r300g: rebuild winsys/pipe buffer handling and add buffer map

This creates a cleaner winsys and drop the simple screen stuff.

It makes r300g use pb_bufmgr structs and adds usage of the cached
bufmgr for vertex/index buffers.

It also avoids mapping too often.

I'm not 100% sure this is perfect but it won't find its own bugs.

Signed-off-by: Dave Airlie <airlied@redhat.com>

23 files changed, 1247 insertions, 347 deletions

diff --git a/src/gallium/drivers/r300/Makefile b/src/gallium/drivers/r300/Makefile
index afddcb161f..5707309bdf 100644
--- a/src/gallium/drivers/r300/Makefile
+++ b/src/gallium/drivers/r300/Makefile
@@ -14,6 +14,7 @@ C_SOURCES = \
 	r300_query.c \
 	r300_render.c \
 	r300_screen.c \
+	r300_screen_buffer.c \
 	r300_state.c \
 	r300_state_derived.c \
 	r300_state_invariant.c \
diff --git a/src/gallium/drivers/r300/r300_context.c b/src/gallium/drivers/r300/r300_context.c
index f631b4ed27..a8a6a07a84 100644
--- a/src/gallium/drivers/r300/r300_context.c
+++ b/src/gallium/drivers/r300/r300_context.c
@@ -24,6 +24,7 @@
 
 #include "util/u_memory.h"
 #include "util/u_simple_list.h"
+#include "util/u_upload_mgr.h"
 
 #include "r300_blit.h"
 #include "r300_context.h"
@@ -54,6 +55,9 @@ static void r300_destroy_context(struct pipe_context* context)
         FREE(query);
     }
 
+    u_upload_destroy(r300->upload_vb);
+    u_upload_destroy(r300->upload_ib);
+
     FREE(r300->blend_color_state.state);
     FREE(r300->clip_state.state);
     FREE(r300->fb_state.state);
@@ -70,11 +74,7 @@ r300_is_texture_referenced(struct pipe_context *pipe,
                            struct pipe_texture *texture,
                            unsigned face, unsigned level)
 {
-    struct pipe_buffer* buf = 0;
-
-    r300_get_texture_buffer(pipe->screen, texture, &buf, NULL);
-
-    return pipe->is_buffer_referenced(pipe, buf);
+    return 0;
 }
 
 static unsigned int
@@ -136,14 +136,14 @@ struct pipe_context* r300_create_context(struct pipe_screen* screen,
 {
     struct r300_context* r300 = CALLOC_STRUCT(r300_context);
     struct r300_screen* r300screen = r300_screen(screen);
-    struct radeon_winsys* radeon_winsys = r300screen->radeon_winsys;
+    struct r300_winsys_screen *rws = r300screen->rws;
 
     if (!r300)
         return NULL;
 
-    r300->winsys = radeon_winsys;
+    r300->rws = rws;
 
-    r300->context.winsys = (struct pipe_winsys*)radeon_winsys;
+    r300->context.winsys = (struct pipe_winsys*)rws;
     r300->context.screen = screen;
     r300->context.priv = priv;
 
@@ -201,11 +201,31 @@ struct pipe_context* r300_create_context(struct pipe_screen* screen,
 
     r300->invariant_state.dirty = TRUE;
 
-    r300->winsys->set_flush_cb(r300->winsys, r300_flush_cb, r300);
+    rws->set_flush_cb(r300->rws, r300_flush_cb, r300);
     r300->dirty_state = R300_NEW_KITCHEN_SINK;
     r300->dirty_hw++;
 
     r300->blitter = util_blitter_create(&r300->context);
 
+    r300->upload_ib = u_upload_create(screen,
+				      32 * 1024, 16,
+				      PIPE_BUFFER_USAGE_INDEX);
+
+    if (r300->upload_ib == NULL)
+        goto no_upload_ib;
+
+    r300->upload_vb = u_upload_create(screen,
+				      128 * 1024, 16,
+				      PIPE_BUFFER_USAGE_VERTEX);
+    if (r300->upload_vb == NULL)
+        goto no_upload_vb;
+
     return &r300->context;
+
+    //    u_upload_destroy(r300->upload_vb);
+ no_upload_ib:
+    u_upload_destroy(r300->upload_ib);
+ no_upload_vb:
+    FREE(r300);
+    return NULL;
 }
diff --git a/src/gallium/drivers/r300/r300_context.h b/src/gallium/drivers/r300/r300_context.h
index 443af4ec2e..749bac0498 100644
--- a/src/gallium/drivers/r300/r300_context.h
+++ b/src/gallium/drivers/r300/r300_context.h
@@ -32,6 +32,7 @@
 
 #include "r300_screen.h"
 
+struct u_upload_mgr;
 struct r300_context;
 
 struct r300_fragment_shader;
@@ -238,7 +239,7 @@ struct r300_texture {
     boolean is_npot;
 
     /* Pipe buffer backing this texture. */
-    struct pipe_buffer* buffer;
+    struct r300_winsys_buffer *buffer;
 
     /* Registers carrying texture format data. */
     struct r300_texture_state state;
@@ -265,7 +266,7 @@ struct r300_context {
     struct pipe_context context;
 
     /* The interface to the windowing system, etc. */
-    struct radeon_winsys* winsys;
+    struct r300_winsys_screen *rws;
     /* Draw module. Used mostly for SW TCL. */
     struct draw_context* draw;
     /* Accelerated blit support. */
@@ -330,6 +331,7 @@ struct r300_context {
     /* Vertex elements for Gallium. */
     struct pipe_vertex_element vertex_element[PIPE_MAX_ATTRIBS];
     int vertex_element_count;
+    bool any_user_vbs;
 
     struct pipe_stencil_ref stencil_ref;
 
@@ -343,6 +345,9 @@ struct r300_context {
     boolean polygon_offset_enabled;
     /* Z buffer bit depth. */
     uint32_t zbuffer_bpp;
+
+    struct u_upload_mgr *upload_vb;
+    struct u_upload_mgr *upload_ib;
 };
 
 /* Convenience cast wrapper. */
diff --git a/src/gallium/drivers/r300/r300_cs.h b/src/gallium/drivers/r300/r300_cs.h
index 151f72b0fe..ad07efbffd 100644
--- a/src/gallium/drivers/r300/r300_cs.h
+++ b/src/gallium/drivers/r300/r300_cs.h
@@ -51,7 +51,7 @@
 
 #define CS_LOCALS(context) \
     struct r300_context* const cs_context_copy = (context); \
-    struct radeon_winsys* cs_winsys = cs_context_copy->winsys; \
+    struct r300_winsys_screen *cs_winsys = cs_context_copy->rws; \
     int cs_count = 0; (void) cs_count;
 
 #define CHECK_CS(size) \
@@ -105,22 +105,34 @@
     cs_count--; \
 } while (0)
 
-#define OUT_CS_RELOC(bo, offset, rd, wd, flags) do { \
+#define OUT_CS_BUF_RELOC(bo, offset, rd, wd, flags) do { \
     DBG(cs_context_copy, DBG_CS, "r300: writing relocation for buffer %p, offset %d, " \
             "domains (%d, %d, %d)\n", \
         bo, offset, rd, wd, flags); \
     assert(bo); \
     cs_winsys->write_cs_dword(cs_winsys, offset); \
-    cs_winsys->write_cs_reloc(cs_winsys, bo, rd, wd, flags); \
+    r300_buffer_write_reloc(cs_winsys, r300_buffer(bo), rd, wd, flags);	\
     cs_count -= 3; \
 } while (0)
 
-#define OUT_CS_RELOC_NO_OFFSET(bo, rd, wd, flags) do { \
+
+#define OUT_CS_TEX_RELOC(tex, offset, rd, wd, flags) do { \
+    DBG(cs_context_copy, DBG_CS, "r300: writing relocation for texture %p, offset %d, " \
+            "domains (%d, %d, %d)\n", \
+        tex, offset, rd, wd, flags); \
+    assert(tex); \
+    cs_winsys->write_cs_dword(cs_winsys, offset); \
+    r300_texture_write_reloc(cs_winsys, tex, rd, wd, flags);	\
+    cs_count -= 3; \
+} while (0)
+
+
+#define OUT_CS_BUF_RELOC_NO_OFFSET(bo, rd, wd, flags) do { \
     DBG(cs_context_copy, DBG_CS, "r300: writing relocation for buffer %p, " \
             "domains (%d, %d, %d)\n", \
         bo, rd, wd, flags); \
     assert(bo); \
-    cs_winsys->write_cs_reloc(cs_winsys, bo, rd, wd, flags); \
+    r300_buffer_write_reloc(cs_winsys, r300_buffer(bo), rd, wd, flags);	\
     cs_count -= 2; \
 } while (0)
 
diff --git a/src/gallium/drivers/r300/r300_emit.c b/src/gallium/drivers/r300/r300_emit.c
index f7dcd8dc52..6fb1d7bd92 100644
--- a/src/gallium/drivers/r300/r300_emit.c
+++ b/src/gallium/drivers/r300/r300_emit.c
@@ -32,6 +32,8 @@
 #include "r300_emit.h"
 #include "r300_fs.h"
 #include "r300_screen.h"
+#include "r300_screen_buffer.h"
+#include "r300_state_inlines.h"
 #include "r300_vs.h"
 
 void r300_emit_blend_state(struct r300_context* r300, void* state)
@@ -418,10 +420,10 @@ void r300_emit_fb_state(struct r300_context* r300, void* state)
         assert(tex && tex->buffer && "cbuf is marked, but NULL!");
 
         OUT_CS_REG_SEQ(R300_RB3D_COLOROFFSET0 + (4 * i), 1);
-        OUT_CS_RELOC(tex->buffer, surf->offset, 0, RADEON_GEM_DOMAIN_VRAM, 0);
+        OUT_CS_TEX_RELOC(tex, surf->offset, 0, RADEON_GEM_DOMAIN_VRAM, 0);
 
         OUT_CS_REG_SEQ(R300_RB3D_COLORPITCH0 + (4 * i), 1);
-        OUT_CS_RELOC(tex->buffer, tex->fb_state.colorpitch[surf->level],
+        OUT_CS_TEX_RELOC(tex, tex->fb_state.colorpitch[surf->level],
                      0, RADEON_GEM_DOMAIN_VRAM, 0);
 
         OUT_CS_REG(R300_US_OUT_FMT_0 + (4 * i), tex->fb_state.us_out_fmt);
@@ -434,12 +436,12 @@ void r300_emit_fb_state(struct r300_context* r300, void* state)
         assert(tex && tex->buffer && "zsbuf is marked, but NULL!");
 
         OUT_CS_REG_SEQ(R300_ZB_DEPTHOFFSET, 1);
-        OUT_CS_RELOC(tex->buffer, surf->offset, 0, RADEON_GEM_DOMAIN_VRAM, 0);
+        OUT_CS_TEX_RELOC(tex, surf->offset, 0, RADEON_GEM_DOMAIN_VRAM, 0);
 
         OUT_CS_REG(R300_ZB_FORMAT, tex->fb_state.zb_format);
 
         OUT_CS_REG_SEQ(R300_ZB_DEPTHPITCH, 1);
-        OUT_CS_RELOC(tex->buffer, tex->fb_state.depthpitch[surf->level],
+        OUT_CS_TEX_RELOC(tex, tex->fb_state.depthpitch[surf->level],
                      0, RADEON_GEM_DOMAIN_VRAM, 0);
     }
 
@@ -489,13 +491,13 @@ static void r300_emit_query_finish(struct r300_context *r300,
             /* pipe 3 only */
             OUT_CS_REG(R300_SU_REG_DEST, 1 << 3);
             OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
-            OUT_CS_RELOC(r300->oqbo, query->offset + (sizeof(uint32_t) * 3),
+            OUT_CS_BUF_RELOC(r300->oqbo, query->offset + (sizeof(uint32_t) * 3),
                     0, RADEON_GEM_DOMAIN_GTT, 0);
         case 3:
             /* pipe 2 only */
             OUT_CS_REG(R300_SU_REG_DEST, 1 << 2);
             OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
-            OUT_CS_RELOC(r300->oqbo, query->offset + (sizeof(uint32_t) * 2),
+            OUT_CS_BUF_RELOC(r300->oqbo, query->offset + (sizeof(uint32_t) * 2),
                     0, RADEON_GEM_DOMAIN_GTT, 0);
         case 2:
             /* pipe 1 only */
@@ -503,13 +505,13 @@ static void r300_emit_query_finish(struct r300_context *r300,
             OUT_CS_REG(R300_SU_REG_DEST,
                     1 << (caps->high_second_pipe ? 3 : 1));
             OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
-            OUT_CS_RELOC(r300->oqbo, query->offset + (sizeof(uint32_t) * 1),
+            OUT_CS_BUF_RELOC(r300->oqbo, query->offset + (sizeof(uint32_t) * 1),
                     0, RADEON_GEM_DOMAIN_GTT, 0);
         case 1:
             /* pipe 0 only */
             OUT_CS_REG(R300_SU_REG_DEST, 1 << 0);
             OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
-            OUT_CS_RELOC(r300->oqbo, query->offset + (sizeof(uint32_t) * 0),
+            OUT_CS_BUF_RELOC(r300->oqbo, query->offset + (sizeof(uint32_t) * 0),
                     0, RADEON_GEM_DOMAIN_GTT, 0);
             break;
         default:
@@ -531,7 +533,7 @@ static void rv530_emit_query_single(struct r300_context *r300,
     BEGIN_CS(8);
     OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_0);
     OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
-    OUT_CS_RELOC(r300->oqbo, query->offset, 0, RADEON_GEM_DOMAIN_GTT, 0);
+    OUT_CS_BUF_RELOC(r300->oqbo, query->offset, 0, RADEON_GEM_DOMAIN_GTT, 0);
     OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_ALL);
     END_CS;
 }
@@ -544,10 +546,10 @@ static void rv530_emit_query_double(struct r300_context *r300,
     BEGIN_CS(14);
     OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_0);
     OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
-    OUT_CS_RELOC(r300->oqbo, query->offset, 0, RADEON_GEM_DOMAIN_GTT, 0);
+    OUT_CS_BUF_RELOC(r300->oqbo, query->offset, 0, RADEON_GEM_DOMAIN_GTT, 0);
     OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_1);
     OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
-    OUT_CS_RELOC(r300->oqbo, query->offset + sizeof(uint32_t), 0, RADEON_GEM_DOMAIN_GTT, 0);
+    OUT_CS_BUF_RELOC(r300->oqbo, query->offset + sizeof(uint32_t), 0, RADEON_GEM_DOMAIN_GTT, 0);
     OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_ALL);
     END_CS;
 }
@@ -759,7 +761,7 @@ void r300_emit_texture(struct r300_context* r300,
     OUT_CS_REG(R300_TX_FORMAT1_0 + (offset * 4), tex->state.format1);
     OUT_CS_REG(R300_TX_FORMAT2_0 + (offset * 4), tex->state.format2);
     OUT_CS_REG_SEQ(R300_TX_OFFSET_0 + (offset * 4), 1);
-    OUT_CS_RELOC(tex->buffer,
+    OUT_CS_TEX_RELOC(tex,
                  R300_TXO_MACRO_TILE(tex->macrotile) |
                  R300_TXO_MICRO_TILE(tex->microtile),
                  RADEON_GEM_DOMAIN_GTT | RADEON_GEM_DOMAIN_VRAM, 0, 0);
@@ -800,7 +802,7 @@ void r300_emit_aos(struct r300_context* r300, unsigned offset)
     }
 
     for (i = 0; i < aos_count; i++) {
-        OUT_CS_RELOC_NO_OFFSET(vbuf[velem[i].vertex_buffer_index].buffer,
+        OUT_CS_BUF_RELOC_NO_OFFSET(vbuf[velem[i].vertex_buffer_index].buffer,
                                RADEON_GEM_DOMAIN_GTT, 0, 0);
     }
     END_CS;
@@ -1012,15 +1014,15 @@ void r300_emit_buffer_validate(struct r300_context *r300)
     boolean invalid = FALSE;
 
     /* Clean out BOs. */
-    r300->winsys->reset_bos(r300->winsys);
+    r300->rws->reset_bos(r300->rws);
 
 validate:
     /* Color buffers... */
     for (i = 0; i < fb->nr_cbufs; i++) {
         tex = (struct r300_texture*)fb->cbufs[i]->texture;
         assert(tex && tex->buffer && "cbuf is marked, but NULL!");
-        if (!r300->winsys->add_buffer(r300->winsys, tex->buffer,
-                    0, RADEON_GEM_DOMAIN_VRAM)) {
+        if (!r300_add_texture(r300->rws, tex,
+			      0, RADEON_GEM_DOMAIN_VRAM)) {
             r300->context.flush(&r300->context, 0, NULL);
             goto validate;
         }
@@ -1029,8 +1031,8 @@ validate:
     if (fb->zsbuf) {
         tex = (struct r300_texture*)fb->zsbuf->texture;
         assert(tex && tex->buffer && "zsbuf is marked, but NULL!");
-        if (!r300->winsys->add_buffer(r300->winsys, tex->buffer,
-                    0, RADEON_GEM_DOMAIN_VRAM)) {
+        if (!r300_add_texture(r300->rws, tex,
+			      0, RADEON_GEM_DOMAIN_VRAM)) {
             r300->context.flush(&r300->context, 0, NULL);
             goto validate;
         }
@@ -1040,31 +1042,31 @@ validate:
         tex = r300->textures[i];
         if (!tex)
             continue;
-        if (!r300->winsys->add_buffer(r300->winsys, tex->buffer,
-                    RADEON_GEM_DOMAIN_GTT | RADEON_GEM_DOMAIN_VRAM, 0)) {
+        if (!r300_add_texture(r300->rws, tex,
+			      RADEON_GEM_DOMAIN_GTT | RADEON_GEM_DOMAIN_VRAM, 0)) {
             r300->context.flush(&r300->context, 0, NULL);
             goto validate;
         }
     }
     /* ...occlusion query buffer... */
     if (r300->dirty_state & R300_NEW_QUERY) {
-        if (!r300->winsys->add_buffer(r300->winsys, r300->oqbo,
-                    0, RADEON_GEM_DOMAIN_GTT)) {
+        if (!r300_add_buffer(r300->rws, r300->oqbo,
+			     0, RADEON_GEM_DOMAIN_GTT)) {
             r300->context.flush(&r300->context, 0, NULL);
             goto validate;
         }
     }
     /* ...and vertex buffer. */
     if (r300->vbo) {
-        if (!r300->winsys->add_buffer(r300->winsys, r300->vbo,
-                    RADEON_GEM_DOMAIN_GTT, 0)) {
+        if (!r300_add_buffer(r300->rws, r300->vbo,
+			     RADEON_GEM_DOMAIN_GTT, 0)) {
             r300->context.flush(&r300->context, 0, NULL);
             goto validate;
         }
     } else {
         /* debug_printf("No VBO while emitting dirty state!\n"); */
     }
-    if (!r300->winsys->validate(r300->winsys)) {
+    if (!r300->rws->validate(r300->rws)) {
         r300->context.flush(&r300->context, 0, NULL);
         if (invalid) {
             /* Well, hell. */
@@ -1096,7 +1098,7 @@ void r300_emit_dirty_state(struct r300_context* r300)
     /* Make sure we have at least 2*1024 spare dwords. */
     /* XXX It would be nice to know the number of dwords we really need to
      * XXX emit. */
-    while (!r300->winsys->check_cs(r300->winsys, dwords)) {
+    while (!r300->rws->check_cs(r300->rws, dwords)) {
         r300->context.flush(&r300->context, 0, NULL);
     }
 
diff --git a/src/gallium/drivers/r300/r300_render.c b/src/gallium/drivers/r300/r300_render.c
index 754eb4dc76..47f5d6f7f1 100644
--- a/src/gallium/drivers/r300/r300_render.c
+++ b/src/gallium/drivers/r300/r300_render.c
@@ -30,10 +30,12 @@
 
 #include "util/u_format.h"
 #include "util/u_memory.h"
+#include "util/u_upload_mgr.h"
 #include "util/u_prim.h"
 
 #include "r300_cs.h"
 #include "r300_context.h"
+#include "r300_screen_buffer.h"
 #include "r300_emit.h"
 #include "r300_reg.h"
 #include "r300_render.h"
@@ -297,7 +299,7 @@ static void r300_emit_draw_elements(struct r300_context *r300,
     OUT_CS(R300_INDX_BUFFER_ONE_REG_WR | (R300_VAP_PORT_IDX0 >> 2) |
            (0 << R300_INDX_BUFFER_SKIP_SHIFT));
     OUT_CS(offset_dwords << 2);
-    OUT_CS_RELOC(indexBuffer, count_dwords,
+    OUT_CS_BUF_RELOC(indexBuffer, count_dwords,
         RADEON_GEM_DOMAIN_GTT, 0, 0);
 
     END_CS;
@@ -308,21 +310,28 @@ static boolean r300_setup_vertex_buffers(struct r300_context *r300)
     struct pipe_vertex_buffer *vbuf = r300->vertex_buffer;
     struct pipe_vertex_element *velem = r300->vertex_element;
     struct pipe_buffer *pbuf;
+    int ret;
+    
+    /* upload buffers first */
+    if (r300->any_user_vbs) {
+        ret = r300_upload_user_buffers(r300);
+        r300->any_user_vbs = false;
+    }
 
 validate:
     for (int i = 0; i < r300->vertex_element_count; i++) {
         pbuf = vbuf[velem[i].vertex_buffer_index].buffer;
 
-        if (!r300->winsys->add_buffer(r300->winsys, pbuf,
-                                      RADEON_GEM_DOMAIN_GTT, 0)) {
+        if (!r300_add_buffer(r300->rws, pbuf,
+			     RADEON_GEM_DOMAIN_GTT, 0)) {
             r300->context.flush(&r300->context, 0, NULL);
             goto validate;
         }
     }
 
-    if (!r300->winsys->validate(r300->winsys)) {
+    if (!r300->rws->validate(r300->rws)) {
         r300->context.flush(&r300->context, 0, NULL);
-        return r300->winsys->validate(r300->winsys);
+        return r300->rws->validate(r300->rws);
     }
 
     return TRUE;
@@ -396,15 +405,20 @@ void r300_draw_range_elements(struct pipe_context* pipe,
         indexSize = 2;
     }
 
-    if (!r300->winsys->add_buffer(r300->winsys, indexBuffer,
-                                  RADEON_GEM_DOMAIN_GTT, 0)) {
+    r300_upload_index_buffer(r300, &indexBuffer,
+			     indexSize, start, count);
+    
+    if (!r300_add_buffer(r300->rws, indexBuffer,
+			 RADEON_GEM_DOMAIN_GTT, 0)) {
         goto cleanup;
     }
 
-    if (!r300->winsys->validate(r300->winsys)) {
+    if (!r300->rws->validate(r300->rws)) {
         goto cleanup;
     }
 
+    u_upload_flush(r300->upload_vb);
+    u_upload_flush(r300->upload_ib);
     r300_emit_dirty_state(r300);
 
     r300_emit_aos(r300, 0);
@@ -425,7 +439,7 @@ void r300_draw_range_elements(struct pipe_context* pipe,
 
 cleanup:
     if (indexBuffer != orgIndexBuffer) {
-        pipe->screen->buffer_destroy(indexBuffer);
+        pipe_buffer_reference( &indexBuffer, NULL );
     }
 }
 
@@ -466,6 +480,7 @@ void r300_draw_arrays(struct pipe_context* pipe, unsigned mode,
             return;
         }
 
+    	u_upload_flush(r300->upload_vb);
         r300_emit_dirty_state(r300);
 
         if (alt_num_verts || count <= 65535) {
diff --git a/src/gallium/drivers/r300/r300_screen.c b/src/gallium/drivers/r300/r300_screen.c
index 6a55570571..3b70312c82 100644
--- a/src/gallium/drivers/r300/r300_screen.c
+++ b/src/gallium/drivers/r300/r300_screen.c
@@ -23,7 +23,6 @@
 #include "util/u_inlines.h"
 #include "util/u_format.h"
 #include "util/u_memory.h"
-#include "util/u_simple_screen.h"
 
 #include "r300_context.h"
 #include "r300_screen.h"
@@ -32,6 +31,8 @@
 #include "radeon_winsys.h"
 #include "r300_winsys.h"
 
+#include "r300_screen_buffer.h"
+
 /* Return the identifier behind whom the brave coders responsible for this
  * amalgamation of code, sweat, and duct tape, routinely obscure their names.
  *
@@ -292,10 +293,11 @@ static void* r300_transfer_map(struct pipe_screen* screen,
                               struct pipe_transfer* transfer)
 {
     struct r300_texture* tex = (struct r300_texture*)transfer->texture;
+    struct r300_winsys_screen *rws = r300_winsys_screen(screen);
     char* map;
     enum pipe_format format = tex->tex.format;
 
-    map = pipe_buffer_map(screen, tex->buffer,
+    map = rws->buffer_map(rws, tex->buffer,
                           pipe_transfer_buffer_flags(transfer));
 
     if (!map) {
@@ -311,21 +313,26 @@ static void r300_transfer_unmap(struct pipe_screen* screen,
                                 struct pipe_transfer* transfer)
 {
     struct r300_texture* tex = (struct r300_texture*)transfer->texture;
-    pipe_buffer_unmap(screen, tex->buffer);
+    struct r300_winsys_screen *rws = r300_winsys_screen(screen);
+    rws->buffer_unmap(rws, tex->buffer);
 }
 
 static void r300_destroy_screen(struct pipe_screen* pscreen)
 {
     struct r300_screen* r300screen = r300_screen(pscreen);
+    struct r300_winsys_screen *rws = r300_winsys_screen(pscreen);
+
+    if (rws)
+      rws->destroy(rws);
 
     FREE(r300screen->caps);
     FREE(r300screen);
 }
 
-struct pipe_screen* r300_create_screen(struct radeon_winsys* radeon_winsys)
+struct pipe_screen* r300_create_screen(struct r300_winsys_screen *rws)
 {
-    struct r300_screen* r300screen = CALLOC_STRUCT(r300_screen);
-    struct r300_capabilities* caps = CALLOC_STRUCT(r300_capabilities);
+    struct r300_screen *r300screen = CALLOC_STRUCT(r300_screen);
+    struct r300_capabilities *caps = CALLOC_STRUCT(r300_capabilities);
 
     if (!r300screen || !caps) {
         FREE(r300screen);
@@ -333,16 +340,16 @@ struct pipe_screen* r300_create_screen(struct radeon_winsys* radeon_winsys)
         return NULL;
     }
 
-    caps->pci_id = radeon_winsys->pci_id;
-    caps->num_frag_pipes = radeon_winsys->gb_pipes;
-    caps->num_z_pipes = radeon_winsys->z_pipes;
+    caps->pci_id = rws->get_value(rws, R300_VID_PCI_ID);
+    caps->num_frag_pipes = rws->get_value(rws, R300_VID_GB_PIPES);
+    caps->num_z_pipes = rws->get_value(rws, R300_VID_Z_PIPES);
 
     r300_init_debug(r300screen);
     r300_parse_chipset(caps);
 
     r300screen->caps = caps;
-    r300screen->radeon_winsys = radeon_winsys;
-    r300screen->screen.winsys = (struct pipe_winsys*)radeon_winsys;
+    r300screen->rws = rws;
+    r300screen->screen.winsys = (struct pipe_winsys*)rws;
     r300screen->screen.destroy = r300_destroy_screen;
     r300screen->screen.get_name = r300_get_name;
     r300screen->screen.get_vendor = r300_get_vendor;
@@ -356,7 +363,12 @@ struct pipe_screen* r300_create_screen(struct radeon_winsys* radeon_winsys)
     r300screen->screen.transfer_unmap = r300_transfer_unmap;
 
     r300_init_screen_texture_functions(&r300screen->screen);
-    u_simple_screen_init(&r300screen->screen);
-
+    r300_screen_init_buffer_functions(r300screen);
     return &r300screen->screen;
 }
+
+struct r300_winsys_screen *
+r300_winsys_screen(struct pipe_screen *screen)
+{
+    return r300_screen(screen)->rws;
+}
diff --git a/src/gallium/drivers/r300/r300_screen.h b/src/gallium/drivers/r300/r300_screen.h
index 502fbfa5a2..87fc6190ce 100644
--- a/src/gallium/drivers/r300/r300_screen.h
+++ b/src/gallium/drivers/r300/r300_screen.h
@@ -33,7 +33,7 @@ struct r300_screen {
     /* Parent class */
     struct pipe_screen screen;
 
-    struct radeon_winsys* radeon_winsys;
+    struct r300_winsys_screen *rws;
 
     /* Chipset capabilities */
     struct r300_capabilities* caps;
diff --git a/src/gallium/drivers/r300/r300_screen_buffer.c b/src/gallium/drivers/r300/r300_screen_buffer.c
new file mode 100644
index 0000000000..358582ea21
--- /dev/null
+++ b/src/gallium/drivers/r300/r300_screen_buffer.c
@@ -0,0 +1,222 @@
+#include <stdio.h>
+
+#include "util/u_inlines.h"
+#include "util/u_format.h"
+#include "util/u_memory.h"
+#include "util/u_upload_mgr.h"
+
+#include "r300_screen_buffer.h"
+
+#include "r300_winsys.h"
+
+int r300_upload_index_buffer(struct r300_context *r300,
+			     struct pipe_buffer **index_buffer,
+			     unsigned index_size,
+			     unsigned start,
+			     unsigned count)
+{
+   struct pipe_buffer *upload_buffer = NULL;
+   unsigned index_offset = start * index_size;
+   int ret = 0;
+
+    if (r300_buffer_is_user_buffer(*index_buffer)) {
+	ret = u_upload_buffer(r300->upload_ib,
+			      index_offset,
+			      count * index_size,
+			      *index_buffer,
+			      &index_offset,
+			      &upload_buffer);
+	if (ret) {
+	    goto done;
+	}
+	*index_buffer = upload_buffer;
+    }
+ done:
+    //    if (upload_buffer)
+    //	pipe_buffer_reference(&upload_buffer, NULL);
+    return ret;
+}
+
+int r300_upload_user_buffers(struct r300_context *r300)
+{
+    enum pipe_error ret = PIPE_OK;
+    int i, nr;
+
+    nr = r300->vertex_buffer_count;
+
+    for (i = 0; i < nr; i++) {
+
+	if (r300_buffer_is_user_buffer(r300->vertex_buffer[i].buffer)) {
+	    struct pipe_buffer *upload_buffer = NULL;
+	    unsigned offset = 0;
+	    unsigned size = r300->vertex_buffer[i].buffer->size;
+	    unsigned upload_offset;
+
+	    ret = u_upload_buffer(r300->upload_vb,
+				  offset, size,
+				  r300->vertex_buffer[i].buffer,
+				  &upload_offset, &upload_buffer);
+	    if (ret)
+		return ret;
+
+	    pipe_buffer_reference(&r300->vertex_buffer[i].buffer, NULL);
+	    r300->vertex_buffer[i].buffer = upload_buffer;
+	    r300->vertex_buffer[i].buffer_offset = upload_offset;
+	}
+    }
+    return ret;
+}
+
+static struct r300_winsys_buffer *
+r300_winsys_buffer_create(struct r300_screen *r300screen,
+			  unsigned alignment,
+			  unsigned usage,
+			  unsigned size)
+{
+    struct r300_winsys_screen *rws = r300screen->rws;
+    struct r300_winsys_buffer *buf;
+
+    buf = rws->buffer_create(rws, alignment, usage, size);
+    return buf;
+}
+
+static void r300_winsys_buffer_destroy(struct r300_screen *r300screen,
+				       struct r300_buffer *rbuf)
+{
+    struct r300_winsys_screen *rws = r300screen->rws;
+
+    if (rbuf->buf) {
+	rws->buffer_destroy(rbuf->buf);
+	rbuf->buf = NULL;
+    }
+}
+
+static struct pipe_buffer *r300_buffer_create(struct pipe_screen *screen,
+					      unsigned alignment,
+					      unsigned usage,
+					      unsigned size)
+{
+    struct r300_screen *r300screen = r300_screen(screen);
+    struct r300_buffer *rbuf;
+
+    rbuf = CALLOC_STRUCT(r300_buffer);
+    if (!rbuf)
+	goto error1;
+
+    rbuf->magic = R300_BUFFER_MAGIC;
+
+    pipe_reference_init(&rbuf->base.reference, 1);
+    rbuf->base.screen = screen;
+    rbuf->base.alignment = alignment;
+    rbuf->base.usage = usage;
+    rbuf->base.size = size;
+
+    rbuf->buf = r300_winsys_buffer_create(r300screen,
+					  alignment,
+					  usage,
+					  size);
+
+    if (!rbuf->buf)
+	goto error2;
+
+    return &rbuf->base;
+error2:
+    FREE(rbuf);
+error1:
+    return NULL;
+}
+
+
+static struct pipe_buffer *r300_user_buffer_create(struct pipe_screen *screen,
+						   void *ptr,
+						   unsigned bytes)
+{
+    struct r300_buffer *rbuf;
+
+    rbuf = CALLOC_STRUCT(r300_buffer);
+    if (!rbuf)
+	goto no_rbuf;
+
+    rbuf->magic = R300_BUFFER_MAGIC;
+
+    pipe_reference_init(&rbuf->base.reference, 1);
+    rbuf->base.screen = screen;
+    rbuf->base.alignment = 1;
+    rbuf->base.usage = 0;
+    rbuf->base.size = bytes;
+
+    rbuf->user_buffer = ptr;
+    return &rbuf->base;
+
+no_rbuf:
+    return NULL;
+}
+
+static void r300_buffer_destroy(struct pipe_buffer *buf)
+{
+    struct r300_screen *r300screen = r300_screen(buf->screen);
+    struct r300_buffer *rbuf = r300_buffer(buf);
+
+    r300_winsys_buffer_destroy(r300screen, rbuf);
+    FREE(rbuf);
+}
+
+static void *
+r300_buffer_map_range(struct pipe_screen *screen,
+		      struct pipe_buffer *buf,
+		      unsigned offset, unsigned length,
+		      unsigned usage )
+{
+    struct r300_screen *r300screen = r300_screen(screen);
+    struct r300_winsys_screen *rws = r300screen->rws;
+    struct r300_buffer *rbuf = r300_buffer(buf);
+    void *map;
+
+   if (rbuf->user_buffer)
+      return rbuf->user_buffer;
+
+    map = rws->buffer_map(rws, rbuf->buf, usage);
+
+    return map;
+}
+
+static void *
+r300_buffer_map(struct pipe_screen *screen,
+		struct pipe_buffer *buf,
+		unsigned usage)
+{
+    struct r300_screen *r300screen = r300_screen(screen);
+    struct r300_winsys_screen *rws = r300screen->rws;
+    struct r300_buffer *rbuf = r300_buffer(buf);
+    void *map;
+
+   if (rbuf->user_buffer)
+      return rbuf->user_buffer;
+
+    map = rws->buffer_map(rws, rbuf->buf, usage);
+
+    return map;
+}
+
+static void
+r300_buffer_unmap(struct pipe_screen *screen,
+		  struct pipe_buffer *buf)
+{
+    struct r300_screen *r300screen = r300_screen(screen);
+    struct r300_winsys_screen *rws = r300screen->rws;
+    struct r300_buffer *rbuf = r300_buffer(buf);
+
+    if (rbuf->buf)
+      rws->buffer_unmap(rws, rbuf->buf);
+}
+
+void r300_screen_init_buffer_functions(struct r300_screen *r300screen)
+{
+    r300screen->screen.buffer_create = r300_buffer_create;
+    r300screen->screen.user_buffer_create = r300_user_buffer_create;
+    r300screen->screen.buffer_map = r300_buffer_map;
+    r300screen->screen.buffer_map_range = r300_buffer_map_range;
+//    r300screen->screen.buffer_flush_mapped_range = r300_buffer_flush_mapped_range;
+    r300screen->screen.buffer_unmap = r300_buffer_unmap;
+    r300screen->screen.buffer_destroy = r300_buffer_destroy;
+}
diff --git a/src/gallium/drivers/r300/r300_screen_buffer.h b/src/gallium/drivers/r300/r300_screen_buffer.h
new file mode 100644
index 0000000000..7c026508b5
--- /dev/null
+++ b/src/gallium/drivers/r300/r300_screen_buffer.h
@@ -0,0 +1,85 @@
+#ifndef R300_SCREEN_BUFFER_H
+#define R300_SCREEN_BUFFER_H
+#include "pipe/p_compiler.h"
+#include "pipe/p_state.h"
+#include "r300_screen.h"
+
+#include "r300_winsys.h"
+#include "r300_context.h"
+
+#define R300_BUFFER_MAGIC 0xabcd1234
+
+struct r300_buffer
+{
+    struct pipe_buffer base;
+
+    uint32_t magic;
+
+    struct r300_winsys_buffer *buf;
+
+    void *user_buffer;
+};
+
+static INLINE struct r300_buffer *
+r300_buffer(struct pipe_buffer *buffer)
+{
+    if (buffer) {
+	assert(((struct r300_buffer *)buffer)->magic == R300_BUFFER_MAGIC);
+	return (struct r300_buffer *)buffer;
+    }
+    return NULL;
+}
+
+static INLINE boolean 
+r300_buffer_is_user_buffer(struct pipe_buffer *buffer)
+{
+    return r300_buffer(buffer)->user_buffer ? true : false;
+}
+
+static INLINE boolean r300_add_buffer(struct r300_winsys_screen *rws,
+				      struct pipe_buffer *buffer,
+				      int rd, int wr)
+{
+    struct r300_buffer *buf = r300_buffer(buffer);
+
+    if (!buf->buf)
+	return true;
+
+    return rws->add_buffer(rws, buf->buf, rd, wr);
+}
+
+
+static INLINE boolean r300_add_texture(struct r300_winsys_screen *rws,
+				       struct r300_texture *tex,
+				       int rd, int wr)
+{
+    return rws->add_buffer(rws, tex->buffer, rd, wr);
+}
+
+void r300_screen_init_buffer_functions(struct r300_screen *r300screen);
+
+static INLINE void r300_buffer_write_reloc(struct r300_winsys_screen *rws,
+				      struct r300_buffer *buf,
+				      uint32_t rd, uint32_t wd, uint32_t flags)
+{
+    if (!buf->buf)
+	return;
+
+    rws->write_cs_reloc(rws, buf->buf, rd, wd, flags);
+}
+
+static INLINE void r300_texture_write_reloc(struct r300_winsys_screen *rws,
+					    struct r300_texture *texture,
+					    uint32_t rd, uint32_t wd, uint32_t flags)
+{
+    rws->write_cs_reloc(rws, texture->buffer, rd, wd, flags);
+}
+
+int r300_upload_user_buffers(struct r300_context *r300);
+
+int r300_upload_index_buffer(struct r300_context *r300,
+			     struct pipe_buffer **index_buffer,
+			     unsigned index_size,
+			     unsigned start,
+			     unsigned count);
+#endif
diff --git a/src/gallium/drivers/r300/r300_state.c b/src/gallium/drivers/r300/r300_state.c
index 34bf81c193..d8a87dec15 100644
--- a/src/gallium/drivers/r300/r300_state.c
+++ b/src/gallium/drivers/r300/r300_state.c
@@ -34,6 +34,7 @@
 #include "r300_context.h"
 #include "r300_reg.h"
 #include "r300_screen.h"
+#include "r300_screen_buffer.h"
 #include "r300_state_inlines.h"
 #include "r300_fs.h"
 #include "r300_vs.h"
@@ -521,7 +522,7 @@ static void r300_fb_update_tiling_flags(struct r300_context *r300,
         tex = (struct r300_texture*)old_state->cbufs[i]->texture;
 
         if (tex) {
-            r300->winsys->buffer_set_tiling(r300->winsys, tex->buffer,
+            r300->rws->buffer_set_tiling(r300->rws, tex->buffer,
                                             tex->pitch[0],
                                             tex->microtile != 0,
                                             tex->macrotile != 0);
@@ -533,7 +534,7 @@ static void r300_fb_update_tiling_flags(struct r300_context *r300,
         tex = (struct r300_texture*)old_state->zsbuf->texture;
 
         if (tex) {
-            r300->winsys->buffer_set_tiling(r300->winsys, tex->buffer,
+            r300->rws->buffer_set_tiling(r300->rws, tex->buffer,
                                             tex->pitch[0],
                                             tex->microtile != 0,
                                             tex->macrotile != 0);
@@ -545,7 +546,7 @@ static void r300_fb_update_tiling_flags(struct r300_context *r300,
         tex = (struct r300_texture*)new_state->cbufs[i]->texture;
         level = new_state->cbufs[i]->level;
 
-        r300->winsys->buffer_set_tiling(r300->winsys, tex->buffer,
+        r300->rws->buffer_set_tiling(r300->rws, tex->buffer,
                                         tex->pitch[level],
                                         tex->microtile != 0,
                                         tex->mip_macrotile[level] != 0);
@@ -554,7 +555,7 @@ static void r300_fb_update_tiling_flags(struct r300_context *r300,
         tex = (struct r300_texture*)new_state->zsbuf->texture;
         level = new_state->zsbuf->level;
 
-        r300->winsys->buffer_set_tiling(r300->winsys, tex->buffer,
+        r300->rws->buffer_set_tiling(r300->rws, tex->buffer,
                                         tex->pitch[level],
                                         tex->microtile != 0,
                                         tex->mip_macrotile[level] != 0);
@@ -1028,10 +1029,26 @@ static void r300_set_vertex_buffers(struct pipe_context* pipe,
                                     const struct pipe_vertex_buffer* buffers)
 {
     struct r300_context* r300 = r300_context(pipe);
+    boolean any_user_buffer = false;
+    int i;
+
+    if (count == r300->vertex_buffer_count &&
+	memcmp(r300->vertex_buffer, buffers, count * sizeof(buffers[0])) == 0)
+        return;
+
+    for (i = 0; i < count; i++) {
+	pipe_buffer_reference(&r300->vertex_buffer[i].buffer, buffers[i].buffer);
+	if (r300_buffer_is_user_buffer(buffers[i].buffer))
+	    any_user_buffer = true;
+    }
+
+    for ( ; i < r300->vertex_buffer_count; i++)
+	pipe_buffer_reference(&r300->vertex_buffer[i].buffer, NULL);
 
     memcpy(r300->vertex_buffer, buffers,
         sizeof(struct pipe_vertex_buffer) * count);
     r300->vertex_buffer_count = count;
+    r300->any_user_vbs = any_user_buffer;
 
     if (r300->draw) {
         draw_flush(r300->draw);
diff --git a/src/gallium/drivers/r300/r300_texture.c b/src/gallium/drivers/r300/r300_texture.c
index ed2be06254..9016e86ccd 100644
--- a/src/gallium/drivers/r300/r300_texture.c
+++ b/src/gallium/drivers/r300/r300_texture.c
@@ -731,7 +731,7 @@ static struct pipe_texture*
 {
     struct r300_texture* tex = CALLOC_STRUCT(r300_texture);
     struct r300_screen* rscreen = r300_screen(screen);
-    struct radeon_winsys* winsys = (struct radeon_winsys*)screen->winsys;
+    struct r300_winsys_screen *rws = (struct r300_winsys_screen *)screen->winsys;
 
     if (!tex) {
         return NULL;
@@ -745,13 +745,13 @@ static struct pipe_texture*
     r300_setup_miptree(rscreen, tex);
     r300_setup_texture_state(rscreen, tex);
 
-    tex->buffer = screen->buffer_create(screen, 2048,
-                                        PIPE_BUFFER_USAGE_PIXEL,
-                                        tex->size);
-    winsys->buffer_set_tiling(winsys, tex->buffer,
-                              tex->pitch[0],
-                              tex->microtile != R300_BUFFER_LINEAR,
-                              tex->macrotile != R300_BUFFER_LINEAR);
+    tex->buffer = rws->buffer_create(rws, 2048,
+				     PIPE_BUFFER_USAGE_PIXEL,
+				     tex->size);
+    rws->buffer_set_tiling(rws, tex->buffer,
+			   tex->pitch[0],
+			   tex->microtile != R300_BUFFER_LINEAR,
+			   tex->macrotile != R300_BUFFER_LINEAR);
 
     if (!tex->buffer) {
         FREE(tex);
@@ -764,9 +764,9 @@ static struct pipe_texture*
 static void r300_texture_destroy(struct pipe_texture* texture)
 {
     struct r300_texture* tex = (struct r300_texture*)texture;
+    struct r300_winsys_screen *rws = (struct r300_winsys_screen *)texture->screen->winsys;
 
-    pipe_buffer_reference(&tex->buffer, NULL);
-
+    rws->buffer_reference(rws, &tex->buffer, NULL);
     FREE(tex);
 }
 
@@ -806,14 +806,14 @@ static void r300_tex_surface_destroy(struct pipe_surface* s)
     FREE(s);
 }
 
-static struct pipe_texture*
-    r300_texture_blanket(struct pipe_screen* screen,
-                         const struct pipe_texture* base,
-                         const unsigned* stride,
-                         struct pipe_buffer* buffer)
+struct pipe_texture *r300_texture_blanket_winsys_buffer(struct pipe_screen *screen,
+							const struct pipe_texture *base,
+							const unsigned *stride,
+							struct r300_winsys_buffer *buffer)
 {
     struct r300_texture* tex;
     struct r300_screen* rscreen = r300_screen(screen);
+    struct r300_winsys_screen *rws = (struct r300_winsys_screen *)screen->winsys;
 
     /* Support only 2D textures without mipmaps */
     if (base->target != PIPE_TEXTURE_2D ||
@@ -837,7 +837,7 @@ static struct pipe_texture*
     r300_setup_flags(tex);
     r300_setup_texture_state(rscreen, tex);
 
-    pipe_buffer_reference(&tex->buffer, buffer);
+    rws->buffer_reference(rws, &tex->buffer, buffer);
 
     return (struct pipe_texture*)tex;
 }
@@ -896,7 +896,6 @@ void r300_init_screen_texture_functions(struct pipe_screen* screen)
     screen->texture_destroy = r300_texture_destroy;
     screen->get_tex_surface = r300_get_tex_surface;
     screen->tex_surface_destroy = r300_tex_surface_destroy;
-    screen->texture_blanket = r300_texture_blanket;
 
     screen->video_surface_create = r300_video_surface_create;
     screen->video_surface_destroy= r300_video_surface_destroy;
@@ -904,19 +903,23 @@ void r300_init_screen_texture_functions(struct pipe_screen* screen)
 
 boolean r300_get_texture_buffer(struct pipe_screen* screen,
                                 struct pipe_texture* texture,
-                                struct pipe_buffer** buffer,
+                                struct r300_winsys_buffer** buffer,
                                 unsigned* stride)
 {
     struct r300_texture* tex = (struct r300_texture*)texture;
+    struct r300_winsys_screen *rws = (struct r300_winsys_screen *)screen->winsys;
+    struct r300_winsys_buffer *buf;
+
     if (!tex) {
         return FALSE;
     }
 
-    pipe_buffer_reference(buffer, tex->buffer);
+    rws->buffer_reference(rws, &buf, tex->buffer);
 
     if (stride) {
         *stride = r300_texture_get_stride(r300_screen(screen), tex, 0);
     }
 
+    *buffer = buf;
     return TRUE;
 }
diff --git a/src/gallium/drivers/r300/r300_texture.h b/src/gallium/drivers/r300/r300_texture.h
index 46a5fb6188..97724f790c 100644
--- a/src/gallium/drivers/r300/r300_texture.h
+++ b/src/gallium/drivers/r300/r300_texture.h
@@ -63,8 +63,8 @@ r300_video_surface(struct pipe_video_surface *pvs)
 #ifndef R300_WINSYS_H
 
 boolean r300_get_texture_buffer(struct pipe_screen* screen,
-                                struct pipe_texture* texture,
-                                struct pipe_buffer** buffer,
+                                struct pipe_texture *texture,
+                                struct r300_winsys_buffer** buffer,
                                 unsigned* stride);
 
 #endif /* R300_WINSYS_H */
diff --git a/src/gallium/drivers/r300/r300_winsys.h b/src/gallium/drivers/r300/r300_winsys.h
index 40fb8a95ca..c6f9174496 100644
--- a/src/gallium/drivers/r300/r300_winsys.h
+++ b/src/gallium/drivers/r300/r300_winsys.h
@@ -23,10 +23,6 @@
 #ifndef R300_WINSYS_H
 #define R300_WINSYS_H
 
-#ifdef __cplusplus
-extern "C" {
-#endif
-
 /* The public interface header for the r300 pipe driver.
  * Any winsys hosting this pipe needs to implement r300_winsys and then
  * call r300_create_screen to start things. */
@@ -34,19 +30,128 @@ extern "C" {
 #include "pipe/p_defines.h"
 #include "pipe/p_state.h"
 
-struct radeon_winsys;
+struct r300_winsys_screen;
 
 /* Creates a new r300 screen. */
-struct pipe_screen* r300_create_screen(struct radeon_winsys* radeon_winsys);
+struct pipe_screen* r300_create_screen(struct r300_winsys_screen *rws);
+
+struct r300_winsys_buffer;
 
 
 boolean r300_get_texture_buffer(struct pipe_screen* screen,
                                 struct pipe_texture* texture,
-                                struct pipe_buffer** buffer,
-                                unsigned* stride);
+                                struct r300_winsys_buffer** buffer,
+                                unsigned *stride);
+
+enum r300_value_id {
+    R300_VID_PCI_ID,
+    R300_VID_GB_PIPES,
+    R300_VID_Z_PIPES,
+};
+
+struct r300_winsys_screen {
+    void (*destroy)(struct r300_winsys_screen *ws);
+    
+    /**
+     * Buffer management. Buffer attributes are mostly fixed over its lifetime.
+     *
+     * Remember that gallium gets to choose the interface it needs, and the
+     * window systems must then implement that interface (rather than the
+     * other way around...).
+     *
+     * usage is a bitmask of R300_WINSYS_BUFFER_USAGE_PIXEL/VERTEX/INDEX/CONSTANT. This
+     * usage argument is only an optimization hint, not a guarantee, therefore
+     * proper behavior must be observed in all circumstances.
+     *
+     * alignment indicates the client's alignment requirements, eg for
+     * SSE instructions.
+     */
+    struct r300_winsys_buffer *(*buffer_create)(struct r300_winsys_screen *ws,
+						unsigned alignment,
+						unsigned usage,
+						unsigned size);
+    
+    /**
+     * Map the entire data store of a buffer object into the client's address.
+     * flags is bitmask of R300_WINSYS_BUFFER_USAGE_CPU_READ/WRITE flags.
+     */
+    void *(*buffer_map)( struct r300_winsys_screen *ws,
+			 struct r300_winsys_buffer *buf,
+			 unsigned usage );
+
+    void (*buffer_unmap)( struct r300_winsys_screen *ws,
+			  struct r300_winsys_buffer *buf );
+
+    void (*buffer_destroy)( struct r300_winsys_buffer *buf );
+
+
+    void (*buffer_reference)(struct r300_winsys_screen *rws,
+			     struct r300_winsys_buffer **pdst,
+			     struct r300_winsys_buffer *src);
+
+    boolean (*buffer_references)(struct r300_winsys_buffer *a,
+				 struct r300_winsys_buffer *b);
+
+    /* Add a pipe_buffer to the list of buffer objects to validate. */
+    boolean (*add_buffer)(struct r300_winsys_screen *winsys,
+                          struct r300_winsys_buffer *buf,
+                          uint32_t rd,
+                          uint32_t wd);
+
+    /* Revalidate all currently setup pipe_buffers.
+     * Returns TRUE if a flush is required. */
+    boolean (*validate)(struct r300_winsys_screen* winsys);
+
+    /* Check to see if there's room for commands. */
+    boolean (*check_cs)(struct r300_winsys_screen* winsys, int size);
+
+    /* Start a command emit. */
+    void (*begin_cs)(struct r300_winsys_screen* winsys,
+                     int size,
+                     const char* file,
+                     const char* function,
+                     int line);
+
+    /* Write a dword to the command buffer. */
+    void (*write_cs_dword)(struct r300_winsys_screen* winsys, uint32_t dword);
+
+    /* Write a relocated dword to the command buffer. */
+    void (*write_cs_reloc)(struct r300_winsys_screen *winsys,
+                           struct r300_winsys_buffer *buf,
+                           uint32_t rd,
+                           uint32_t wd,
+                           uint32_t flags);
+
+    /* Finish a command emit. */
+    void (*end_cs)(struct r300_winsys_screen* winsys,
+                   const char* file,
+                   const char* function,
+                   int line);
+
+    /* Flush the CS. */
+    void (*flush_cs)(struct r300_winsys_screen* winsys);
+
+    /* winsys flush - callback from winsys when flush required */
+    void (*set_flush_cb)(struct r300_winsys_screen *winsys,
+			 void (*flush_cb)(void *), void *data);
+
+    void (*reset_bos)(struct r300_winsys_screen *winsys);
+
+    void (*buffer_set_tiling)(struct r300_winsys_screen *winsys,
+                              struct r300_winsys_buffer *buffer,
+                              uint32_t pitch,
+                              boolean microtiled,
+                              boolean macrotiled);
+
+    uint32_t (*get_value)(struct r300_winsys_screen *winsys,
+			  enum r300_value_id vid);
+};
 
-#ifdef __cplusplus
-}
-#endif
+struct r300_winsys_screen *
+r300_winsys_screen(struct pipe_screen *screen);
 
+struct pipe_texture *r300_texture_blanket_winsys_buffer(struct pipe_screen *screen,
+							const struct pipe_texture *base,
+							const unsigned *stride,
+							struct r300_winsys_buffer *buffer);
 #endif /* R300_WINSYS_H */
diff --git a/src/gallium/winsys/drm/i965/gem/i965_drm_winsys.h b/src/gallium/winsys/drm/i965/gem/i965_drm_winsys.h
index c6a7d4a8c5..b36680eb66 100644
--- a/src/gallium/winsys/drm/i965/gem/i965_drm_winsys.h
+++ b/src/gallium/winsys/drm/i965/gem/i965_drm_winsys.h
@@ -32,8 +32,6 @@ i965_libdrm_winsys(struct brw_winsys_screen *iws)
    return (struct i965_libdrm_winsys *)iws;
 }
 
-struct i965_libdrm_winsys *i965_libdrm_winsys_create(int fd, unsigned pci_id);
-
 void i965_libdrm_winsys_init_buffer_functions(struct i965_libdrm_winsys *idws);
 
 
diff --git a/src/gallium/winsys/drm/radeon/core/Makefile b/src/gallium/winsys/drm/radeon/core/Makefile
index 860cbb6dbf..13bbbf730d 100644
--- a/src/gallium/winsys/drm/radeon/core/Makefile
+++ b/src/gallium/winsys/drm/radeon/core/Makefile
@@ -5,7 +5,7 @@ include $(TOP)/configs/current
 LIBNAME = radeonwinsys
 
 C_SOURCES = \
-	radeon_buffer.c \
+	radeon_drm_buffer.c \
 	radeon_drm.c \
 	radeon_r300.c
 
diff --git a/src/gallium/winsys/drm/radeon/core/radeon_buffer.h b/src/gallium/winsys/drm/radeon/core/radeon_buffer.h
index f1c8fc2a3b..8e0274e33d 100644
--- a/src/gallium/winsys/drm/radeon/core/radeon_buffer.h
+++ b/src/gallium/winsys/drm/radeon/core/radeon_buffer.h
@@ -32,11 +32,11 @@
 
 #include <stdio.h>
 
-#include "util/u_simple_screen.h"
 #include "pipe/p_defines.h"
 #include "util/u_inlines.h"
 
 #include "pipebuffer/pb_buffer.h"
+#include "pipebuffer/pb_bufmgr.h"
 
 #include "util/u_memory.h"
 
@@ -47,41 +47,41 @@
 
 #include "radeon_winsys.h"
 
-struct radeon_pipe_buffer {
-    struct pipe_buffer  base;
-    /* Pointer to GPU-backed BO. */
-    struct radeon_bo    *bo;
-    /* Pointer to fallback PB buffer. */
-    struct pb_buffer    *pb;
-    boolean flinked;
-    uint32_t flink;
-};
 
 #define RADEON_MAX_BOS 24
 
-struct radeon_winsys_priv {
-    /* DRM FD */
-    int fd;
+static INLINE struct pb_buffer *
+radeon_pb_buffer(struct r300_winsys_buffer *buffer)
+{
+    return (struct pb_buffer *)buffer;
+}
 
-    /* Radeon BO manager. */
-    struct radeon_bo_manager* bom;
+static INLINE struct r300_winsys_buffer *
+radeon_libdrm_winsys_buffer(struct pb_buffer *buffer)
+{
+    return (struct r300_winsys_buffer *)buffer;
+}
 
-    /* Radeon CS manager. */
-    struct radeon_cs_manager* csm;
+struct pb_manager *
+radeon_drm_bufmgr_create(struct radeon_libdrm_winsys *rws);
 
-    /* Current CS. */
-    struct radeon_cs* cs;
+boolean radeon_drm_bufmgr_add_buffer(struct pb_buffer *_buf,
+				     uint32_t rd, uint32_t wd);
 
-    /* Flush CB */
-    void (*flush_cb)(void *);
-    void *flush_data;
-};
 
-struct radeon_winsys* radeon_pipe_winsys(int fb);
-#if 0
-struct pipe_surface *radeon_surface_from_handle(struct radeon_context *radeon_context,
-                                             uint32_t handle,
-                                             enum pipe_format format,
-                                             int w, int h, int pitch);
-#endif
+void radeon_drm_bufmgr_write_reloc(struct pb_buffer *_buf,
+				   uint32_t rd, uint32_t wd,
+				   uint32_t flags);
+
+struct radeon_libdrm_winsys* radeon_pipe_winsys(int fd);
+
+boolean radeon_drm_bufmgr_shared_handle_from_buffer(struct pb_buffer *_buf,
+						    uint32_t *handle);
+uint32_t radeon_drm_bufmgr_handle_from_buffer(struct pb_buffer *_buf);
+struct pb_buffer *radeon_drm_bufmgr_create_buffer_from_handle(struct pb_manager *_mgr,
+							      uint32_t handle);
+
+void radeon_drm_bufmgr_set_tiling(struct pb_buffer *_buf, boolean microtiled, boolean macrotiled, uint32_t pitch);
+
+void radeon_drm_bufmgr_flush_maps(struct pb_manager *_mgr);
 #endif
diff --git a/src/gallium/winsys/drm/radeon/core/radeon_drm.c b/src/gallium/winsys/drm/radeon/core/radeon_drm.c
index 0c0e118ba3..f886fda2f2 100644
--- a/src/gallium/winsys/drm/radeon/core/radeon_drm.c
+++ b/src/gallium/winsys/drm/radeon/core/radeon_drm.c
@@ -31,8 +31,22 @@
 
 #include "radeon_drm.h"
 
+static struct radeon_libdrm_winsys *
+radeon_winsys_create(int fd)
+{
+    struct radeon_libdrm_winsys *rws;
+
+    rws = CALLOC_STRUCT(radeon_libdrm_winsys);
+    if (rws == NULL) {
+        return NULL;
+    }
+
+    rws->fd = fd;
+    return rws;
+}
+
 /* Helper function to do the ioctls needed for setup and init. */
-static void do_ioctls(int fd, struct radeon_winsys* winsys)
+static void do_ioctls(int fd, struct radeon_libdrm_winsys* winsys)
 {
     struct drm_radeon_gem_info gem_info = {0};
     struct drm_radeon_info info = {0};
@@ -123,62 +137,26 @@ struct pipe_screen* radeon_create_screen(struct drm_api* api,
                                          int drmFB,
                                          struct drm_create_screen_arg *arg)
 {
-    struct radeon_winsys* rwinsys = radeon_pipe_winsys(drmFB);
-    do_ioctls(drmFB, rwinsys);
+    struct radeon_libdrm_winsys* rws; 
+
+    rws = radeon_winsys_create(drmFB);
+    if (!rws)
+	return NULL;
+
+    do_ioctls(drmFB, rws);
 
     /* The state tracker can organize a softpipe fallback if no hw
      * driver is found.
      */
-    if (is_r3xx(rwinsys->pci_id)) {
-        radeon_setup_winsys(drmFB, rwinsys);
-        return r300_create_screen(rwinsys);
+    if (is_r3xx(rws->pci_id)) {
+        radeon_setup_winsys(drmFB, rws);
+        return r300_create_screen(&rws->base);
     } else {
-        FREE(rwinsys);
+        FREE(rws);
         return NULL;
     }
 }
 
-
-boolean radeon_buffer_from_texture(struct drm_api* api,
-                                   struct pipe_screen* screen,
-                                   struct pipe_texture* texture,
-                                   struct pipe_buffer** buffer,
-                                   unsigned* stride)
-{
-    /* XXX fix this */
-    return r300_get_texture_buffer(screen, texture, buffer, stride);
-}
-
-/* Create a buffer from a handle. */
-/* XXX what's up with name? */
-struct pipe_buffer* radeon_buffer_from_handle(struct drm_api* api,
-                                              struct pipe_screen* screen,
-                                              const char* name,
-                                              unsigned handle)
-{
-    struct radeon_bo_manager* bom =
-        ((struct radeon_winsys*)screen->winsys)->priv->bom;
-    struct radeon_pipe_buffer* radeon_buffer;
-    struct radeon_bo* bo = NULL;
-
-    bo = radeon_bo_open(bom, handle, 0, 0, 0, 0);
-    if (bo == NULL) {
-        return NULL;
-    }
-
-    radeon_buffer = CALLOC_STRUCT(radeon_pipe_buffer);
-    if (radeon_buffer == NULL) {
-        radeon_bo_unref(bo);
-        return NULL;
-    }
-
-    pipe_reference_init(&radeon_buffer->base.reference, 1);
-    radeon_buffer->base.screen = screen;
-    radeon_buffer->base.usage = PIPE_BUFFER_USAGE_PIXEL;
-    radeon_buffer->bo = bo;
-    return &radeon_buffer->base;
-}
-
 static struct pipe_texture*
 radeon_texture_from_shared_handle(struct drm_api *api,
                                   struct pipe_screen *screen,
@@ -187,18 +165,20 @@ radeon_texture_from_shared_handle(struct drm_api *api,
                                   unsigned stride,
                                   unsigned handle)
 {
-    struct pipe_buffer *buffer;
+    struct pb_buffer *_buf;
+    struct r300_winsys_buffer *buf;
     struct pipe_texture *blanket;
+    struct radeon_libdrm_winsys *ws = radeon_winsys_screen(r300_winsys_screen(screen));
 
-    buffer = radeon_buffer_from_handle(api, screen, name, handle);
-    if (!buffer) {
-        return NULL;
+    _buf = radeon_drm_bufmgr_create_buffer_from_handle(ws->kman, handle);
+    if (!_buf) {
+	return NULL;
     }
 
-    blanket = screen->texture_blanket(screen, templ, &stride, buffer);
-
-    pipe_buffer_reference(&buffer, NULL);
+    buf = radeon_libdrm_winsys_buffer(_buf);
+    blanket = r300_texture_blanket_winsys_buffer(screen, templ, &stride, buf);
 
+    pb_reference(&_buf, NULL);
     return blanket;
 }
 
@@ -208,34 +188,14 @@ static boolean radeon_shared_handle_from_texture(struct drm_api *api,
                                                  unsigned *stride,
                                                  unsigned *handle)
 {
-    int retval, fd;
-    struct drm_gem_flink flink;
-    struct radeon_pipe_buffer* radeon_buffer;
-    struct pipe_buffer *buffer = NULL;
-
-    if (!radeon_buffer_from_texture(api, screen, texture, &buffer, stride)) {
+    struct r300_winsys_buffer *radeon_buffer;
+    struct pb_buffer *_buf;
+    if (!r300_get_texture_buffer(screen, texture, &radeon_buffer, stride)) {
         return FALSE;
     }
 
-    radeon_buffer = (struct radeon_pipe_buffer*)buffer;
-    if (!radeon_buffer->flinked) {
-        fd = ((struct radeon_winsys*)screen->winsys)->priv->fd;
-
-        flink.handle = radeon_buffer->bo->handle;
-
-        retval = ioctl(fd, DRM_IOCTL_GEM_FLINK, &flink);
-        if (retval) {
-            debug_printf("radeon: DRM_IOCTL_GEM_FLINK failed, error %d\n",
-                    retval);
-            return FALSE;
-        }
-
-        radeon_buffer->flink = flink.name;
-        radeon_buffer->flinked = TRUE;
-    }
-
-    *handle = radeon_buffer->flink;
-    return TRUE;
+    _buf = radeon_pb_buffer(radeon_buffer);
+    return radeon_drm_bufmgr_shared_handle_from_buffer(_buf, handle);
 }
 
 static boolean radeon_local_handle_from_texture(struct drm_api *api,
@@ -244,16 +204,18 @@ static boolean radeon_local_handle_from_texture(struct drm_api *api,
                                                 unsigned *stride,
                                                 unsigned *handle)
 {
-    struct pipe_buffer *buffer = NULL;
-    if (!radeon_buffer_from_texture(api, screen, texture, &buffer, stride)) {
+    struct r300_winsys_buffer *radeon_buffer;
+    struct pb_buffer *_buf;
+
+    if (!r300_get_texture_buffer(screen, texture, &radeon_buffer, stride)) {
         return FALSE;
     }
 
-    *handle = ((struct radeon_pipe_buffer*)buffer)->bo->handle;
-
-    pipe_buffer_reference(&buffer, NULL);
+    _buf = radeon_pb_buffer(radeon_buffer);
+    *handle = radeon_drm_bufmgr_handle_from_buffer(_buf);
 
-    return TRUE;
+    pb_reference(&_buf, NULL);
+    return true;
 }
 
 static void radeon_drm_api_destroy(struct drm_api *api)
diff --git a/src/gallium/winsys/drm/radeon/core/radeon_drm.h b/src/gallium/winsys/drm/radeon/core/radeon_drm.h
index 8d74cbafc2..2f55061330 100644
--- a/src/gallium/winsys/drm/radeon/core/radeon_drm.h
+++ b/src/gallium/winsys/drm/radeon/core/radeon_drm.h
@@ -59,11 +59,6 @@ boolean radeon_buffer_from_texture(struct drm_api* api,
                                    struct pipe_buffer** buffer,
                                    unsigned* stride);
 
-struct pipe_buffer* radeon_buffer_from_handle(struct drm_api* api,
-                                              struct pipe_screen* screen,
-                                              const char* name,
-                                              unsigned handle);
-
 boolean radeon_handle_from_buffer(struct drm_api* api,
                                   struct pipe_screen* screen,
                                   struct pipe_buffer* buffer,
diff --git a/src/gallium/winsys/drm/radeon/core/radeon_drm_buffer.c b/src/gallium/winsys/drm/radeon/core/radeon_drm_buffer.c
new file mode 100644
index 0000000000..5f2fd2d7fc
--- /dev/null
+++ b/src/gallium/winsys/drm/radeon/core/radeon_drm_buffer.c
@@ -0,0 +1,361 @@
+
+#include "radeon_drm.h"
+#include "radeon_bo_gem.h"
+#include "radeon_cs_gem.h"
+
+#include "util/u_inlines.h"
+#include "util/u_memory.h"
+#include "util/u_simple_list.h"
+#include "pipebuffer/pb_buffer.h"
+#include "pipebuffer/pb_bufmgr.h"
+
+#include "radeon_winsys.h"
+struct radeon_drm_bufmgr;
+
+struct radeon_drm_buffer {
+    struct pb_buffer base;
+    struct radeon_drm_bufmgr *mgr;
+
+    struct radeon_bo *bo;
+
+    boolean flinked;
+    uint32_t flink;
+
+    boolean mapped;
+    struct radeon_drm_buffer *next, *prev;
+};
+
+extern const struct pb_vtbl radeon_drm_buffer_vtbl;
+
+
+static INLINE struct radeon_drm_buffer *
+radeon_drm_buffer(struct pb_buffer *buf)
+{
+    assert(buf);
+    assert(buf->vtbl == &radeon_drm_buffer_vtbl);
+    return (struct radeon_drm_buffer *)buf;
+}
+
+struct radeon_drm_bufmgr {
+    struct pb_manager base;
+    struct radeon_libdrm_winsys *rws;
+    struct radeon_drm_buffer buffer_map_list;
+};
+
+static INLINE struct radeon_drm_bufmgr *
+radeon_drm_bufmgr(struct pb_manager *mgr)
+{
+    assert(mgr);
+    return (struct radeon_drm_bufmgr *)mgr;
+}
+
+static void
+radeon_drm_buffer_destroy(struct pb_buffer *_buf)
+{
+    struct radeon_drm_buffer *buf = radeon_drm_buffer(_buf);
+
+    if (buf->mapped) {
+	remove_from_list(buf);
+	radeon_bo_unmap(buf->bo);
+	buf->mapped = false;
+    }
+    radeon_bo_unref(buf->bo);
+
+    FREE(buf);
+}
+
+static void *
+radeon_drm_buffer_map(struct pb_buffer *_buf,
+		      unsigned flags)
+{
+    struct radeon_drm_buffer *buf = radeon_drm_buffer(_buf);
+    int write;
+    
+    if (flags & PIPE_BUFFER_USAGE_DONTBLOCK) {
+        uint32_t domain;
+
+        if (radeon_bo_is_busy(buf->bo, &domain))
+            return NULL;
+    }
+
+    if (radeon_bo_is_referenced_by_cs(buf->bo, buf->mgr->rws->cs)) {
+        buf->mgr->rws->flush_cb(buf->mgr->rws->flush_data);
+    }
+
+    if (buf->mapped)
+	return buf->bo->ptr;
+
+    if (flags & PIPE_BUFFER_USAGE_CPU_WRITE) {
+        write = 1;
+    }
+
+    if (radeon_bo_map(buf->bo, write)) {
+        return NULL;
+    }
+    buf->mapped = true;
+    insert_at_tail(&buf->mgr->buffer_map_list, buf);
+    return buf->bo->ptr;
+}
+
+static void
+radeon_drm_buffer_unmap(struct pb_buffer *_buf)
+{
+    (void)_buf;
+}
+
+static void
+radeon_drm_buffer_get_base_buffer(struct pb_buffer *buf,
+				  struct pb_buffer **base_buf,
+				  unsigned *offset)
+{
+    *base_buf = buf;
+    *offset = 0;
+}
+
+
+static enum pipe_error
+radeon_drm_buffer_validate(struct pb_buffer *_buf, 
+			   struct pb_validate *vl,
+			   unsigned flags)
+{
+   /* Always pinned */
+   return PIPE_OK;
+}
+
+static void
+radeon_drm_buffer_fence(struct pb_buffer *buf,
+			struct pipe_fence_handle *fence)
+{
+}
+
+const struct pb_vtbl radeon_drm_buffer_vtbl = {
+    radeon_drm_buffer_destroy,
+    radeon_drm_buffer_map,
+    radeon_drm_buffer_unmap,
+    radeon_drm_buffer_validate,
+    radeon_drm_buffer_fence,
+    radeon_drm_buffer_get_base_buffer,
+};
+
+
+static uint32_t radeon_domain_from_usage(unsigned usage)
+{
+    uint32_t domain = 0;
+
+    if (usage & PIPE_BUFFER_USAGE_GPU_WRITE) {
+        domain |= RADEON_GEM_DOMAIN_VRAM;
+    }
+    if (usage & PIPE_BUFFER_USAGE_PIXEL) {
+        domain |= RADEON_GEM_DOMAIN_VRAM;
+    }
+    if (usage & PIPE_BUFFER_USAGE_VERTEX) {
+        domain |= RADEON_GEM_DOMAIN_GTT;
+    }
+    if (usage & PIPE_BUFFER_USAGE_INDEX) {
+        domain |= RADEON_GEM_DOMAIN_GTT;
+    }
+
+    return domain;
+}
+
+struct pb_buffer *radeon_drm_bufmgr_create_buffer_from_handle(struct pb_manager *_mgr,
+							      uint32_t handle)
+{
+    struct radeon_drm_bufmgr *mgr = radeon_drm_bufmgr(_mgr);
+    struct radeon_libdrm_winsys *rws = mgr->rws;
+    struct radeon_drm_buffer *buf;
+    struct radeon_bo *bo;
+
+    bo = radeon_bo_open(rws->bom, handle, 0,
+			0, 0, 0);
+    if (bo == NULL)
+	return NULL;
+
+    buf = CALLOC_STRUCT(radeon_drm_buffer);
+    if (!buf) {
+	radeon_bo_unref(bo);
+	return NULL;
+    }
+
+    make_empty_list(buf);
+
+    pipe_reference_init(&buf->base.base.reference, 1);
+    buf->base.base.alignment = 0;
+    buf->base.base.usage = PIPE_BUFFER_USAGE_PIXEL;
+    buf->base.base.size = 0;
+    buf->base.vtbl = &radeon_drm_buffer_vtbl;
+    buf->mgr = mgr;
+
+    buf->bo = bo;
+
+    //    fprintf(stderr,"hand %p : %d\n", &buf->base, handle);
+    return &buf->base;
+
+}
+
+static struct pb_buffer *
+radeon_drm_bufmgr_create_buffer(struct pb_manager *_mgr,
+				pb_size size,
+				const struct pb_desc *desc)
+{
+    struct radeon_drm_bufmgr *mgr = radeon_drm_bufmgr(_mgr);
+    struct radeon_libdrm_winsys *rws = mgr->rws;
+    struct radeon_drm_buffer *buf;
+    uint32_t domain;
+
+    buf = CALLOC_STRUCT(radeon_drm_buffer);
+    if (!buf)
+	goto error1;
+
+    pipe_reference_init(&buf->base.base.reference, 1);
+    buf->base.base.alignment = desc->alignment;
+    buf->base.base.usage = desc->usage;
+    buf->base.base.size = size;
+    buf->base.vtbl = &radeon_drm_buffer_vtbl;
+    buf->mgr = mgr;
+
+    make_empty_list(buf);
+    domain = radeon_domain_from_usage(desc->usage);
+    buf->bo = radeon_bo_open(rws->bom, 0, size,
+			     desc->alignment, domain, 0);
+    if (buf->bo == NULL)
+	goto error2;
+
+    //    fprintf(stderr,"%p : %d\n", &buf->base, size);
+    return &buf->base;
+
+ error2:
+    FREE(buf);
+ error1:
+    return NULL; 
+}
+
+static void
+radeon_drm_bufmgr_flush(struct pb_manager *mgr)
+{
+    /* NOP */
+}
+
+static void
+radeon_drm_bufmgr_destroy(struct pb_manager *_mgr)
+{
+    struct radeon_drm_bufmgr *mgr = radeon_drm_bufmgr(_mgr);
+    FREE(mgr);
+}
+
+struct pb_manager *
+radeon_drm_bufmgr_create(struct radeon_libdrm_winsys *rws)
+{
+    struct radeon_drm_bufmgr *mgr;
+
+    mgr = CALLOC_STRUCT(radeon_drm_bufmgr);
+    if (!mgr)
+	return NULL;
+
+    mgr->base.destroy = radeon_drm_bufmgr_destroy;
+    mgr->base.create_buffer = radeon_drm_bufmgr_create_buffer;
+    mgr->base.flush = radeon_drm_bufmgr_flush;
+
+    mgr->rws = rws;
+    make_empty_list(&mgr->buffer_map_list);
+    return &mgr->base;
+}
+
+uint32_t radeon_drm_bufmgr_handle_from_buffer(struct pb_buffer *_buf)
+{
+    struct radeon_drm_buffer *buf = radeon_drm_buffer(_buf);
+
+    return buf->bo->handle;
+}
+
+boolean radeon_drm_bufmgr_shared_handle_from_buffer(struct pb_buffer *_buf,
+						     uint32_t *handle)
+{
+    struct radeon_drm_buffer *buf = radeon_drm_buffer(_buf);
+    struct drm_gem_flink flink;
+    int retval, fd;
+    if (!buf->flinked) {
+	fd = buf->mgr->rws->fd;
+	flink.handle = buf->bo->handle;
+
+	retval = ioctl(fd, DRM_IOCTL_GEM_FLINK, &flink);
+	if (retval) {
+	    return false;
+	}
+
+	buf->flinked = TRUE;
+	buf->flink = flink.name;
+    }
+    *handle = buf->flink;
+    return TRUE;
+}
+
+static struct radeon_drm_buffer *get_drm_buffer(struct pb_buffer *_buf)
+{
+    struct radeon_drm_buffer *buf;
+    if (_buf->vtbl == &radeon_drm_buffer_vtbl) {
+        buf = radeon_drm_buffer(_buf);
+    } else {
+	struct pb_buffer *base_buf;
+	pb_size offset;
+	pb_get_base_buffer(_buf, &base_buf, &offset);
+
+	buf = radeon_drm_buffer(base_buf);
+    }
+    return buf;
+}
+
+void radeon_drm_bufmgr_set_tiling(struct pb_buffer *_buf, boolean microtiled, boolean macrotiled, uint32_t pitch)
+{
+    struct radeon_drm_buffer *buf = get_drm_buffer(_buf);
+    uint32_t flags = 0;
+
+    if (microtiled)
+	flags |= RADEON_BO_FLAGS_MICRO_TILE;
+    if (macrotiled)
+	flags |= RADEON_BO_FLAGS_MACRO_TILE;
+
+    radeon_bo_set_tiling(buf->bo, flags, pitch);
+
+}
+
+boolean radeon_drm_bufmgr_add_buffer(struct pb_buffer *_buf,
+				     uint32_t rd, uint32_t wd)
+{
+    struct radeon_drm_buffer *buf = get_drm_buffer(_buf);
+    radeon_cs_space_add_persistent_bo(buf->mgr->rws->cs, buf->bo,
+					  rd, wd);
+
+    return true;
+}
+
+void radeon_drm_bufmgr_write_reloc(struct pb_buffer *_buf,
+				   uint32_t rd, uint32_t wd,
+				   uint32_t flags)
+{
+    struct radeon_drm_buffer *buf = get_drm_buffer(_buf);
+    int retval;
+
+    retval = radeon_cs_write_reloc(buf->mgr->rws->cs,
+				   buf->bo, rd, wd, flags);
+    if (retval) {
+        debug_printf("radeon: Relocation of %p (%d, %d, %d) failed!\n",
+		     buf, rd, wd, flags);
+    }
+}
+
+void radeon_drm_bufmgr_flush_maps(struct pb_manager *_mgr)
+{
+    struct radeon_drm_bufmgr *mgr = radeon_drm_bufmgr(_mgr);
+    struct radeon_drm_buffer *rpb, *t_rpb;
+
+    foreach_s(rpb, t_rpb, &mgr->buffer_map_list) {
+	rpb->mapped = 0;
+	radeon_bo_unmap(rpb->bo);
+	remove_from_list(rpb);
+    }
+
+    make_empty_list(&mgr->buffer_map_list);
+
+    
+}
diff --git a/src/gallium/winsys/drm/radeon/core/radeon_r300.c b/src/gallium/winsys/drm/radeon/core/radeon_r300.c
index d759beaba1..997abb8f5c 100644
--- a/src/gallium/winsys/drm/radeon/core/radeon_r300.c
+++ b/src/gallium/winsys/drm/radeon/core/radeon_r300.c
@@ -22,29 +22,104 @@
 
 #include "radeon_r300.h"
 
-static void radeon_set_flush_cb(struct radeon_winsys *winsys,
+static struct r300_winsys_buffer *
+radeon_r300_winsys_buffer_create(struct r300_winsys_screen *rws,
+				 unsigned alignment,
+				 unsigned usage,
+				 unsigned size)
+{
+    struct radeon_libdrm_winsys *ws = radeon_winsys_screen(rws);
+    struct pb_desc desc;
+    struct pb_manager *provider;
+    struct pb_buffer *buffer;
+
+    memset(&desc, 0, sizeof(desc));
+    desc.alignment = alignment;
+    desc.usage = usage;
+
+    if (usage == PIPE_BUFFER_USAGE_CONSTANT)
+        provider = ws->mman;
+    else if (usage == PIPE_BUFFER_USAGE_VERTEX ||
+	     usage == PIPE_BUFFER_USAGE_INDEX)
+        provider = ws->cman;
+    else
+        provider = ws->kman;
+    buffer = provider->create_buffer(provider, size, &desc);
+    if (!buffer)
+	return NULL;
+
+    return radeon_libdrm_winsys_buffer(buffer);
+}
+
+static void radeon_r300_winsys_buffer_destroy(struct r300_winsys_buffer *buf)
+{
+    struct pb_buffer *_buf = radeon_pb_buffer(buf);
+
+    pb_destroy(_buf);
+}
+static void radeon_r300_winsys_buffer_set_tiling(struct r300_winsys_screen *rws,
+						  struct r300_winsys_buffer *buf,
+						  uint32_t pitch,
+						  boolean microtiled,
+						  boolean macrotiled)
+{
+    struct pb_buffer *_buf = radeon_pb_buffer(buf);
+    radeon_drm_bufmgr_set_tiling(_buf, microtiled, macrotiled, pitch);
+}
+
+static void *radeon_r300_winsys_buffer_map(struct r300_winsys_screen *ws,
+					  struct r300_winsys_buffer *buf,
+					  unsigned usage)
+{
+    struct pb_buffer *_buf = radeon_pb_buffer(buf);
+
+    return pb_map(_buf, 0);
+}
+
+static void radeon_r300_winsys_buffer_unmap(struct r300_winsys_screen *ws,
+					    struct r300_winsys_buffer *buf)
+{
+    struct pb_buffer *_buf = radeon_pb_buffer(buf);
+
+    pb_unmap(_buf);
+}
+
+static void radeon_r300_winsys_buffer_reference(struct r300_winsys_screen *rws,
+						struct r300_winsys_buffer **pdst,
+						struct r300_winsys_buffer *src)
+{
+    struct pb_buffer *_src = radeon_pb_buffer(src);
+    struct pb_buffer *_dst = radeon_pb_buffer(*pdst);
+
+    pb_reference(&_dst, _src);
+
+    *pdst = radeon_libdrm_winsys_buffer(_dst);
+}
+
+static void radeon_set_flush_cb(struct r300_winsys_screen *rws,
                                 void (*flush_cb)(void *),
                                 void *data)
 {
-    winsys->priv->flush_cb = flush_cb;
-    winsys->priv->flush_data = data;
-    radeon_cs_space_set_flush(winsys->priv->cs, flush_cb, data);
+    struct radeon_libdrm_winsys *ws = radeon_winsys_screen(rws);
+    ws->flush_cb = flush_cb;
+    ws->flush_data = data;
+    radeon_cs_space_set_flush(ws->cs, flush_cb, data);
 }
 
-static boolean radeon_add_buffer(struct radeon_winsys* winsys,
-                                 struct pipe_buffer* pbuffer,
+static boolean radeon_add_buffer(struct r300_winsys_screen *rws,
+                                 struct r300_winsys_buffer *buf,
                                  uint32_t rd,
                                  uint32_t wd)
 {
-    struct radeon_bo* bo = ((struct radeon_pipe_buffer*)pbuffer)->bo;
+    struct pb_buffer *_buf = radeon_pb_buffer(buf);
 
-    radeon_cs_space_add_persistent_bo(winsys->priv->cs, bo, rd, wd);
-    return TRUE;
+    return radeon_drm_bufmgr_add_buffer(_buf, rd, wd);
 }
 
-static boolean radeon_validate(struct radeon_winsys* winsys)
+static boolean radeon_validate(struct r300_winsys_screen *rws)
 {
-    if (radeon_cs_space_check(winsys->priv->cs) < 0) {
+    struct radeon_libdrm_winsys *ws = radeon_winsys_screen(rws);
+    if (radeon_cs_space_check(ws->cs) < 0) {
         return FALSE;
     }
 
@@ -52,108 +127,146 @@ static boolean radeon_validate(struct radeon_winsys* winsys)
     return TRUE;
 }
 
-static boolean radeon_check_cs(struct radeon_winsys* winsys, int size)
+static boolean radeon_check_cs(struct r300_winsys_screen *rws, int size)
 {
-    struct radeon_cs* cs = winsys->priv->cs;
+    struct radeon_libdrm_winsys *ws = radeon_winsys_screen(rws);
+    struct radeon_cs *cs = ws->cs;
 
-    return radeon_validate(winsys) && cs->cdw + size <= cs->ndw;
+    return radeon_validate(rws) && cs->cdw + size <= cs->ndw;
 }
 
-static void radeon_begin_cs(struct radeon_winsys* winsys,
+static void radeon_begin_cs(struct r300_winsys_screen *rws,
                             int size,
                             const char* file,
                             const char* function,
                             int line)
 {
-    radeon_cs_begin(winsys->priv->cs, size, file, function, line);
+    struct radeon_libdrm_winsys *ws = radeon_winsys_screen(rws);
+    radeon_cs_begin(ws->cs, size, file, function, line);
 }
 
-static void radeon_write_cs_dword(struct radeon_winsys* winsys,
+static void radeon_write_cs_dword(struct r300_winsys_screen *rws,
                                   uint32_t dword)
 {
-    radeon_cs_write_dword(winsys->priv->cs, dword);
+    struct radeon_libdrm_winsys *ws = radeon_winsys_screen(rws);
+    radeon_cs_write_dword(ws->cs, dword);
 }
 
-static void radeon_write_cs_reloc(struct radeon_winsys* winsys,
-                                  struct pipe_buffer* pbuffer,
+static void radeon_write_cs_reloc(struct r300_winsys_screen *rws,
+                                  struct r300_winsys_buffer *buf,
                                   uint32_t rd,
                                   uint32_t wd,
                                   uint32_t flags)
 {
-    int retval = 0;
-    struct radeon_pipe_buffer* radeon_buffer =
-        (struct radeon_pipe_buffer*)pbuffer;
-
-    assert(!radeon_buffer->pb);
-
-    retval = radeon_cs_write_reloc(winsys->priv->cs, radeon_buffer->bo,
-                                   rd, wd, flags);
-
-    if (retval) {
-        debug_printf("radeon: Relocation of %p (%d, %d, %d) failed!\n",
-                pbuffer, rd, wd, flags);
-    }
+    struct pb_buffer *_buf = radeon_pb_buffer(buf);
+    radeon_drm_bufmgr_write_reloc(_buf, rd, wd, flags);
 }
 
-static void radeon_reset_bos(struct radeon_winsys *winsys)
+static void radeon_reset_bos(struct r300_winsys_screen *rws)
 {
-    radeon_cs_space_reset_bos(winsys->priv->cs);
+    struct radeon_libdrm_winsys *ws = radeon_winsys_screen(rws);
+    radeon_cs_space_reset_bos(ws->cs);
 }
 
-static void radeon_end_cs(struct radeon_winsys* winsys,
+static void radeon_end_cs(struct r300_winsys_screen *rws,
                           const char* file,
                           const char* function,
                           int line)
 {
-    radeon_cs_end(winsys->priv->cs, file, function, line);
+    struct radeon_libdrm_winsys *ws = radeon_winsys_screen(rws);
+    radeon_cs_end(ws->cs, file, function, line);
 }
 
-static void radeon_flush_cs(struct radeon_winsys* winsys)
+static void radeon_flush_cs(struct r300_winsys_screen *rws)
 {
+    struct radeon_libdrm_winsys *ws = radeon_winsys_screen(rws);
     int retval;
 
     /* Don't flush a zero-sized CS. */
-    if (!winsys->priv->cs->cdw) {
+    if (!ws->cs->cdw) {
         return;
     }
 
+    radeon_drm_bufmgr_flush_maps(ws->kman);
     /* Emit the CS. */
-    retval = radeon_cs_emit(winsys->priv->cs);
+    retval = radeon_cs_emit(ws->cs);
     if (retval) {
         debug_printf("radeon: Bad CS, dumping...\n");
-        radeon_cs_print(winsys->priv->cs, stderr);
+        radeon_cs_print(ws->cs, stderr);
     }
 
     /* Reset CS.
      * Someday, when we care about performance, we should really find a way
      * to rotate between two or three CS objects so that the GPU can be
      * spinning through one CS while another one is being filled. */
-    radeon_cs_erase(winsys->priv->cs);
+    radeon_cs_erase(ws->cs);
+}
+
+static uint32_t radeon_get_value(struct r300_winsys_screen *rws,
+			     enum r300_value_id id)
+{
+    struct radeon_libdrm_winsys *ws = (struct radeon_libdrm_winsys *)rws;
+
+    switch(id) {
+    case R300_VID_PCI_ID:
+	return ws->pci_id;
+    case R300_VID_GB_PIPES:
+	return ws->gb_pipes;
+    case R300_VID_Z_PIPES:
+	return ws->z_pipes;
+    }
+    return 0;
+}
+
+static void
+radeon_winsys_destroy(struct r300_winsys_screen *rws)
+{
+    struct radeon_libdrm_winsys *ws = (struct radeon_libdrm_winsys *)rws;
+    radeon_cs_destroy(ws->cs);
+
+    ws->cman->destroy(ws->cman);
+    ws->kman->destroy(ws->kman);
+    ws->mman->destroy(ws->mman);
+
+    radeon_bo_manager_gem_dtor(ws->bom);
+    radeon_cs_manager_gem_dtor(ws->csm);
 }
 
 void
-radeon_setup_winsys(int fd, struct radeon_winsys* winsys)
+radeon_setup_winsys(int fd, struct radeon_libdrm_winsys* ws)
 {
-    struct radeon_winsys_priv* priv = winsys->priv;
+    ws->csm = radeon_cs_manager_gem_ctor(fd);
+
+    ws->bom = radeon_bo_manager_gem_ctor(fd);
+    ws->kman = radeon_drm_bufmgr_create(ws);
 
-    priv->csm = radeon_cs_manager_gem_ctor(fd);
+    ws->cman = pb_cache_manager_create(ws->kman, 100000);
+    ws->mman = pb_malloc_bufmgr_create();
 
     /* Size limit on IBs is 64 kibibytes. */
-    priv->cs = radeon_cs_create(priv->csm, 1024 * 64 / 4);
-    radeon_cs_set_limit(priv->cs,
-            RADEON_GEM_DOMAIN_GTT, winsys->gart_size);
-    radeon_cs_set_limit(priv->cs,
-            RADEON_GEM_DOMAIN_VRAM, winsys->vram_size);
-
-    winsys->add_buffer = radeon_add_buffer;
-    winsys->validate = radeon_validate;
-
-    winsys->check_cs = radeon_check_cs;
-    winsys->begin_cs = radeon_begin_cs;
-    winsys->write_cs_dword = radeon_write_cs_dword;
-    winsys->write_cs_reloc = radeon_write_cs_reloc;
-    winsys->end_cs = radeon_end_cs;
-    winsys->flush_cs = radeon_flush_cs;
-    winsys->reset_bos = radeon_reset_bos;
-    winsys->set_flush_cb = radeon_set_flush_cb;
+    ws->cs = radeon_cs_create(ws->csm, 1024 * 64 / 4);
+    radeon_cs_set_limit(ws->cs,
+            RADEON_GEM_DOMAIN_GTT, ws->gart_size);
+    radeon_cs_set_limit(ws->cs,
+            RADEON_GEM_DOMAIN_VRAM, ws->vram_size);
+
+    ws->base.add_buffer = radeon_add_buffer;
+    ws->base.validate = radeon_validate;
+    ws->base.destroy = radeon_winsys_destroy;
+    ws->base.check_cs = radeon_check_cs;
+    ws->base.begin_cs = radeon_begin_cs;
+    ws->base.write_cs_dword = radeon_write_cs_dword;
+    ws->base.write_cs_reloc = radeon_write_cs_reloc;
+    ws->base.end_cs = radeon_end_cs;
+    ws->base.flush_cs = radeon_flush_cs;
+    ws->base.reset_bos = radeon_reset_bos;
+    ws->base.set_flush_cb = radeon_set_flush_cb;
+    ws->base.get_value = radeon_get_value;
+
+    ws->base.buffer_create = radeon_r300_winsys_buffer_create;
+    ws->base.buffer_destroy = radeon_r300_winsys_buffer_destroy;
+    ws->base.buffer_set_tiling = radeon_r300_winsys_buffer_set_tiling;
+    ws->base.buffer_map = radeon_r300_winsys_buffer_map;
+    ws->base.buffer_unmap = radeon_r300_winsys_buffer_unmap;
+    ws->base.buffer_reference = radeon_r300_winsys_buffer_reference;
 }
diff --git a/src/gallium/winsys/drm/radeon/core/radeon_r300.h b/src/gallium/winsys/drm/radeon/core/radeon_r300.h
index cfbdb30266..82a8aad715 100644
--- a/src/gallium/winsys/drm/radeon/core/radeon_r300.h
+++ b/src/gallium/winsys/drm/radeon/core/radeon_r300.h
@@ -28,12 +28,12 @@
 #include <stdlib.h>
 #include "drm.h"
 #include "radeon_drm.h"
+#include "radeon_bo_gem.h"
 #include "radeon_cs_gem.h"
 
-#include "r300_winsys.h"
-
+#include "radeon_winsys.h"
 #include "radeon_buffer.h"
 
-void radeon_setup_winsys(int fd, struct radeon_winsys* winsys);
+void radeon_setup_winsys(int fd, struct radeon_libdrm_winsys* winsys);
 
 #endif /* RADEON_R300_H */
diff --git a/src/gallium/winsys/drm/radeon/core/radeon_winsys.h b/src/gallium/winsys/drm/radeon/core/radeon_winsys.h
index 4901080ca7..c8e725cff0 100644
--- a/src/gallium/winsys/drm/radeon/core/radeon_winsys.h
+++ b/src/gallium/winsys/drm/radeon/core/radeon_winsys.h
@@ -30,16 +30,17 @@
 #ifndef RADEON_WINSYS_H
 #define RADEON_WINSYS_H
 
-#include "util/u_simple_screen.h"
+#include "r300_winsys.h"
 
-struct radeon_winsys_priv;
-
-struct radeon_winsys {
+struct radeon_libdrm_winsys {
     /* Parent class. */
-    struct pipe_winsys base;
+    struct r300_winsys_screen base;
+
+    struct pb_manager *kman;
 
-    /* Winsys private */
-    struct radeon_winsys_priv* priv;
+    struct pb_manager *mman;
+
+    struct pb_manager *cman;
 
     /* PCI ID */
     uint32_t pci_id;
@@ -56,56 +57,27 @@ struct radeon_winsys {
     /* VRAM size. */
     uint32_t vram_size;
 
-    /* Add a pipe_buffer to the list of buffer objects to validate. */
-    boolean (*add_buffer)(struct radeon_winsys* winsys,
-                          struct pipe_buffer* pbuffer,
-                          uint32_t rd,
-                          uint32_t wd);
-
-    /* Revalidate all currently setup pipe_buffers.
-     * Returns TRUE if a flush is required. */
-    boolean (*validate)(struct radeon_winsys* winsys);
-
-    /* Check to see if there's room for commands. */
-    boolean (*check_cs)(struct radeon_winsys* winsys, int size);
-
-    /* Start a command emit. */
-    void (*begin_cs)(struct radeon_winsys* winsys,
-                     int size,
-                     const char* file,
-                     const char* function,
-                     int line);
-
-    /* Write a dword to the command buffer. */
-    void (*write_cs_dword)(struct radeon_winsys* winsys, uint32_t dword);
-
-    /* Write a relocated dword to the command buffer. */
-    void (*write_cs_reloc)(struct radeon_winsys* winsys,
-                           struct pipe_buffer* bo,
-                           uint32_t rd,
-                           uint32_t wd,
-                           uint32_t flags);
-
-    /* Finish a command emit. */
-    void (*end_cs)(struct radeon_winsys* winsys,
-                   const char* file,
-                   const char* function,
-                   int line);
-
-    /* Flush the CS. */
-    void (*flush_cs)(struct radeon_winsys* winsys);
-
-    /* winsys flush - callback from winsys when flush required */
-    void (*set_flush_cb)(struct radeon_winsys *winsys,
-			 void (*flush_cb)(void *), void *data);
-
-    void (*reset_bos)(struct radeon_winsys *winsys);
-
-    void (*buffer_set_tiling)(struct radeon_winsys* winsys,
-                              struct pipe_buffer* buffer,
-                              uint32_t pitch,
-                              boolean microtiled,
-                              boolean macrotiled);
+    /* DRM FD */
+    int fd;
+
+    /* Radeon BO manager. */
+    struct radeon_bo_manager *bom;
+
+    /* Radeon CS manager. */
+    struct radeon_cs_manager *csm;
+
+    /* Current CS. */
+    struct radeon_cs *cs;
+
+    /* Flush CB */
+    void (*flush_cb)(void *);
+    void *flush_data;
 };
 
+static INLINE struct radeon_libdrm_winsys *
+radeon_winsys_screen(struct r300_winsys_screen *base)
+{
+  return (struct radeon_libdrm_winsys *)base;
+}
+
 #endif