diff --git a/libdrm/radeon/radeon_bo.h b/libdrm/radeon/radeon_bo.h
index 44dc090..3cabdfc 100644
--- a/libdrm/radeon/radeon_bo.h
+++ b/libdrm/radeon/radeon_bo.h
@@ -52,6 +52,7 @@ struct radeon_bo {
 #endif
     void                        *ptr;
     struct radeon_bo_manager    *bom;
+    uint32_t                    space_accounted;
 };
 
 /* bo functions */
@@ -66,6 +67,7 @@ struct radeon_bo_funcs {
     struct radeon_bo *(*bo_unref)(struct radeon_bo *bo);
     int (*bo_map)(struct radeon_bo *bo, int write);
     int (*bo_unmap)(struct radeon_bo *bo);
+    int (*bo_wait)(struct radeon_bo *bo);
 };
 
 struct radeon_bo_manager {
@@ -151,6 +153,14 @@ static inline int _radeon_bo_unmap(struct radeon_bo *bo,
     return bo->bom->funcs->bo_unmap(bo);
 }
 
+static inline int _radeon_bo_wait(struct radeon_bo *bo,
+                                  const char *file,
+                                  const char *func,
+                                  int line)
+{
+    return bo->bom->funcs->bo_wait(bo);
+}
+
 #define radeon_bo_open(bom, h, s, a, d, f)\
     _radeon_bo_open(bom, h, s, a, d, f, __FILE__, __FUNCTION__, __LINE__)
 #define radeon_bo_ref(bo)\
@@ -163,5 +173,7 @@ static inline int _radeon_bo_unmap(struct radeon_bo *bo,
     _radeon_bo_unmap(bo, __FILE__, __FUNCTION__, __LINE__)
 #define radeon_bo_debug(bo, opcode)\
     _radeon_bo_debug(bo, opcode, __FILE__, __FUNCTION__, __LINE__)
+#define radeon_bo_wait(bo) \
+    _radeon_bo_wait(bo, __FILE__, __func__, __LINE__)
 
 #endif
diff --git a/libdrm/radeon/radeon_bo_gem.c b/libdrm/radeon/radeon_bo_gem.c
index fdf852a..932afeb 100644
--- a/libdrm/radeon/radeon_bo_gem.c
+++ b/libdrm/radeon/radeon_bo_gem.c
@@ -35,6 +35,7 @@
 #include <string.h>
 #include <sys/mman.h>
 #include <sys/ioctl.h>
+#include <errno.h>
 #include "xf86drm.h"
 #include "drm.h"
 #include "radeon_drm.h"
@@ -176,12 +177,26 @@ static int bo_unmap(struct radeon_bo *bo)
     return 0;
 }
 
+static int bo_wait(struct radeon_bo *bo)
+{
+    struct drm_radeon_gem_wait_rendering args;
+    int ret;
+
+    args.handle = bo->handle;
+    do {
+        ret = drmCommandWriteRead(bo->bom->fd, DRM_RADEON_GEM_WAIT_RENDERING,
+                                  &args, sizeof(args));
+    } while (ret == -EAGAIN);
+    return ret;
+}
+
 static struct radeon_bo_funcs bo_gem_funcs = {
     bo_open,
     bo_ref,
     bo_unref,
     bo_map,
-    bo_unmap
+    bo_unmap,
+    bo_wait
 };
 
 struct radeon_bo_manager *radeon_bo_manager_gem_ctor(int fd)
diff --git a/libdrm/radeon/radeon_cs.h b/libdrm/radeon/radeon_cs.h
index 8b1a999..8b5a2a9 100644
--- a/libdrm/radeon/radeon_cs.h
+++ b/libdrm/radeon/radeon_cs.h
@@ -33,6 +33,8 @@
 #define RADEON_CS_H
 
 #include <stdint.h>
+#include "drm.h"
+#include "radeon_drm.h"
 #include "radeon_bo.h"
 
 struct radeon_cs_reloc {
@@ -42,6 +44,18 @@ struct radeon_cs_reloc {
     uint32_t            flags;
 };
 
+
+#define RADEON_CS_SPACE_OK 0
+#define RADEON_CS_SPACE_OP_TO_BIG 1
+#define RADEON_CS_SPACE_FLUSH 2
+
+struct radeon_cs_space_check {
+    struct radeon_bo *bo;
+    uint32_t read_domains;
+    uint32_t write_domain;
+    uint32_t new_accounted;
+};
+
 struct radeon_cs_manager;
 
 struct radeon_cs {
@@ -58,6 +72,7 @@ struct radeon_cs {
     const char                  *section_file;
     const char                  *section_func;
     int                         section_line;
+
 };
 
 /* cs functions */
@@ -84,11 +99,16 @@ struct radeon_cs_funcs {
     int (*cs_erase)(struct radeon_cs *cs);
     int (*cs_need_flush)(struct radeon_cs *cs);
     void (*cs_print)(struct radeon_cs *cs, FILE *file);
+    int (*cs_space_check)(struct radeon_cs *cs, struct radeon_cs_space_check *bos,
+			  int num_bo);
 };
 
 struct radeon_cs_manager {
     struct radeon_cs_funcs  *funcs;
     int                     fd;
+    uint32_t vram_limit, gart_limit;
+    uint32_t vram_write_used, gart_write_used;
+    uint32_t read_used;
 };
 
 static inline struct radeon_cs *radeon_cs_create(struct radeon_cs_manager *csm,
@@ -157,4 +177,19 @@ static inline void radeon_cs_print(struct radeon_cs *cs, FILE *file)
     cs->csm->funcs->cs_print(cs, file);
 }
 
+static inline int radeon_cs_space_check(struct radeon_cs *cs,
+					    struct radeon_cs_space_check *bos,
+					    int num_bo)
+{
+    return cs->csm->funcs->cs_space_check(cs, bos, num_bo);
+}
+
+static inline void radeon_cs_set_limit(struct radeon_cs *cs, uint32_t domain, uint32_t limit)
+{
+    
+    if (domain == RADEON_GEM_DOMAIN_VRAM)
+	cs->csm->vram_limit = limit;
+    else
+	cs->csm->gart_limit = limit;
+}
 #endif
diff --git a/libdrm/radeon/radeon_cs_gem.c b/libdrm/radeon/radeon_cs_gem.c
index bfc5aca..c9cf345 100644
--- a/libdrm/radeon/radeon_cs_gem.c
+++ b/libdrm/radeon/radeon_cs_gem.c
@@ -229,6 +229,16 @@ static int cs_gem_end(struct radeon_cs *cs,
     return 0;
 }
 
+
+static void dump_cmdbuf(struct radeon_cs *cs)
+{
+  int i;
+  for (i = 0; i < cs->cdw; i++){
+    fprintf(stderr,"%x: %08x\n", i, cs->packets[i]);
+  }
+
+}
+
 static int cs_gem_emit(struct radeon_cs *cs)
 {
     struct cs_gem *csg = (struct cs_gem*)cs;
@@ -236,6 +246,7 @@ static int cs_gem_emit(struct radeon_cs *cs)
     unsigned i;
     int r;
 
+    dump_cmdbuf(cs);
     csg->chunks[0].length_dw = cs->cdw;
 
     chunk_array[0] = (uint64_t)(intptr_t)&csg->chunks[0];
@@ -387,6 +398,97 @@ static void cs_gem_print(struct radeon_cs *cs, FILE *file)
     }
 }
 
+static int cs_gem_check_space(struct radeon_cs *cs, struct radeon_cs_space_check *bos, int num_bo)
+{
+    struct radeon_cs_manager *csm = cs->csm;
+    int this_op_read = 0, this_op_gart_write = 0, this_op_vram_write = 0;
+    uint32_t read_domains, write_domain;
+    int i;
+    struct radeon_bo *bo;
+
+    /* check the totals for this operation */
+
+    if (num_bo == 0)
+        return 0;
+
+    /* prepare */
+    for (i = 0; i < num_bo; i++) {
+      bo = bos[i].bo;
+
+      bos[i].new_accounted = 0;
+      read_domains = bos[i].read_domains;
+      write_domain = bos[i].write_domain;
+		
+      /* already accounted this bo */
+      if (write_domain && (write_domain == bo->space_accounted))
+	  continue;
+
+      if (read_domains && ((read_domains << 16) == bo->space_accounted))
+	  continue;
+      
+      if (bo->space_accounted == 0) {
+	  if (write_domain == RADEON_GEM_DOMAIN_VRAM)
+	      this_op_vram_write += bo->size;
+	  else if (write_domain == RADEON_GEM_DOMAIN_GTT)
+	      this_op_gart_write += bo->size;
+	  else
+	      this_op_read += bo->size;
+	  bos[i].new_accounted = (read_domains << 16) | write_domain;
+      } else {
+	  uint16_t old_read, old_write;
+	  
+	  old_read = bo->space_accounted >> 16;
+	  old_write = bo->space_accounted & 0xffff;
+
+	  if (write_domain && (old_read & write_domain)) {
+	      bos[i].new_accounted = write_domain;
+	      /* moving from read to a write domain */
+	      if (write_domain == RADEON_GEM_DOMAIN_VRAM) {
+		  this_op_read -= bo->size;
+		  this_op_vram_write += bo->size;
+	      } else if (write_domain == RADEON_GEM_DOMAIN_VRAM) {
+		  this_op_read -= bo->size;
+		  this_op_gart_write += bo->size;
+	      }
+	  } else if (read_domains & old_write) {
+	      bos[i].new_accounted = bo->space_accounted & 0xffff;
+	  } else {
+	      /* rewrite the domains */
+	      if (write_domain != old_write)
+		  fprintf(stderr,"WRITE DOMAIN RELOC FAILURE 0x%x %d %d\n", bo->handle, write_domain, old_write);
+	      if (read_domains != old_read)
+		  fprintf(stderr,"READ DOMAIN RELOC FAILURE 0x%x %d %d\n", bo->handle, read_domains, old_read);
+	      return RADEON_CS_SPACE_FLUSH;
+	  }
+      }
+	}
+	
+	if (this_op_read < 0)
+		this_op_read = 0;
+
+	/* check sizes - operation first */
+	if ((this_op_read + this_op_gart_write > csm->gart_limit) ||
+	    (this_op_vram_write > csm->vram_limit)) {
+	    return RADEON_CS_SPACE_OP_TO_BIG;
+	}
+
+	if (((csm->vram_write_used + this_op_vram_write) > csm->vram_limit) ||
+	    ((csm->read_used + csm->gart_write_used + this_op_gart_write + this_op_read) > csm->gart_limit)) {
+		return RADEON_CS_SPACE_FLUSH;
+	}
+
+	csm->gart_write_used += this_op_gart_write;
+	csm->vram_write_used += this_op_vram_write;
+	csm->read_used += this_op_read;
+	/* commit */
+	for (i = 0; i < num_bo; i++) {
+		bo = bos[i].bo;
+		bo->space_accounted = bos[i].new_accounted;
+	}
+
+	return RADEON_CS_SPACE_OK;
+}
+
 static struct radeon_cs_funcs radeon_cs_gem_funcs = {
     cs_gem_create,
     cs_gem_write_dword,
@@ -397,7 +499,8 @@ static struct radeon_cs_funcs radeon_cs_gem_funcs = {
     cs_gem_destroy,
     cs_gem_erase,
     cs_gem_need_flush,
-    cs_gem_print
+    cs_gem_print,
+    cs_gem_check_space,
 };
 
 struct radeon_cs_manager *radeon_cs_manager_gem_ctor(int fd)
diff --git a/shared-core/radeon_drm.h b/shared-core/radeon_drm.h
index fd26eee..b989649 100644
--- a/shared-core/radeon_drm.h
+++ b/shared-core/radeon_drm.h
@@ -711,6 +711,7 @@ typedef struct drm_radeon_indirect {
 #define RADEON_PARAM_FB_LOCATION           14   /* FB location */
 #define RADEON_PARAM_NUM_GB_PIPES          15   /* num GB pipes */
 #define RADEON_PARAM_KERNEL_MM             16
+#define RADEON_PARAM_DEVICE_ID            17
 
 typedef struct drm_radeon_getparam {
 	int param;