_context *intel,
 			GLuint width, GLuint height,
 			GLuint color );
 
+/* Helpers for zerocopy uploads, particularly texture image uploads:
+ */
+void intel_region_attach_pbo( struct intel_context *intel,
+			       struct intel_region *region,
+			      struct intel_buffer_object *pbo );
+void intel_region_release_pbo( struct intel_context *intel,
+			       struct intel_region *region );
+void intel_region_cow( struct intel_context *intel,
+		       struct intel_region *region );
+
+struct buffer *intel_region_buffer( struct intel_context *intel,
+				    struct intel_region *region,
+				    GLuint flag );
 
 #endif
diff --git a/src/mesa/drivers/dri/i915/intel_tex_image.c b/src/mesa/drivers/dri/i915/intel_tex_image.c
index c67e58533fa..26bd0003c3c 100644
--- a/src/mesa/drivers/dri/i915/intel_tex_image.c
+++ b/src/mesa/drivers/dri/i915/intel_tex_image.c
@@ -153,9 +153,9 @@ static GLuint target_to_face( GLenum target )
 /* There are actually quite a few combinations this will work for,
  * more than what I've listed here.
  */
-static GLboolean check_pbo_blit( GLint internalFormat,
-				 GLenum format, GLenum type, 
-				 const struct gl_texture_format *mesa_format )
+static GLboolean check_pbo_format( GLint internalFormat,
+				   GLenum format, GLenum type, 
+				   const struct gl_texture_format *mesa_format )
 {
    switch (internalFormat) {
    case 4:
@@ -178,7 +178,8 @@ static GLboolean check_pbo_blit( GLint internalFormat,
 }
 
 
-
+/* XXX: Do this for TexSubImage also:
+ */
 static GLboolean try_pbo_upload( struct intel_context *intel,
 				 struct intel_texture_image *intelImage,
 				 const struct gl_pixelstore_attrib *unpack,
@@ -187,11 +188,11 @@ static GLboolean try_pbo_upload( struct intel_context *intel,
 				 GLenum format, GLenum type, 
 				 const void *pixels)
 {
-   struct intel_buffer_object *intelObj = intel_buffer_object(unpack->BufferObj);
+   struct intel_buffer_object *pbo = intel_buffer_object(unpack->BufferObj);
    GLuint src_offset, src_stride;
    GLuint dst_offset, dst_stride;
 
-   if (!intelObj ||
+   if (!pbo ||
        intel->ctx._ImageTransferState ||
        unpack->SkipPixels ||
        unpack->SkipRows) {
@@ -199,11 +200,6 @@ static GLboolean try_pbo_upload( struct intel_context *intel,
       return GL_FALSE;
    }
     
-   if (!check_pbo_blit(internalFormat, format, type, intelImage->base.TexFormat)) {
-      _mesa_printf("%s - bad format for blit\n", __FUNCTION__);
-      return GL_FALSE;
-   }
-
    src_offset = (GLuint) pixels;
 
    if (unpack->RowLength > 0)
@@ -219,19 +215,25 @@ static GLboolean try_pbo_upload( struct intel_context *intel,
 
    intelFlush( &intel->ctx );
    LOCK_HARDWARE( intel );
+   {
+      struct buffer *src_buffer = intel_bufferobj_buffer(intel, pbo, INTEL_READ);
+      struct buffer *dst_buffer = intel_region_buffer(intel, intelImage->mt->region, 
+						      INTEL_WRITE_FULL);
 
-   intelEmitCopyBlit( intel,
-		      intelImage->mt->cpp,
-		      src_stride, intel_bufferobj_buffer(intelObj), src_offset,
-		      dst_stride, intelImage->mt->region->buffer, dst_offset,
-		      0, 
-		      0,
-		      0, 
-		      0,
-		      width,
-		      height );
 
-   intel_batchbuffer_flush( intel->batch );   
+      intelEmitCopyBlit( intel,
+			 intelImage->mt->cpp,
+			 src_stride, src_buffer, src_offset,
+			 dst_stride, dst_buffer, dst_offset,
+			 0, 
+			 0,
+			 0, 
+			 0,
+			 width,
+			 height );
+
+      intel_batchbuffer_flush( intel->batch );   
+   }
    UNLOCK_HARDWARE( intel );
 
    return GL_TRUE;
@@ -239,6 +241,58 @@ static GLboolean try_pbo_upload( struct intel_context *intel,
 
 
 
+static GLboolean try_pbo_zcopy( struct intel_context *intel,
+				 struct intel_texture_image *intelImage,
+				 const struct gl_pixelstore_attrib *unpack,
+				 GLint internalFormat,
+				 GLint width, GLint height,
+				 GLenum format, GLenum type, 
+				 const void *pixels)
+{
+   struct intel_buffer_object *pbo = intel_buffer_object(unpack->BufferObj);
+   GLuint src_offset, src_stride;
+   GLuint dst_offset, dst_stride;
+
+   if (!pbo ||
+       intel->ctx._ImageTransferState ||
+       unpack->SkipPixels ||
+       unpack->SkipRows) {
+      _mesa_printf("%s: failure 1\n", __FUNCTION__);
+      return GL_FALSE;
+   }
+    
+   src_offset = (GLuint) pixels;
+
+   if (unpack->RowLength > 0)
+      src_stride = unpack->RowLength;
+   else
+      src_stride = width;
+
+   dst_offset = intel_miptree_image_offset(intelImage->mt, 
+					   intelImage->face, 
+					   intelImage->level);
+
+   dst_stride = intelImage->mt->pitch;
+
+   if (src_stride != dst_stride ||
+       dst_offset != 0 ||
+       src_offset != 0) {
+      _mesa_printf("%s: failure 2\n", __FUNCTION__);
+      return GL_FALSE;
+   }
+
+   intel_region_attach_pbo( intel,
+			    intelImage->mt->region,
+			    pbo );
+
+   return GL_TRUE;
+}
+    
+      
+
+
+
+
 static void intelTexImage(GLcontext *ctx, 
 			  GLint dims,
 			  GLenum target, GLint level,
@@ -345,43 +399,57 @@ static void intelTexImage(GLcontext *ctx,
    }
 
 
+   assert(!intelImage->mt);
+
    if (intelObj->mt && 
-       intelObj->mt != intelImage->mt &&
        intel_miptree_match_image(intelObj->mt, &intelImage->base,
 				 intelImage->face, intelImage->level)) {
       
-      if (intelImage->mt) {
-	 intel_miptree_release(intel, &intelImage->mt);
-      }
-
       intel_miptree_reference(&intelImage->mt, intelObj->mt);
       assert(intelImage->mt);
    }
 
-   if (!intelImage->mt) {
-      if (INTEL_DEBUG & DEBUG_TEXTURE)
-	 _mesa_printf("XXX: Image did not fit into tree - storing in local memory!\n");
-   }
+   if (!intelImage->mt) 
+      DBG("XXX: Image did not fit into tree - storing in local memory!\n");
 
-
-   /* Attempt to use the blitter for PBO image uploads: 
-    *
-    * Next step would be texturing directly from PBO's.
+   /* PBO fastpaths:
     */
    if (dims <= 2 &&
        intelImage->mt &&
-       intel_buffer_object(unpack->BufferObj)) {
+       intel_buffer_object(unpack->BufferObj) &&
+       check_pbo_format(internalFormat, format, 
+			type, intelImage->base.TexFormat)) {
+
+      DBG("trying pbo upload\n");
+
+      /* Attempt to texture directly from PBO data (zero copy upload).
+       * This is about twice as fast as regular uploads:
+       */
+      if (intelObj->mt == intelImage->mt && 
+	  intelObj->mt->first_level == level &&
+	  intelObj->mt->last_level == level) {
+
+	 if (try_pbo_zcopy(intel, intelImage, unpack,
+			   internalFormat,
+			   width, height, format, type, pixels)) {
+
+	    DBG("pbo zcopy upload succeeded\n");
+	    return;
+	 }
+      }	   
 
-      _mesa_printf("trying pbo upload\n");
       
+      /* Otherwise, attempt to use the blitter for PBO image uploads.
+       * This is about 20% faster than regular uploads:
+       */
       if (try_pbo_upload(intel, intelImage, unpack, 
 			 internalFormat,
 			 width, height, format, type, pixels)) {
-	 _mesa_printf("pbo upload succeeded\n");
+	 DBG("pbo upload succeeded\n");
 	 return;
       }
 
-      _mesa_printf("pbo upload failed\n");
+      DBG("pbo upload failed\n");
    }