diff --git a/src/gallium/drivers/i965/brw_batchbuffer.c b/src/gallium/drivers/i965/brw_batchbuffer.c
index a55be6faab0..d725e8b27e0 100644
--- a/src/gallium/drivers/i965/brw_batchbuffer.c
+++ b/src/gallium/drivers/i965/brw_batchbuffer.c
@@ -35,7 +35,6 @@
 #include "brw_structs.h"
 #include "intel_decode.h"
 
-#define USE_MALLOC_BUFFER 1
 #define ALWAYS_EMIT_MI_FLUSH 1
 
 enum pipe_error
@@ -50,14 +49,18 @@ brw_batchbuffer_reset(struct brw_batchbuffer *batch)
    if (ret)
       return ret;
 
-   if (batch->malloc_buffer)
-      batch->map = batch->malloc_buffer;
-   else 
-      batch->map = batch->sws->bo_map(batch->buf,
-                                      BRW_DATA_BATCH_BUFFER,
-                                      GL_TRUE);
-
    batch->size = BRW_BATCH_SIZE;
+
+   /* With map_range semantics, the winsys can decide whether to
+    * inject a malloc'ed bounce buffer instead of mapping directly.
+    */
+   batch->map = batch->sws->bo_map(batch->buf,
+                                   BRW_DATA_BATCH_BUFFER,
+                                   0, batch->size,
+                                   GL_TRUE,
+                                   GL_TRUE,
+                                   GL_TRUE);
+
    batch->ptr = batch->map;
    return PIPE_OK;
 }
@@ -68,11 +71,6 @@ brw_batchbuffer_alloc(struct brw_winsys_screen *sws,
 {
    struct brw_batchbuffer *batch = CALLOC_STRUCT(brw_batchbuffer);
 
-   batch->use_malloc_buffer = USE_MALLOC_BUFFER;
-   if (batch->use_malloc_buffer) {
-      batch->malloc_buffer = MALLOC(BRW_BATCH_SIZE);
-   }
-
    batch->sws = sws;
    batch->chipset = chipset;
    brw_batchbuffer_reset(batch);
@@ -83,11 +81,7 @@ brw_batchbuffer_alloc(struct brw_winsys_screen *sws,
 void
 brw_batchbuffer_free(struct brw_batchbuffer *batch)
 {
-   if (batch->malloc_buffer) {
-      FREE(batch->malloc_buffer);
-      batch->map = NULL;
-   }
-   else if (batch->map) {
+   if (batch->map) {
       batch->sws->bo_unmap(batch->buf);
       batch->map = NULL;
    }
@@ -134,18 +128,9 @@ _brw_batchbuffer_flush(struct brw_batchbuffer *batch,
    batch->ptr += 4;
    used = batch->ptr - batch->map;
 
-   if (batch->use_malloc_buffer) {
-      batch->sws->bo_subdata(batch->buf, 
-                             BRW_DATA_BATCH_BUFFER,
-                             0, used,
-                             batch->map );
-      batch->map = NULL;
-   }
-   else {
-      batch->sws->bo_unmap(batch->buf);
-      batch->map = NULL;
-   }
-
+   batch->sws->bo_flush_range(batch->buf, 0, used);
+   batch->sws->bo_unmap(batch->buf);
+   batch->map = NULL;
    batch->ptr = NULL;
       
    batch->sws->bo_exec(batch->buf, used );
diff --git a/src/gallium/drivers/i965/brw_batchbuffer.h b/src/gallium/drivers/i965/brw_batchbuffer.h
index 288a9d27555..7473f5bea4d 100644
--- a/src/gallium/drivers/i965/brw_batchbuffer.h
+++ b/src/gallium/drivers/i965/brw_batchbuffer.h
@@ -28,15 +28,6 @@ struct brw_batchbuffer {
    struct brw_winsys_buffer *buf;
    struct brw_chipset chipset;
 
-   /* Main-memory copy of the batch-buffer, built up incrementally &
-    * then copied as one to the true buffer.
-    *
-    * XXX: is this still necessary?
-    * XXX: if so, can this be hidden inside the GEM-specific winsys code?
-    */
-   boolean use_malloc_buffer;
-   uint8_t *malloc_buffer;
-
    /**
     * Values exported to speed up the writing the batchbuffer,
     * instead of having to go trough a accesor function for
diff --git a/src/gallium/drivers/i965/brw_pipe_query.c b/src/gallium/drivers/i965/brw_pipe_query.c
index 6a011737879..2eb862635cc 100644
--- a/src/gallium/drivers/i965/brw_pipe_query.c
+++ b/src/gallium/drivers/i965/brw_pipe_query.c
@@ -63,7 +63,7 @@ brw_query_get_result(struct pipe_context *pipe,
       if (brw->sws->bo_is_busy(query->bo) && !wait)
 	 return FALSE;
       
-      map = brw->sws->bo_map(query->bo, BRW_DATA_OTHER, GL_FALSE);
+      map = bo_map_read(brw->sws, query->bo);
       if (map == NULL)
 	 return FALSE;
       
diff --git a/src/gallium/drivers/i965/brw_screen_buffers.c b/src/gallium/drivers/i965/brw_screen_buffers.c
index 7ae386ffb34..d8141a3f5b9 100644
--- a/src/gallium/drivers/i965/brw_screen_buffers.c
+++ b/src/gallium/drivers/i965/brw_screen_buffers.c
@@ -11,6 +11,29 @@
 
 
 
+static void *
+brw_buffer_map_range( struct pipe_screen *screen,
+                      struct pipe_buffer *buffer,
+                      unsigned offset,
+                      unsigned length,
+                      unsigned usage )
+{
+   struct brw_screen *bscreen = brw_screen(screen); 
+   struct brw_winsys_screen *sws = bscreen->sws;
+   struct brw_buffer *buf = brw_buffer( buffer );
+
+   if (buf->user_buffer)
+      return buf->user_buffer;
+
+   return sws->bo_map( buf->bo, 
+                       BRW_DATA_OTHER,
+                       offset,
+                       length,
+                       (usage & PIPE_BUFFER_USAGE_CPU_WRITE) ? TRUE : FALSE,
+                       (usage & PIPE_BUFFER_USAGE_DISCARD) ? TRUE : FALSE,
+                       (usage & PIPE_BUFFER_USAGE_FLUSH_EXPLICIT) ? TRUE : FALSE);
+}
+
 static void *
 brw_buffer_map( struct pipe_screen *screen,
                 struct pipe_buffer *buffer,
@@ -25,9 +48,33 @@ brw_buffer_map( struct pipe_screen *screen,
 
    return sws->bo_map( buf->bo, 
                        BRW_DATA_OTHER,
-                       (usage & PIPE_BUFFER_USAGE_CPU_WRITE) ? TRUE : FALSE );
+                       0,
+                       buf->base.size,
+                       (usage & PIPE_BUFFER_USAGE_CPU_WRITE) ? TRUE : FALSE,
+                       FALSE,
+                       FALSE);
 }
 
+
+static void 
+brw_buffer_flush_mapped_range( struct pipe_screen *screen,
+                               struct pipe_buffer *buffer,
+                               unsigned offset,
+                               unsigned length )
+{
+   struct brw_screen *bscreen = brw_screen(screen); 
+   struct brw_winsys_screen *sws = bscreen->sws;
+   struct brw_buffer *buf = brw_buffer( buffer );
+
+   if (buf->user_buffer)
+      return;
+
+   sws->bo_flush_range( buf->bo, 
+                        offset,
+                        length );
+}
+
+
 static void 
 brw_buffer_unmap( struct pipe_screen *screen,
                    struct pipe_buffer *buffer )
@@ -148,6 +195,8 @@ void brw_screen_buffer_init(struct brw_screen *brw_screen)
    brw_screen->base.buffer_create = brw_buffer_create;
    brw_screen->base.user_buffer_create = brw_user_buffer_create;
    brw_screen->base.buffer_map = brw_buffer_map;
+   brw_screen->base.buffer_map_range = brw_buffer_map_range;
+   brw_screen->base.buffer_flush_mapped_range = brw_buffer_flush_mapped_range;
    brw_screen->base.buffer_unmap = brw_buffer_unmap;
    brw_screen->base.buffer_destroy = brw_buffer_destroy;
 }
diff --git a/src/gallium/drivers/i965/brw_winsys.h b/src/gallium/drivers/i965/brw_winsys.h
index f4a1e9d8edf..e72b928b064 100644
--- a/src/gallium/drivers/i965/brw_winsys.h
+++ b/src/gallium/drivers/i965/brw_winsys.h
@@ -169,7 +169,15 @@ struct brw_winsys_screen {
     */
    void *(*bo_map)(struct brw_winsys_buffer *buffer,
                    enum brw_buffer_data_type data_type,
-		   boolean write);
+                   unsigned offset,
+                   unsigned length,
+                   boolean write,
+                   boolean discard,
+                   boolean flush_explicit );
+
+   void (*bo_flush_range)( struct brw_winsys_buffer *buffer,
+                           unsigned offset,
+                           unsigned length );
 
    /**
     * Unmap a buffer.
@@ -189,6 +197,14 @@ struct brw_winsys_screen {
    void (*destroy)(struct brw_winsys_screen *iws);
 };
 
+static INLINE void *
+bo_map_read( struct brw_winsys_screen *sws, struct brw_winsys_buffer *buf )
+{
+   return sws->bo_map( buf,
+                       BRW_DATA_OTHER,
+                       0, buf->size,
+                       FALSE, FALSE, FALSE );
+}
 
 static INLINE void
 bo_reference(struct brw_winsys_buffer **ptr, struct brw_winsys_buffer *buf)
diff --git a/src/gallium/winsys/drm/i965/xlib/xlib_i965.c b/src/gallium/winsys/drm/i965/xlib/xlib_i965.c
index f46d9961c6e..ab5df56bc05 100644
--- a/src/gallium/winsys/drm/i965/xlib/xlib_i965.c
+++ b/src/gallium/winsys/drm/i965/xlib/xlib_i965.c
@@ -350,7 +350,11 @@ xlib_brw_check_aperture_space( struct brw_winsys_screen *iws,
 static void *
 xlib_brw_bo_map(struct brw_winsys_buffer *buffer,
                 enum brw_buffer_data_type data_type,
-                boolean write)
+                unsigned offset,
+                unsigned length,
+                boolean write,
+                boolean discard,
+                boolean explicit)
 {
    struct xlib_brw_buffer *buf = xlib_brw_buffer(buffer);
 
@@ -365,6 +369,15 @@ xlib_brw_bo_map(struct brw_winsys_buffer *buffer,
    return buf->virtual;
 }
 
+
+static void
+xlib_brw_bo_flush_range( struct brw_winsys_buffer *buffer,
+                         unsigned offset,
+                         unsigned length )
+{
+}
+
+
 static void 
 xlib_brw_bo_unmap(struct brw_winsys_buffer *buffer)
 {
@@ -380,7 +393,8 @@ xlib_brw_bo_unmap(struct brw_winsys_buffer *buffer)
 
       buf->modified = 0;
       
-      /* Consider dumping new buffer contents here.
+      /* Consider dumping new buffer contents here, using the
+       * flush-range info to minimize verbosity.
        */
    }
 }
@@ -421,6 +435,7 @@ xlib_create_brw_winsys_screen( void )
    ws->base.bo_references        = xlib_brw_bo_references;
    ws->base.check_aperture_space = xlib_brw_check_aperture_space;
    ws->base.bo_map               = xlib_brw_bo_map;
+   ws->base.bo_flush_range       = xlib_brw_bo_flush_range;
    ws->base.bo_unmap             = xlib_brw_bo_unmap;
    ws->base.bo_wait_idle         = xlib_brw_bo_wait_idle;