r300g: rebuild winsys and command submission to support multiple contexts

4 files changed, 275 insertions, 258 deletions

diff --git a/src/gallium/winsys/radeon/drm/radeon_buffer.h b/src/gallium/winsys/radeon/drm/radeon_buffer.h
index 73cb6a579b4..a8137d85e83 100644
--- a/src/gallium/winsys/radeon/drm/radeon_buffer.h
+++ b/src/gallium/winsys/radeon/drm/radeon_buffer.h
@@ -43,10 +43,9 @@
 
 #include "radeon_winsys.h"
 
-#define RADEON_USAGE_DOMAIN_GTT  (1 << 29)
-#define RADEON_USAGE_DOMAIN_VRAM (1 << 30)
-
-#define RADEON_MAX_BOS 24
+#define RADEON_PB_USAGE_VERTEX      (1 << 28)
+#define RADEON_PB_USAGE_DOMAIN_GTT  (1 << 29)
+#define RADEON_PB_USAGE_DOMAIN_VRAM (1 << 30)
 
 static INLINE struct pb_buffer *
 radeon_pb_buffer(struct r300_winsys_buffer *buffer)
@@ -63,24 +62,26 @@ radeon_libdrm_winsys_buffer(struct pb_buffer *buffer)
 struct pb_manager *
 radeon_drm_bufmgr_create(struct radeon_libdrm_winsys *rws);
 
-boolean radeon_drm_bufmgr_add_buffer(struct pb_buffer *_buf,
-				     enum r300_buffer_domain rd,
-                                     enum r300_buffer_domain wd);
-
+void radeon_drm_bufmgr_add_buffer(struct r300_winsys_cs *cs,
+                                  struct r300_winsys_buffer *buf,
+                                  enum r300_buffer_domain rd,
+                                  enum r300_buffer_domain wd);
 
-void radeon_drm_bufmgr_write_reloc(struct pb_buffer *_buf,
+void radeon_drm_bufmgr_write_reloc(struct r300_winsys_cs *cs,
+                                   struct r300_winsys_buffer *buf,
 				   enum r300_buffer_domain rd,
-                                   enum r300_buffer_domain wd,
-				   uint32_t flags);
+                                   enum r300_buffer_domain wd);
 
 struct pb_buffer *radeon_drm_bufmgr_create_buffer_from_handle(struct pb_manager *_mgr,
 							      uint32_t handle);
 
-void radeon_drm_bufmgr_get_tiling(struct pb_buffer *_buf,
+void radeon_drm_bufmgr_get_tiling(struct r300_winsys_screen *ws,
+                                  struct r300_winsys_buffer *buf,
                                   enum r300_buffer_tiling *microtiled,
                                   enum r300_buffer_tiling *macrotiled);
 
-void radeon_drm_bufmgr_set_tiling(struct pb_buffer *_buf,
+void radeon_drm_bufmgr_set_tiling(struct r300_winsys_screen *ws,
+                                  struct r300_winsys_buffer *buf,
                                   enum r300_buffer_tiling microtiled,
                                   enum r300_buffer_tiling macrotiled,
                                   uint32_t pitch);
@@ -90,9 +91,19 @@ void radeon_drm_bufmgr_flush_maps(struct pb_manager *_mgr);
 boolean radeon_drm_bufmgr_get_handle(struct pb_buffer *_buf,
 				     struct winsys_handle *whandle);
 
-boolean radeon_drm_bufmgr_is_buffer_referenced(struct pb_buffer *_buf,
+boolean radeon_drm_bufmgr_is_buffer_referenced(struct r300_winsys_cs *cs,
+                                               struct r300_winsys_buffer *buf,
                                                enum r300_reference_domain domain);
 
-void radeon_drm_bufmgr_wait(struct pb_buffer *_buf);
+void radeon_drm_bufmgr_wait(struct r300_winsys_screen *ws,
+                            struct r300_winsys_buffer *buf);
+
+void *radeon_drm_buffer_map(struct r300_winsys_screen *ws,
+                            struct r300_winsys_buffer *buf,
+                            struct r300_winsys_cs *cs,
+                            enum pipe_transfer_usage usage);
+
+void radeon_drm_buffer_unmap(struct r300_winsys_screen *ws,
+                             struct r300_winsys_buffer *buf);
 
 #endif
diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_buffer.c b/src/gallium/winsys/radeon/drm/radeon_drm_buffer.c
index c5f133e7b28..5ea5912089e 100644
--- a/src/gallium/winsys/radeon/drm/radeon_drm_buffer.c
+++ b/src/gallium/winsys/radeon/drm/radeon_drm_buffer.c
@@ -20,6 +20,9 @@ struct radeon_drm_buffer {
 
     struct radeon_bo *bo;
 
+    /* The CS associated with the last buffer_map. */
+    struct radeon_libdrm_cs *cs;
+
     boolean flinked;
     uint32_t flink;
 
@@ -65,34 +68,53 @@ radeon_drm_buffer_destroy(struct pb_buffer *_buf)
     FREE(buf);
 }
 
+static unsigned get_pb_usage_from_transfer_flags(enum pipe_transfer_usage usage)
+{
+    unsigned res = 0;
+
+    if (usage & PIPE_TRANSFER_READ)
+        res |= PB_USAGE_CPU_READ;
+
+    if (usage & PIPE_TRANSFER_WRITE)
+        res |= PB_USAGE_CPU_WRITE;
+
+    if (usage & PIPE_TRANSFER_DONTBLOCK)
+        res |= PB_USAGE_DONTBLOCK;
+
+    if (usage & PIPE_TRANSFER_UNSYNCHRONIZED)
+        res |= PB_USAGE_UNSYNCHRONIZED;
+
+    return res;
+}
+
 static void *
-radeon_drm_buffer_map(struct pb_buffer *_buf,
+radeon_drm_buffer_map_internal(struct pb_buffer *_buf,
 		      unsigned flags)
 {
     struct radeon_drm_buffer *buf = radeon_drm_buffer(_buf);
+    struct radeon_libdrm_cs *cs = buf->cs;
     int write = 0;
 
-    if (flags & PIPE_TRANSFER_DONTBLOCK) {
-	if ((_buf->base.usage & PIPE_BIND_VERTEX_BUFFER) ||
-	    (_buf->base.usage & PIPE_BIND_INDEX_BUFFER))
-	    if (radeon_bo_is_referenced_by_cs(buf->bo, buf->mgr->rws->cs))
+    if (flags & PB_USAGE_DONTBLOCK) {
+        if (_buf->base.usage & RADEON_PB_USAGE_VERTEX)
+            if (cs && radeon_bo_is_referenced_by_cs(buf->bo, cs->cs))
 		return NULL;
     }
 
     if (buf->bo->ptr != NULL)
 	return buf->bo->ptr;
 
-    if (flags & PIPE_TRANSFER_DONTBLOCK) {
+    if (flags & PB_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 (cs && radeon_bo_is_referenced_by_cs(buf->bo, cs->cs)) {
+        cs->flush_cs(cs->flush_data);
     }
 
-    if (flags & PIPE_TRANSFER_WRITE) {
+    if (flags & PB_USAGE_CPU_WRITE) {
         write = 1;
     }
 
@@ -104,7 +126,7 @@ radeon_drm_buffer_map(struct pb_buffer *_buf,
 }
 
 static void
-radeon_drm_buffer_unmap(struct pb_buffer *_buf)
+radeon_drm_buffer_unmap_internal(struct pb_buffer *_buf)
 {
     (void)_buf;
 }
@@ -136,8 +158,8 @@ radeon_drm_buffer_fence(struct pb_buffer *buf,
 
 const struct pb_vtbl radeon_drm_buffer_vtbl = {
     radeon_drm_buffer_destroy,
-    radeon_drm_buffer_map,
-    radeon_drm_buffer_unmap,
+    radeon_drm_buffer_map_internal,
+    radeon_drm_buffer_unmap_internal,
     radeon_drm_buffer_validate,
     radeon_drm_buffer_fence,
     radeon_drm_buffer_get_base_buffer,
@@ -166,7 +188,7 @@ struct pb_buffer *radeon_drm_bufmgr_create_buffer_from_handle(struct pb_manager
 
     pipe_reference_init(&buf->base.base.reference, 1);
     buf->base.base.alignment = 0;
-    buf->base.base.usage = PIPE_BIND_SAMPLER_VIEW;
+    buf->base.base.usage = PB_USAGE_GPU_WRITE | PB_USAGE_GPU_READ;
     buf->base.base.size = 0;
     buf->base.vtbl = &radeon_drm_buffer_vtbl;
     buf->mgr = mgr;
@@ -200,8 +222,8 @@ radeon_drm_bufmgr_create_buffer(struct pb_manager *_mgr,
     make_empty_list(buf);
 
     domain =
-        (desc->usage & RADEON_USAGE_DOMAIN_GTT  ? RADEON_GEM_DOMAIN_GTT  : 0) |
-        (desc->usage & RADEON_USAGE_DOMAIN_VRAM ? RADEON_GEM_DOMAIN_VRAM : 0);
+        (desc->usage & RADEON_PB_USAGE_DOMAIN_GTT  ? RADEON_GEM_DOMAIN_GTT  : 0) |
+        (desc->usage & RADEON_PB_USAGE_DOMAIN_VRAM ? RADEON_GEM_DOMAIN_VRAM : 0);
 
     buf->bo = radeon_bo_open(rws->bom, 0, size,
 			     desc->alignment, domain, 0);
@@ -249,7 +271,8 @@ radeon_drm_bufmgr_create(struct radeon_libdrm_winsys *rws)
 
 static struct radeon_drm_buffer *get_drm_buffer(struct pb_buffer *_buf)
 {
-    struct radeon_drm_buffer *buf;
+    struct radeon_drm_buffer *buf = NULL;
+
     if (_buf->vtbl == &radeon_drm_buffer_vtbl) {
         buf = radeon_drm_buffer(_buf);
     } else {
@@ -257,11 +280,35 @@ static struct radeon_drm_buffer *get_drm_buffer(struct pb_buffer *_buf)
 	pb_size offset;
 	pb_get_base_buffer(_buf, &base_buf, &offset);
 
-	buf = radeon_drm_buffer(base_buf);
+        if (base_buf->vtbl == &radeon_drm_buffer_vtbl)
+            buf = radeon_drm_buffer(base_buf);
     }
+
     return buf;
 }
 
+void *radeon_drm_buffer_map(struct r300_winsys_screen *ws,
+                            struct r300_winsys_buffer *buf,
+                            struct r300_winsys_cs *cs,
+                            enum pipe_transfer_usage usage)
+{
+    struct pb_buffer *_buf = radeon_pb_buffer(buf);
+    struct radeon_drm_buffer *rbuf = get_drm_buffer(_buf);
+
+    if (rbuf)
+        rbuf->cs = radeon_libdrm_cs(cs);
+
+    return pb_map(_buf, get_pb_usage_from_transfer_flags(usage));
+}
+
+void radeon_drm_buffer_unmap(struct r300_winsys_screen *ws,
+                             struct r300_winsys_buffer *buf)
+{
+    struct pb_buffer *_buf = radeon_pb_buffer(buf);
+
+    pb_unmap(_buf);
+}
+
 boolean radeon_drm_bufmgr_get_handle(struct pb_buffer *_buf,
 				     struct winsys_handle *whandle)
 {
@@ -286,11 +333,12 @@ boolean radeon_drm_bufmgr_get_handle(struct pb_buffer *_buf,
     return TRUE;
 }
 
-void radeon_drm_bufmgr_get_tiling(struct pb_buffer *_buf,
+void radeon_drm_bufmgr_get_tiling(struct r300_winsys_screen *ws,
+                                  struct r300_winsys_buffer *_buf,
                                   enum r300_buffer_tiling *microtiled,
                                   enum r300_buffer_tiling *macrotiled)
 {
-    struct radeon_drm_buffer *buf = get_drm_buffer(_buf);
+    struct radeon_drm_buffer *buf = get_drm_buffer(radeon_pb_buffer(_buf));
     uint32_t flags = 0, pitch;
 
     radeon_bo_get_tiling(buf->bo, &flags, &pitch);
@@ -304,12 +352,13 @@ void radeon_drm_bufmgr_get_tiling(struct pb_buffer *_buf,
 	*macrotiled = R300_BUFFER_TILED;
 }
 
-void radeon_drm_bufmgr_set_tiling(struct pb_buffer *_buf,
+void radeon_drm_bufmgr_set_tiling(struct r300_winsys_screen *ws,
+                                  struct r300_winsys_buffer *_buf,
                                   enum r300_buffer_tiling microtiled,
                                   enum r300_buffer_tiling macrotiled,
                                   uint32_t pitch)
 {
-    struct radeon_drm_buffer *buf = get_drm_buffer(_buf);
+    struct radeon_drm_buffer *buf = get_drm_buffer(radeon_pb_buffer(_buf));
     uint32_t flags = 0;
     if (microtiled == R300_BUFFER_TILED)
         flags |= RADEON_BO_FLAGS_MICRO_TILE;
@@ -324,56 +373,60 @@ void radeon_drm_bufmgr_set_tiling(struct pb_buffer *_buf,
     radeon_bo_set_tiling(buf->bo, flags, pitch);
 }
 
-static uint32_t gem_domain(enum r300_buffer_domain dom)
+static uint32_t get_gem_domain(enum r300_buffer_domain domain)
 {
     uint32_t res = 0;
 
-    if (dom & R300_DOMAIN_GTT)
+    if (domain & R300_DOMAIN_GTT)
         res |= RADEON_GEM_DOMAIN_GTT;
-    if (dom & R300_DOMAIN_VRAM)
+    if (domain & R300_DOMAIN_VRAM)
         res |= RADEON_GEM_DOMAIN_VRAM;
     return res;
 }
 
-boolean radeon_drm_bufmgr_add_buffer(struct pb_buffer *_buf,
-				     enum r300_buffer_domain rd,
-                                     enum r300_buffer_domain wd)
+void radeon_drm_bufmgr_add_buffer(struct r300_winsys_cs *rcs,
+                                  struct r300_winsys_buffer *_buf,
+                                  enum r300_buffer_domain rd,
+                                  enum r300_buffer_domain wd)
 {
-    struct radeon_drm_buffer *buf = get_drm_buffer(_buf);
-    uint32_t gem_rd = gem_domain(rd);
-    uint32_t gem_wd = gem_domain(wd);
+    struct radeon_libdrm_cs *cs = radeon_libdrm_cs(rcs);
+    struct radeon_drm_buffer *buf = get_drm_buffer(radeon_pb_buffer(_buf));
+    uint32_t gem_rd = get_gem_domain(rd);
+    uint32_t gem_wd = get_gem_domain(wd);
 
-    radeon_cs_space_add_persistent_bo(buf->mgr->rws->cs, buf->bo,
-					  gem_rd, gem_wd);
-    return TRUE;
+    radeon_cs_space_add_persistent_bo(cs->cs, buf->bo, gem_rd, gem_wd);
 }
 
-void radeon_drm_bufmgr_write_reloc(struct pb_buffer *_buf,
+void radeon_drm_bufmgr_write_reloc(struct r300_winsys_cs *rcs,
+                                   struct r300_winsys_buffer *_buf,
 				   enum r300_buffer_domain rd,
-                                   enum r300_buffer_domain wd,
-				   uint32_t flags)
+                                   enum r300_buffer_domain wd)
 {
-    struct radeon_drm_buffer *buf = get_drm_buffer(_buf);
+    struct radeon_libdrm_cs *cs = radeon_libdrm_cs(rcs);
+    struct radeon_drm_buffer *buf = get_drm_buffer(radeon_pb_buffer(_buf));
     int retval;
-    uint32_t gem_rd = gem_domain(rd);
-    uint32_t gem_wd = gem_domain(wd);
+    uint32_t gem_rd = get_gem_domain(rd);
+    uint32_t gem_wd = get_gem_domain(wd);
 
-    retval = radeon_cs_write_reloc(buf->mgr->rws->cs,
-				   buf->bo, gem_rd, gem_wd, flags);
+    cs->cs->cdw = cs->base.cdw;
+    retval = radeon_cs_write_reloc(cs->cs, buf->bo, gem_rd, gem_wd, 0);
+    cs->base.cdw = cs->cs->cdw;
     if (retval) {
         fprintf(stderr, "radeon: Relocation of %p (%d, %d, %d) failed!\n",
-                buf, gem_rd, gem_wd, flags);
+                buf, gem_rd, gem_wd, 0);
     }
 }
 
-boolean radeon_drm_bufmgr_is_buffer_referenced(struct pb_buffer *_buf,
+boolean radeon_drm_bufmgr_is_buffer_referenced(struct r300_winsys_cs *rcs,
+                                               struct r300_winsys_buffer *_buf,
                                                enum r300_reference_domain domain)
 {
-    struct radeon_drm_buffer *buf = get_drm_buffer(_buf);
+    struct radeon_libdrm_cs *cs = radeon_libdrm_cs(rcs);
+    struct radeon_drm_buffer *buf = get_drm_buffer(radeon_pb_buffer(_buf));
     uint32_t tmp;
 
     if (domain & R300_REF_CS) {
-        if (radeon_bo_is_referenced_by_cs(buf->bo, buf->mgr->rws->cs)) {
+        if (radeon_bo_is_referenced_by_cs(buf->bo, cs->cs)) {
             return TRUE;
         }
     }
@@ -387,7 +440,6 @@ boolean radeon_drm_bufmgr_is_buffer_referenced(struct pb_buffer *_buf,
     return FALSE;
 }
 
-
 void radeon_drm_bufmgr_flush_maps(struct pb_manager *_mgr)
 {
     struct radeon_drm_bufmgr *mgr = radeon_drm_bufmgr(_mgr);
@@ -402,9 +454,10 @@ void radeon_drm_bufmgr_flush_maps(struct pb_manager *_mgr)
     make_empty_list(&mgr->buffer_map_list);
 }
 
-void radeon_drm_bufmgr_wait(struct pb_buffer *_buf)
+void radeon_drm_bufmgr_wait(struct r300_winsys_screen *ws,
+                            struct r300_winsys_buffer *_buf)
 {
-    struct radeon_drm_buffer *buf = get_drm_buffer(_buf);
+    struct radeon_drm_buffer *buf = get_drm_buffer(radeon_pb_buffer(_buf));
 
     radeon_bo_wait(buf->bo);
 }
diff --git a/src/gallium/winsys/radeon/drm/radeon_r300.c b/src/gallium/winsys/radeon/drm/radeon_r300.c
index af35497fd7c..4e89c070831 100644
--- a/src/gallium/winsys/radeon/drm/radeon_r300.c
+++ b/src/gallium/winsys/radeon/drm/radeon_r300.c
@@ -27,37 +27,69 @@
 #include "radeon_cs_gem.h"
 #include "state_tracker/drm_driver.h"
 
+#include "util/u_memory.h"
+
+static unsigned get_pb_usage_from_create_flags(unsigned bind, unsigned usage,
+                                               enum r300_buffer_domain domain)
+{
+    unsigned res = 0;
+
+    if (bind & (PIPE_BIND_DEPTH_STENCIL | PIPE_BIND_RENDER_TARGET |
+                PIPE_BIND_DISPLAY_TARGET | PIPE_BIND_SCANOUT))
+        res |= PB_USAGE_GPU_WRITE;
+
+    if (bind & PIPE_BIND_SAMPLER_VIEW)
+        res |= PB_USAGE_GPU_READ | PB_USAGE_GPU_WRITE;
+
+    if (bind & (PIPE_BIND_VERTEX_BUFFER | PIPE_BIND_INDEX_BUFFER))
+        res |= PB_USAGE_GPU_READ;
+
+    if (bind & PIPE_BIND_TRANSFER_WRITE)
+        res |= PB_USAGE_CPU_WRITE;
+
+    if (bind & PIPE_BIND_TRANSFER_READ)
+        res |= PB_USAGE_CPU_READ;
+
+    /* Is usage of any use for us? Probably not. */
+
+    /* Now add driver-specific usage flags. */
+    if (bind & (PIPE_BIND_VERTEX_BUFFER | PIPE_BIND_INDEX_BUFFER))
+        res |= RADEON_PB_USAGE_VERTEX;
+
+    if (domain & R300_DOMAIN_GTT)
+        res |= RADEON_PB_USAGE_DOMAIN_GTT;
+
+    if (domain & R300_DOMAIN_VRAM)
+        res |= RADEON_PB_USAGE_DOMAIN_VRAM;
+
+    return res;
+}
+
 static struct r300_winsys_buffer *
 radeon_r300_winsys_buffer_create(struct r300_winsys_screen *rws,
-				 unsigned alignment,
-				 unsigned usage,
-                                 enum r300_buffer_domain domain,
-				 unsigned size)
+                                 unsigned size,
+                                 unsigned alignment,
+                                 unsigned bind,
+                                 unsigned usage,
+                                 enum r300_buffer_domain domain)
 {
-    struct radeon_libdrm_winsys *ws = radeon_winsys_screen(rws);
+    struct radeon_libdrm_winsys *ws = radeon_libdrm_winsys(rws);
     struct pb_desc desc;
     struct pb_manager *provider;
     struct pb_buffer *buffer;
 
-    /* XXX this is hackish, but it's the only way to pass these flags
-     * to the real create function. */
-    usage &= ~(RADEON_USAGE_DOMAIN_GTT | RADEON_USAGE_DOMAIN_VRAM);
-    if (domain & R300_DOMAIN_GTT)
-        usage |= RADEON_USAGE_DOMAIN_GTT;
-    if (domain & R300_DOMAIN_VRAM)
-        usage |= RADEON_USAGE_DOMAIN_VRAM;
-
     memset(&desc, 0, sizeof(desc));
     desc.alignment = alignment;
-    desc.usage = usage;
+    desc.usage = get_pb_usage_from_create_flags(bind, usage, domain);
 
-    if (usage & PIPE_BIND_CONSTANT_BUFFER)
+    /* Assign a buffer manager. */
+    if (bind & PIPE_BIND_CONSTANT_BUFFER)
         provider = ws->mman;
-    else if ((usage & PIPE_BIND_VERTEX_BUFFER) ||
-	     (usage & PIPE_BIND_INDEX_BUFFER))
+    else if (bind & (PIPE_BIND_VERTEX_BUFFER | PIPE_BIND_INDEX_BUFFER))
 	provider = ws->cman;
     else
         provider = ws->kman;
+
     buffer = provider->create_buffer(provider, size, &desc);
     if (!buffer)
 	return NULL;
@@ -65,55 +97,6 @@ radeon_r300_winsys_buffer_create(struct r300_winsys_screen *rws,
     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,
-						  enum r300_buffer_tiling microtiled,
-						  enum r300_buffer_tiling macrotiled)
-{
-    struct pb_buffer *_buf = radeon_pb_buffer(buf);
-    radeon_drm_bufmgr_set_tiling(_buf, microtiled, macrotiled, pitch);
-}
-
-static void radeon_r300_winsys_buffer_get_tiling(struct r300_winsys_screen *rws,
-						  struct r300_winsys_buffer *buf,
-						  enum r300_buffer_tiling *microtiled,
-						  enum r300_buffer_tiling *macrotiled)
-{
-    struct pb_buffer *_buf = radeon_pb_buffer(buf);
-    radeon_drm_bufmgr_get_tiling(_buf, microtiled, macrotiled);
-}
-
-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, usage);
-}
-
-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_wait(struct r300_winsys_screen *ws,
-                                           struct r300_winsys_buffer *buf)
-{
-    struct pb_buffer *_buf = radeon_pb_buffer(buf);
-    radeon_drm_bufmgr_wait(_buf);
-}
-
 static void radeon_r300_winsys_buffer_reference(struct r300_winsys_screen *rws,
 						struct r300_winsys_buffer **pdst,
 						struct r300_winsys_buffer *src)
@@ -126,20 +109,11 @@ static void radeon_r300_winsys_buffer_reference(struct r300_winsys_screen *rws,
     *pdst = radeon_libdrm_winsys_buffer(_dst);
 }
 
-static boolean radeon_r300_winsys_is_buffer_referenced(struct r300_winsys_screen *rws,
-						       struct r300_winsys_buffer *buf,
-                                                       enum r300_reference_domain domain)
-{
-    struct pb_buffer *_buf = radeon_pb_buffer(buf);
-
-    return radeon_drm_bufmgr_is_buffer_referenced(_buf, domain);
-}
-
 static struct r300_winsys_buffer *radeon_r300_winsys_buffer_from_handle(struct r300_winsys_screen *rws,
                                                                         struct winsys_handle *whandle,
                                                                         unsigned *stride)
 {
-    struct radeon_libdrm_winsys *ws = radeon_winsys_screen(rws);
+    struct radeon_libdrm_winsys *ws = radeon_libdrm_winsys(rws);
     struct pb_buffer *_buf;
 
     *stride = whandle->stride;
@@ -150,111 +124,57 @@ static struct r300_winsys_buffer *radeon_r300_winsys_buffer_from_handle(struct r
 
 static boolean radeon_r300_winsys_buffer_get_handle(struct r300_winsys_screen *rws,
 						    struct r300_winsys_buffer *buffer,
-                                                    struct winsys_handle *whandle,
-                                                    unsigned stride)
+                                                    unsigned stride,
+                                                    struct winsys_handle *whandle)
 {
     struct pb_buffer *_buf = radeon_pb_buffer(buffer);
     whandle->stride = stride;
     return radeon_drm_bufmgr_get_handle(_buf, whandle);
 }
 
-static void radeon_set_flush_cb(struct r300_winsys_screen *rws,
-                                void (*flush_cb)(void *),
-                                void *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 r300_winsys_screen *rws,
-                                 struct r300_winsys_buffer *buf,
-                                 enum r300_buffer_domain rd,
-                                 enum r300_buffer_domain wd)
-{
-    struct pb_buffer *_buf = radeon_pb_buffer(buf);
-
-    return radeon_drm_bufmgr_add_buffer(_buf, rd, wd);
-}
-
-static boolean radeon_validate(struct r300_winsys_screen *rws)
+static void radeon_r300_winsys_cs_set_flush(struct r300_winsys_cs *rcs,
+                                            void (*flush)(void *),
+                                            void *user)
 {
-    struct radeon_libdrm_winsys *ws = radeon_winsys_screen(rws);
-    if (radeon_cs_space_check(ws->cs) < 0) {
-        return FALSE;
-    }
-
-    /* Things are fine, we can proceed as normal. */
-    return TRUE;
+    struct radeon_libdrm_cs *cs = radeon_libdrm_cs(rcs);
+    cs->flush_cs = flush;
+    cs->flush_data = user;
+    radeon_cs_space_set_flush(cs->cs, flush, user);
 }
 
-static unsigned radeon_get_cs_free_dwords(struct r300_winsys_screen *rws)
+static boolean radeon_r300_winsys_cs_validate(struct r300_winsys_cs *rcs)
 {
-    struct radeon_libdrm_winsys *ws = radeon_winsys_screen(rws);
-    struct radeon_cs *cs = ws->cs;
+    struct radeon_libdrm_cs *cs = radeon_libdrm_cs(rcs);
 
-    return cs->ndw - cs->cdw;
+    return radeon_cs_space_check(cs->cs) >= 0;
 }
 
-static uint32_t *radeon_get_cs_pointer(struct r300_winsys_screen *rws,
-                                       unsigned count)
+static void radeon_r300_winsys_cs_reset_buffers(struct r300_winsys_cs *rcs)
 {
-    struct radeon_libdrm_winsys *ws = radeon_winsys_screen(rws);
-    struct radeon_cs *cs = ws->cs;
-    uint32_t *ptr = cs->packets + cs->cdw;
-
-    cs->cdw += count;
-    return ptr;
+    struct radeon_libdrm_cs *cs = radeon_libdrm_cs(rcs);
+    radeon_cs_space_reset_bos(cs->cs);
 }
 
-static void radeon_write_cs_dword(struct r300_winsys_screen *rws,
-                                  uint32_t dword)
+static void radeon_r300_winsys_cs_flush(struct r300_winsys_cs *rcs)
 {
-    struct radeon_libdrm_winsys *ws = radeon_winsys_screen(rws);
-    radeon_cs_write_dword(ws->cs, dword);
-}
-
-static void radeon_write_cs_table(struct r300_winsys_screen *rws,
-                                  const void *table, unsigned count)
-{
-    struct radeon_libdrm_winsys *ws = radeon_winsys_screen(rws);
-    radeon_cs_write_table(ws->cs, table, count);
-}
-
-static void radeon_write_cs_reloc(struct r300_winsys_screen *rws,
-                                  struct r300_winsys_buffer *buf,
-                                  enum r300_buffer_domain rd,
-                                  enum r300_buffer_domain wd,
-                                  uint32_t flags)
-{
-    struct pb_buffer *_buf = radeon_pb_buffer(buf);
-    radeon_drm_bufmgr_write_reloc(_buf, rd, wd, flags);
-}
-
-static void radeon_reset_bos(struct r300_winsys_screen *rws)
-{
-    struct radeon_libdrm_winsys *ws = radeon_winsys_screen(rws);
-    radeon_cs_space_reset_bos(ws->cs);
-}
-
-static void radeon_flush_cs(struct r300_winsys_screen *rws)
-{
-    struct radeon_libdrm_winsys *ws = radeon_winsys_screen(rws);
+    struct radeon_libdrm_cs *cs = radeon_libdrm_cs(rcs);
     int retval;
 
     /* Don't flush a zero-sized CS. */
-    if (!ws->cs->cdw) {
+    if (!cs->base.cdw) {
         return;
     }
 
-    radeon_drm_bufmgr_flush_maps(ws->kman);
+    cs->cs->cdw = cs->base.cdw;
+
+    radeon_drm_bufmgr_flush_maps(cs->ws->kman);
+
     /* Emit the CS. */
-    retval = radeon_cs_emit(ws->cs);
+    retval = radeon_cs_emit(cs->cs);
     if (retval) {
         if (debug_get_bool_option("RADEON_DUMP_CS", FALSE)) {
             fprintf(stderr, "radeon: The kernel rejected CS, dumping...\n");
-            radeon_cs_print(ws->cs, stderr);
+            radeon_cs_print(cs->cs, stderr);
         } else {
             fprintf(stderr, "radeon: The kernel rejected CS, "
                             "see dmesg for more information.\n");
@@ -265,11 +185,15 @@ static void radeon_flush_cs(struct r300_winsys_screen *rws)
      * 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(ws->cs);
+    radeon_cs_erase(cs->cs);
+
+    cs->base.ptr = cs->cs->packets;
+    cs->base.cdw = cs->cs->cdw;
+    cs->base.ndw = cs->cs->ndw;
 }
 
 static uint32_t radeon_get_value(struct r300_winsys_screen *rws,
-			     enum r300_value_id id)
+                                 enum r300_value_id id)
 {
     struct radeon_libdrm_winsys *ws = (struct radeon_libdrm_winsys *)rws;
 
@@ -288,11 +212,42 @@ static uint32_t radeon_get_value(struct r300_winsys_screen *rws,
     return 0;
 }
 
-static void
-radeon_winsys_destroy(struct r300_winsys_screen *rws)
+static struct r300_winsys_cs *radeon_r300_winsys_cs_create(struct r300_winsys_screen *rws)
+{
+    struct radeon_libdrm_winsys *ws = radeon_libdrm_winsys(rws);
+    struct radeon_libdrm_cs *cs = CALLOC_STRUCT(radeon_libdrm_cs);
+
+    if (!cs)
+        return NULL;
+
+    /* Size limit on IBs is 64 kibibytes. */
+    cs->cs = radeon_cs_create(ws->csm, 1024 * 64 / 4);
+    if (!cs->cs) {
+        FREE(cs);
+        return NULL;
+    }
+
+    radeon_cs_set_limit(cs->cs,
+            RADEON_GEM_DOMAIN_GTT, ws->gart_size);
+    radeon_cs_set_limit(cs->cs,
+            RADEON_GEM_DOMAIN_VRAM, ws->vram_size);
+
+    cs->ws = ws;
+    cs->base.ptr = cs->cs->packets;
+    cs->base.cdw = cs->cs->cdw;
+    cs->base.ndw = cs->cs->ndw;
+    return &cs->base;
+}
+
+static void radeon_r300_winsys_cs_destroy(struct r300_winsys_cs *rcs)
+{
+    struct radeon_libdrm_cs *cs = radeon_libdrm_cs(rcs);
+    radeon_cs_destroy(cs->cs);
+}
+
+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);
@@ -302,10 +257,8 @@ radeon_winsys_destroy(struct r300_winsys_screen *rws)
     radeon_cs_manager_gem_dtor(ws->csm);
 }
 
-boolean
-radeon_setup_winsys(int fd, struct radeon_libdrm_winsys* ws)
+boolean radeon_setup_winsys(int fd, struct radeon_libdrm_winsys* ws)
 {
-    
     ws->csm = radeon_cs_manager_gem_ctor(fd);
     if (!ws->csm)
 	goto fail;
@@ -324,39 +277,28 @@ radeon_setup_winsys(int fd, struct radeon_libdrm_winsys* ws)
     if (!ws->mman)
 	goto fail;
 
-    /* Size limit on IBs is 64 kibibytes. */
-    ws->cs = radeon_cs_create(ws->csm, 1024 * 64 / 4);
-    if (!ws->cs)
-	goto fail;
-    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.get_cs_free_dwords = radeon_get_cs_free_dwords;
-    ws->base.get_cs_pointer = radeon_get_cs_pointer;
-    ws->base.write_cs_dword = radeon_write_cs_dword;
-    ws->base.write_cs_table = radeon_write_cs_table;
-    ws->base.write_cs_reloc = radeon_write_cs_reloc;
-    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_get_tiling = radeon_r300_winsys_buffer_get_tiling;
-    ws->base.buffer_map = radeon_r300_winsys_buffer_map;
-    ws->base.buffer_unmap = radeon_r300_winsys_buffer_unmap;
-    ws->base.buffer_wait = radeon_r300_winsys_buffer_wait;
+    ws->base.buffer_set_tiling = radeon_drm_bufmgr_set_tiling;
+    ws->base.buffer_get_tiling = radeon_drm_bufmgr_get_tiling;
+    ws->base.buffer_map = radeon_drm_buffer_map;
+    ws->base.buffer_unmap = radeon_drm_buffer_unmap;
+    ws->base.buffer_wait = radeon_drm_bufmgr_wait;
     ws->base.buffer_reference = radeon_r300_winsys_buffer_reference;
     ws->base.buffer_from_handle = radeon_r300_winsys_buffer_from_handle;
     ws->base.buffer_get_handle = radeon_r300_winsys_buffer_get_handle;
-    ws->base.is_buffer_referenced = radeon_r300_winsys_is_buffer_referenced;
+
+    ws->base.cs_create = radeon_r300_winsys_cs_create;
+    ws->base.cs_destroy = radeon_r300_winsys_cs_destroy;
+    ws->base.cs_add_buffer = radeon_drm_bufmgr_add_buffer;
+    ws->base.cs_validate = radeon_r300_winsys_cs_validate;
+    ws->base.cs_write_reloc = radeon_drm_bufmgr_write_reloc;
+    ws->base.cs_flush = radeon_r300_winsys_cs_flush;
+    ws->base.cs_reset_buffers = radeon_r300_winsys_cs_reset_buffers;
+    ws->base.cs_set_flush = radeon_r300_winsys_cs_set_flush;
+    ws->base.cs_is_buffer_referenced = radeon_drm_bufmgr_is_buffer_referenced;
     return TRUE;
 
 fail:
@@ -373,7 +315,5 @@ fail:
     if (ws->mman)
 	ws->mman->destroy(ws->mman);
 
-    if (ws->cs)
-	radeon_cs_destroy(ws->cs);
     return FALSE;
 }
diff --git a/src/gallium/winsys/radeon/drm/radeon_winsys.h b/src/gallium/winsys/radeon/drm/radeon_winsys.h
index ca789be8e93..9a25e7679c9 100644
--- a/src/gallium/winsys/radeon/drm/radeon_winsys.h
+++ b/src/gallium/winsys/radeon/drm/radeon_winsys.h
@@ -75,19 +75,32 @@ struct radeon_libdrm_winsys {
 
     /* Radeon CS manager. */
     struct radeon_cs_manager *csm;
+};
+
+struct radeon_libdrm_cs {
+    struct r300_winsys_cs base;
+
+    /* The winsys. */
+    struct radeon_libdrm_winsys *ws;
 
-    /* Current CS. */
+    /* The libdrm command stream. */
     struct radeon_cs *cs;
 
-    /* Flush CB */
-    void (*flush_cb)(void *);
+    /* Flush CS. */
+    void (*flush_cs)(void *);
     void *flush_data;
 };
 
+static INLINE struct radeon_libdrm_cs *
+radeon_libdrm_cs(struct r300_winsys_cs *base)
+{
+    return (struct radeon_libdrm_cs*)base;
+}
+
 static INLINE struct radeon_libdrm_winsys *
-radeon_winsys_screen(struct r300_winsys_screen *base)
+radeon_libdrm_winsys(struct r300_winsys_screen *base)
 {
-  return (struct radeon_libdrm_winsys *)base;
+    return (struct radeon_libdrm_winsys*)base;
 }
 
 #endif