svga/winsys: implement GBS support

This is a squash commit of many commits by Thomas Hellstrom.

Reviewed-by: Thomas Hellstrom <thellstrom@vmware.com>
Cc: "10.1" <mesa-stable@lists.freedesktop.org>

19 files changed, 3064 insertions, 323 deletions

diff --git a/src/gallium/winsys/svga/drm/Makefile.sources b/src/gallium/winsys/svga/drm/Makefile.sources
index c300860ff79..08aecd6bfaf 100644
--- a/src/gallium/winsys/svga/drm/Makefile.sources
+++ b/src/gallium/winsys/svga/drm/Makefile.sources
@@ -7,4 +7,6 @@ C_SOURCES := \
         vmw_screen_ioctl.c  \
         vmw_screen_pools.c  \
         vmw_screen_svga.c  \
-        vmw_surface.c
+        vmw_surface.c \
+        vmw_shader.c \
+        pb_buffer_simple_fenced.c
diff --git a/src/gallium/winsys/svga/drm/pb_buffer_simple_fenced.c b/src/gallium/winsys/svga/drm/pb_buffer_simple_fenced.c
new file mode 100644
index 00000000000..888aebb2bdc
--- /dev/null
+++ b/src/gallium/winsys/svga/drm/pb_buffer_simple_fenced.c
@@ -0,0 +1,844 @@
+/**************************************************************************
+ *
+ * Copyright 2007-2010 VMware, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
+ * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+ * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+ * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ **************************************************************************/
+
+/**
+ * \file
+ * Implementation of fenced buffers.
+ *
+ * \author Jose Fonseca <jfonseca-at-vmware-dot-com>
+ * \author Thomas Hellström <thellstrom-at-vmware-dot-com>
+ */
+
+
+#include "pipe/p_config.h"
+
+#if defined(PIPE_OS_LINUX) || defined(PIPE_OS_BSD) || defined(PIPE_OS_SOLARIS)
+#include <unistd.h>
+#include <sched.h>
+#endif
+
+#include "pipe/p_compiler.h"
+#include "pipe/p_defines.h"
+#include "util/u_debug.h"
+#include "os/os_thread.h"
+#include "util/u_memory.h"
+#include "util/u_double_list.h"
+
+#include "pipebuffer/pb_buffer.h"
+#include "pipebuffer/pb_bufmgr.h"
+#include "pipebuffer/pb_buffer_fenced.h"
+#include "vmw_screen.h"
+
+
+/**
+ * Convenience macro (type safe).
+ */
+#define SUPER(__derived) (&(__derived)->base)
+
+
+struct fenced_manager
+{
+   struct pb_manager base;
+   struct pb_manager *provider;
+   struct pb_fence_ops *ops;
+
+   /**
+    * Following members are mutable and protected by this mutex.
+    */
+   pipe_mutex mutex;
+
+   /**
+    * Fenced buffer list.
+    *
+    * All fenced buffers are placed in this listed, ordered from the oldest
+    * fence to the newest fence.
+    */
+   struct list_head fenced;
+   pb_size num_fenced;
+
+   struct list_head unfenced;
+   pb_size num_unfenced;
+
+};
+
+
+/**
+ * Fenced buffer.
+ *
+ * Wrapper around a pipe buffer which adds fencing and reference counting.
+ */
+struct fenced_buffer
+{
+   /*
+    * Immutable members.
+    */
+
+   struct pb_buffer base;
+   struct fenced_manager *mgr;
+
+   /*
+    * Following members are mutable and protected by fenced_manager::mutex.
+    */
+
+   struct list_head head;
+
+   /**
+    * Buffer with storage.
+    */
+   struct pb_buffer *buffer;
+   pb_size size;
+
+   /**
+    * A bitmask of PB_USAGE_CPU/GPU_READ/WRITE describing the current
+    * buffer usage.
+    */
+   unsigned flags;
+
+   unsigned mapcount;
+
+   struct pb_validate *vl;
+   unsigned validation_flags;
+
+   struct pipe_fence_handle *fence;
+};
+
+
+static INLINE struct fenced_manager *
+fenced_manager(struct pb_manager *mgr)
+{
+   assert(mgr);
+   return (struct fenced_manager *)mgr;
+}
+
+
+static INLINE struct fenced_buffer *
+fenced_buffer(struct pb_buffer *buf)
+{
+   assert(buf);
+   return (struct fenced_buffer *)buf;
+}
+
+
+static void
+fenced_buffer_destroy_gpu_storage_locked(struct fenced_buffer *fenced_buf);
+
+static enum pipe_error
+fenced_buffer_create_gpu_storage_locked(struct fenced_manager *fenced_mgr,
+                                        struct fenced_buffer *fenced_buf,
+                                        const struct pb_desc *desc,
+                                        boolean wait);
+/**
+ * Dump the fenced buffer list.
+ *
+ * Useful to understand failures to allocate buffers.
+ */
+static void
+fenced_manager_dump_locked(struct fenced_manager *fenced_mgr)
+{
+#ifdef DEBUG
+   struct pb_fence_ops *ops = fenced_mgr->ops;
+   struct list_head *curr, *next;
+   struct fenced_buffer *fenced_buf;
+
+   debug_printf("%10s %7s %8s %7s %10s %s\n",
+                "buffer", "size", "refcount", "storage", "fence", "signalled");
+
+   curr = fenced_mgr->unfenced.next;
+   next = curr->next;
+   while(curr != &fenced_mgr->unfenced) {
+      fenced_buf = LIST_ENTRY(struct fenced_buffer, curr, head);
+      assert(!fenced_buf->fence);
+      debug_printf("%10p %7u %8u %7s\n",
+                   (void *) fenced_buf,
+                   fenced_buf->base.size,
+                   p_atomic_read(&fenced_buf->base.reference.count),
+                   fenced_buf->buffer ? "gpu" : "none");
+      curr = next;
+      next = curr->next;
+   }
+
+   curr = fenced_mgr->fenced.next;
+   next = curr->next;
+   while(curr != &fenced_mgr->fenced) {
+      int signaled;
+      fenced_buf = LIST_ENTRY(struct fenced_buffer, curr, head);
+      assert(fenced_buf->buffer);
+      signaled = ops->fence_signalled(ops, fenced_buf->fence, 0);
+      debug_printf("%10p %7u %8u %7s %10p %s\n",
+                   (void *) fenced_buf,
+                   fenced_buf->base.size,
+                   p_atomic_read(&fenced_buf->base.reference.count),
+                   "gpu",
+                   (void *) fenced_buf->fence,
+                   signaled == 0 ? "y" : "n");
+      curr = next;
+      next = curr->next;
+   }
+#else
+   (void)fenced_mgr;
+#endif
+}
+
+
+static INLINE void
+fenced_buffer_destroy_locked(struct fenced_manager *fenced_mgr,
+                             struct fenced_buffer *fenced_buf)
+{
+   assert(!pipe_is_referenced(&fenced_buf->base.reference));
+
+   assert(!fenced_buf->fence);
+   assert(fenced_buf->head.prev);
+   assert(fenced_buf->head.next);
+   LIST_DEL(&fenced_buf->head);
+   assert(fenced_mgr->num_unfenced);
+   --fenced_mgr->num_unfenced;
+
+   fenced_buffer_destroy_gpu_storage_locked(fenced_buf);
+
+   FREE(fenced_buf);
+}
+
+
+/**
+ * Add the buffer to the fenced list.
+ *
+ * Reference count should be incremented before calling this function.
+ */
+static INLINE void
+fenced_buffer_add_locked(struct fenced_manager *fenced_mgr,
+                         struct fenced_buffer *fenced_buf)
+{
+   assert(pipe_is_referenced(&fenced_buf->base.reference));
+   assert(fenced_buf->flags & PB_USAGE_GPU_READ_WRITE);
+   assert(fenced_buf->fence);
+
+   p_atomic_inc(&fenced_buf->base.reference.count);
+
+   LIST_DEL(&fenced_buf->head);
+   assert(fenced_mgr->num_unfenced);
+   --fenced_mgr->num_unfenced;
+   LIST_ADDTAIL(&fenced_buf->head, &fenced_mgr->fenced);
+   ++fenced_mgr->num_fenced;
+}
+
+
+/**
+ * Remove the buffer from the fenced list, and potentially destroy the buffer
+ * if the reference count reaches zero.
+ *
+ * Returns TRUE if the buffer was detroyed.
+ */
+static INLINE boolean
+fenced_buffer_remove_locked(struct fenced_manager *fenced_mgr,
+                            struct fenced_buffer *fenced_buf)
+{
+   struct pb_fence_ops *ops = fenced_mgr->ops;
+
+   assert(fenced_buf->fence);
+   assert(fenced_buf->mgr == fenced_mgr);
+
+   ops->fence_reference(ops, &fenced_buf->fence, NULL);
+   fenced_buf->flags &= ~PB_USAGE_GPU_READ_WRITE;
+
+   assert(fenced_buf->head.prev);
+   assert(fenced_buf->head.next);
+
+   LIST_DEL(&fenced_buf->head);
+   assert(fenced_mgr->num_fenced);
+   --fenced_mgr->num_fenced;
+
+   LIST_ADDTAIL(&fenced_buf->head, &fenced_mgr->unfenced);
+   ++fenced_mgr->num_unfenced;
+
+   if (p_atomic_dec_zero(&fenced_buf->base.reference.count)) {
+      fenced_buffer_destroy_locked(fenced_mgr, fenced_buf);
+      return TRUE;
+   }
+
+   return FALSE;
+}
+
+
+/**
+ * Wait for the fence to expire, and remove it from the fenced list.
+ *
+ * This function will release and re-aquire the mutex, so any copy of mutable
+ * state must be discarded after calling it.
+ */
+static INLINE enum pipe_error
+fenced_buffer_finish_locked(struct fenced_manager *fenced_mgr,
+                            struct fenced_buffer *fenced_buf)
+{
+   struct pb_fence_ops *ops = fenced_mgr->ops;
+   enum pipe_error ret = PIPE_ERROR;
+
+#if 0
+   debug_warning("waiting for GPU");
+#endif
+
+   assert(pipe_is_referenced(&fenced_buf->base.reference));
+   assert(fenced_buf->fence);
+
+   if(fenced_buf->fence) {
+      struct pipe_fence_handle *fence = NULL;
+      int finished;
+      boolean proceed;
+
+      ops->fence_reference(ops, &fence, fenced_buf->fence);
+
+      pipe_mutex_unlock(fenced_mgr->mutex);
+
+      finished = ops->fence_finish(ops, fenced_buf->fence, 0);
+
+      pipe_mutex_lock(fenced_mgr->mutex);
+
+      assert(pipe_is_referenced(&fenced_buf->base.reference));
+
+      /*
+       * Only proceed if the fence object didn't change in the meanwhile.
+       * Otherwise assume the work has been already carried out by another
+       * thread that re-aquired the lock before us.
+       */
+      proceed = fence == fenced_buf->fence ? TRUE : FALSE;
+
+      ops->fence_reference(ops, &fence, NULL);
+
+      if(proceed && finished == 0) {
+         /*
+          * Remove from the fenced list
+          */
+
+         boolean destroyed;
+
+         destroyed = fenced_buffer_remove_locked(fenced_mgr, fenced_buf);
+
+         /* TODO: remove consequents buffers with the same fence? */
+
+         assert(!destroyed);
+
+         fenced_buf->flags &= ~PB_USAGE_GPU_READ_WRITE;
+
+         ret = PIPE_OK;
+      }
+   }
+
+   return ret;
+}
+
+
+/**
+ * Remove as many fenced buffers from the fenced list as possible.
+ *
+ * Returns TRUE if at least one buffer was removed.
+ */
+static boolean
+fenced_manager_check_signalled_locked(struct fenced_manager *fenced_mgr,
+                                      boolean wait)
+{
+   struct pb_fence_ops *ops = fenced_mgr->ops;
+   struct list_head *curr, *next;
+   struct fenced_buffer *fenced_buf;
+   struct pipe_fence_handle *prev_fence = NULL;
+   boolean ret = FALSE;
+
+   curr = fenced_mgr->fenced.next;
+   next = curr->next;
+   while(curr != &fenced_mgr->fenced) {
+      fenced_buf = LIST_ENTRY(struct fenced_buffer, curr, head);
+
+      if(fenced_buf->fence != prev_fence) {
+         int signaled;
+
+         if (wait) {
+            signaled = ops->fence_finish(ops, fenced_buf->fence, 0);
+
+            /*
+             * Don't return just now. Instead preemptively check if the
+             * following buffers' fences already expired,
+             * without further waits.
+             */
+            wait = FALSE;
+         }
+         else {
+            signaled = ops->fence_signalled(ops, fenced_buf->fence, 0);
+         }
+
+         if (signaled != 0) {
+            return ret;
+         }
+
+         prev_fence = fenced_buf->fence;
+      }
+      else {
+         /* This buffer's fence object is identical to the previous buffer's
+          * fence object, so no need to check the fence again.
+          */
+         assert(ops->fence_signalled(ops, fenced_buf->fence, 0) == 0);
+      }
+
+      fenced_buffer_remove_locked(fenced_mgr, fenced_buf);
+
+      ret = TRUE;
+
+      curr = next;
+      next = curr->next;
+   }
+
+   return ret;
+}
+
+
+/**
+ * Destroy the GPU storage.
+ */
+static void
+fenced_buffer_destroy_gpu_storage_locked(struct fenced_buffer *fenced_buf)
+{
+   if(fenced_buf->buffer) {
+      pb_reference(&fenced_buf->buffer, NULL);
+   }
+}
+
+
+/**
+ * Try to create GPU storage for this buffer.
+ *
+ * This function is a shorthand around pb_manager::create_buffer for
+ * fenced_buffer_create_gpu_storage_locked()'s benefit.
+ */
+static INLINE boolean
+fenced_buffer_try_create_gpu_storage_locked(struct fenced_manager *fenced_mgr,
+                                            struct fenced_buffer *fenced_buf,
+                                            const struct pb_desc *desc)
+{
+   struct pb_manager *provider = fenced_mgr->provider;
+
+   assert(!fenced_buf->buffer);
+
+   fenced_buf->buffer = provider->create_buffer(fenced_mgr->provider,
+                                                fenced_buf->size, desc);
+   return fenced_buf->buffer ? TRUE : FALSE;
+}
+
+
+/**
+ * Create GPU storage for this buffer.
+ */
+static enum pipe_error
+fenced_buffer_create_gpu_storage_locked(struct fenced_manager *fenced_mgr,
+                                        struct fenced_buffer *fenced_buf,
+                                        const struct pb_desc *desc,
+                                        boolean wait)
+{
+   assert(!fenced_buf->buffer);
+
+   /*
+    * Check for signaled buffers before trying to allocate.
+    */
+   fenced_manager_check_signalled_locked(fenced_mgr, FALSE);
+
+   fenced_buffer_try_create_gpu_storage_locked(fenced_mgr, fenced_buf, desc);
+
+   /*
+    * Keep trying while there is some sort of progress:
+    * - fences are expiring,
+    * - or buffers are being being swapped out from GPU memory into CPU memory.
+    */
+   while(!fenced_buf->buffer &&
+         (fenced_manager_check_signalled_locked(fenced_mgr, FALSE))) {
+     fenced_buffer_try_create_gpu_storage_locked(fenced_mgr, fenced_buf,
+                                                 desc);
+   }
+
+   if(!fenced_buf->buffer && wait) {
+      /*
+       * Same as before, but this time around, wait to free buffers if
+       * necessary.
+       */
+      while(!fenced_buf->buffer &&
+            (fenced_manager_check_signalled_locked(fenced_mgr, TRUE))) {
+        fenced_buffer_try_create_gpu_storage_locked(fenced_mgr, fenced_buf,
+                                                    desc);
+      }
+   }
+
+   if(!fenced_buf->buffer) {
+      if(0)
+         fenced_manager_dump_locked(fenced_mgr);
+
+      /* give up */
+      return PIPE_ERROR_OUT_OF_MEMORY;
+   }
+
+   return PIPE_OK;
+}
+
+
+static void
+fenced_buffer_destroy(struct pb_buffer *buf)
+{
+   struct fenced_buffer *fenced_buf = fenced_buffer(buf);
+   struct fenced_manager *fenced_mgr = fenced_buf->mgr;
+
+   assert(!pipe_is_referenced(&fenced_buf->base.reference));
+
+   pipe_mutex_lock(fenced_mgr->mutex);
+
+   fenced_buffer_destroy_locked(fenced_mgr, fenced_buf);
+
+   pipe_mutex_unlock(fenced_mgr->mutex);
+}
+
+
+static void *
+fenced_buffer_map(struct pb_buffer *buf,
+                  unsigned flags, void *flush_ctx)
+{
+   struct fenced_buffer *fenced_buf = fenced_buffer(buf);
+   struct fenced_manager *fenced_mgr = fenced_buf->mgr;
+   struct pb_fence_ops *ops = fenced_mgr->ops;
+   void *map = NULL;
+
+   pipe_mutex_lock(fenced_mgr->mutex);
+
+   assert(!(flags & PB_USAGE_GPU_READ_WRITE));
+
+   /*
+    * Serialize writes.
+    */
+   while((fenced_buf->flags & PB_USAGE_GPU_WRITE) ||
+         ((fenced_buf->flags & PB_USAGE_GPU_READ) &&
+          (flags & PB_USAGE_CPU_WRITE))) {
+
+      /*
+       * Don't wait for the GPU to finish accessing it,
+       * if blocking is forbidden.
+       */
+      if((flags & PB_USAGE_DONTBLOCK) &&
+          ops->fence_signalled(ops, fenced_buf->fence, 0) != 0) {
+         goto done;
+      }
+
+      if (flags & PB_USAGE_UNSYNCHRONIZED) {
+         break;
+      }
+
+      /*
+       * Wait for the GPU to finish accessing. This will release and re-acquire
+       * the mutex, so all copies of mutable state must be discarded.
+       */
+      fenced_buffer_finish_locked(fenced_mgr, fenced_buf);
+   }
+
+   map = pb_map(fenced_buf->buffer, flags, flush_ctx);
+
+   if(map) {
+      ++fenced_buf->mapcount;
+      fenced_buf->flags |= flags & PB_USAGE_CPU_READ_WRITE;
+   }
+
+done:
+   pipe_mutex_unlock(fenced_mgr->mutex);
+
+   return map;
+}
+
+
+static void
+fenced_buffer_unmap(struct pb_buffer *buf)
+{
+   struct fenced_buffer *fenced_buf = fenced_buffer(buf);
+   struct fenced_manager *fenced_mgr = fenced_buf->mgr;
+
+   pipe_mutex_lock(fenced_mgr->mutex);
+
+   assert(fenced_buf->mapcount);
+   if(fenced_buf->mapcount) {
+      if (fenced_buf->buffer)
+         pb_unmap(fenced_buf->buffer);
+      --fenced_buf->mapcount;
+      if(!fenced_buf->mapcount)
+         fenced_buf->flags &= ~PB_USAGE_CPU_READ_WRITE;
+   }
+
+   pipe_mutex_unlock(fenced_mgr->mutex);
+}
+
+
+static enum pipe_error
+fenced_buffer_validate(struct pb_buffer *buf,
+                       struct pb_validate *vl,
+                       unsigned flags)
+{
+   struct fenced_buffer *fenced_buf = fenced_buffer(buf);
+   struct fenced_manager *fenced_mgr = fenced_buf->mgr;
+   enum pipe_error ret;
+
+   pipe_mutex_lock(fenced_mgr->mutex);
+
+   if(!vl) {
+      /* invalidate */
+      fenced_buf->vl = NULL;
+      fenced_buf->validation_flags = 0;
+      ret = PIPE_OK;
+      goto done;
+   }
+
+   assert(flags & PB_USAGE_GPU_READ_WRITE);
+   assert(!(flags & ~PB_USAGE_GPU_READ_WRITE));
+   flags &= PB_USAGE_GPU_READ_WRITE;
+
+   /* Buffer cannot be validated in two different lists */
+   if(fenced_buf->vl && fenced_buf->vl != vl) {
+      ret = PIPE_ERROR_RETRY;
+      goto done;
+   }
+
+   if(fenced_buf->vl == vl &&
+      (fenced_buf->validation_flags & flags) == flags) {
+      /* Nothing to do -- buffer already validated */
+      ret = PIPE_OK;
+      goto done;
+   }
+
+   ret = pb_validate(fenced_buf->buffer, vl, flags);
+   if (ret != PIPE_OK)
+      goto done;
+
+   fenced_buf->vl = vl;
+   fenced_buf->validation_flags |= flags;
+
+done:
+   pipe_mutex_unlock(fenced_mgr->mutex);
+
+   return ret;
+}
+
+
+static void
+fenced_buffer_fence(struct pb_buffer *buf,
+                    struct pipe_fence_handle *fence)
+{
+   struct fenced_buffer *fenced_buf = fenced_buffer(buf);
+   struct fenced_manager *fenced_mgr = fenced_buf->mgr;
+   struct pb_fence_ops *ops = fenced_mgr->ops;
+
+   pipe_mutex_lock(fenced_mgr->mutex);
+
+   assert(pipe_is_referenced(&fenced_buf->base.reference));
+   assert(fenced_buf->buffer);
+
+   if(fence != fenced_buf->fence) {
+      assert(fenced_buf->vl);
+      assert(fenced_buf->validation_flags);
+
+      if (fenced_buf->fence) {
+         boolean destroyed;
+         destroyed = fenced_buffer_remove_locked(fenced_mgr, fenced_buf);
+         assert(!destroyed);
+      }
+      if (fence) {
+         ops->fence_reference(ops, &fenced_buf->fence, fence);
+         fenced_buf->flags |= fenced_buf->validation_flags;
+         fenced_buffer_add_locked(fenced_mgr, fenced_buf);
+      }
+
+      pb_fence(fenced_buf->buffer, fence);
+
+      fenced_buf->vl = NULL;
+      fenced_buf->validation_flags = 0;
+   }
+
+   pipe_mutex_unlock(fenced_mgr->mutex);
+}
+
+
+static void
+fenced_buffer_get_base_buffer(struct pb_buffer *buf,
+                              struct pb_buffer **base_buf,
+                              pb_size *offset)
+{
+   struct fenced_buffer *fenced_buf = fenced_buffer(buf);
+   struct fenced_manager *fenced_mgr = fenced_buf->mgr;
+
+   pipe_mutex_lock(fenced_mgr->mutex);
+
+   assert(fenced_buf->buffer);
+
+   if(fenced_buf->buffer)
+      pb_get_base_buffer(fenced_buf->buffer, base_buf, offset);
+   else {
+      *base_buf = buf;
+      *offset = 0;
+   }
+
+   pipe_mutex_unlock(fenced_mgr->mutex);
+}
+
+
+static const struct pb_vtbl
+fenced_buffer_vtbl = {
+      fenced_buffer_destroy,
+      fenced_buffer_map,
+      fenced_buffer_unmap,
+      fenced_buffer_validate,
+      fenced_buffer_fence,
+      fenced_buffer_get_base_buffer
+};
+
+
+/**
+ * Wrap a buffer in a fenced buffer.
+ */
+static struct pb_buffer *
+fenced_bufmgr_create_buffer(struct pb_manager *mgr,
+                            pb_size size,
+                            const struct pb_desc *desc)
+{
+   struct fenced_manager *fenced_mgr = fenced_manager(mgr);
+   struct fenced_buffer *fenced_buf;
+   enum pipe_error ret;
+
+   fenced_buf = CALLOC_STRUCT(fenced_buffer);
+   if(!fenced_buf)
+      goto no_buffer;
+
+   pipe_reference_init(&fenced_buf->base.reference, 1);
+   fenced_buf->base.alignment = desc->alignment;
+   fenced_buf->base.usage = desc->usage;
+   fenced_buf->base.size = size;
+   fenced_buf->size = size;
+
+   fenced_buf->base.vtbl = &fenced_buffer_vtbl;
+   fenced_buf->mgr = fenced_mgr;
+
+   pipe_mutex_lock(fenced_mgr->mutex);
+
+   /*
+    * Try to create GPU storage without stalling,
+    */
+   ret = fenced_buffer_create_gpu_storage_locked(fenced_mgr, fenced_buf,
+                                                 desc, TRUE);
+
+   /*
+    * Give up.
+    */
+   if(ret != PIPE_OK) {
+      goto no_storage;
+   }
+
+   assert(fenced_buf->buffer);
+
+   LIST_ADDTAIL(&fenced_buf->head, &fenced_mgr->unfenced);
+   ++fenced_mgr->num_unfenced;
+   pipe_mutex_unlock(fenced_mgr->mutex);
+
+   return &fenced_buf->base;
+
+no_storage:
+   pipe_mutex_unlock(fenced_mgr->mutex);
+   FREE(fenced_buf);
+no_buffer:
+   return NULL;
+}
+
+
+static void
+fenced_bufmgr_flush(struct pb_manager *mgr)
+{
+   struct fenced_manager *fenced_mgr = fenced_manager(mgr);
+
+   pipe_mutex_lock(fenced_mgr->mutex);
+   while(fenced_manager_check_signalled_locked(fenced_mgr, TRUE))
+      ;
+   pipe_mutex_unlock(fenced_mgr->mutex);
+
+   assert(fenced_mgr->provider->flush);
+   if(fenced_mgr->provider->flush)
+      fenced_mgr->provider->flush(fenced_mgr->provider);
+}
+
+
+static void
+fenced_bufmgr_destroy(struct pb_manager *mgr)
+{
+   struct fenced_manager *fenced_mgr = fenced_manager(mgr);
+
+   pipe_mutex_lock(fenced_mgr->mutex);
+
+   /* Wait on outstanding fences */
+   while (fenced_mgr->num_fenced) {
+      pipe_mutex_unlock(fenced_mgr->mutex);
+#if defined(PIPE_OS_LINUX) || defined(PIPE_OS_BSD) || defined(PIPE_OS_SOLARIS)
+      sched_yield();
+#endif
+      pipe_mutex_lock(fenced_mgr->mutex);
+      while(fenced_manager_check_signalled_locked(fenced_mgr, TRUE))
+         ;
+   }
+
+#ifdef DEBUG
+   /*assert(!fenced_mgr->num_unfenced);*/
+#endif
+
+   pipe_mutex_unlock(fenced_mgr->mutex);
+   pipe_mutex_destroy(fenced_mgr->mutex);
+
+   FREE(fenced_mgr);
+}
+
+
+struct pb_manager *
+simple_fenced_bufmgr_create(struct pb_manager *provider,
+                            struct pb_fence_ops *ops)
+{
+   struct fenced_manager *fenced_mgr;
+
+   if(!provider)
+      return NULL;
+
+   fenced_mgr = CALLOC_STRUCT(fenced_manager);
+   if (!fenced_mgr)
+      return NULL;
+
+   fenced_mgr->base.destroy = fenced_bufmgr_destroy;
+   fenced_mgr->base.create_buffer = fenced_bufmgr_create_buffer;
+   fenced_mgr->base.flush = fenced_bufmgr_flush;
+
+   fenced_mgr->provider = provider;
+   fenced_mgr->ops = ops;
+
+   LIST_INITHEAD(&fenced_mgr->fenced);
+   fenced_mgr->num_fenced = 0;
+
+   LIST_INITHEAD(&fenced_mgr->unfenced);
+   fenced_mgr->num_unfenced = 0;
+
+   pipe_mutex_init(fenced_mgr->mutex);
+
+   return &fenced_mgr->base;
+}
diff --git a/src/gallium/winsys/svga/drm/vmw_buffer.c b/src/gallium/winsys/svga/drm/vmw_buffer.c
index 66ab80848ec..c516054b7fc 100644
--- a/src/gallium/winsys/svga/drm/vmw_buffer.c
+++ b/src/gallium/winsys/svga/drm/vmw_buffer.c
@@ -51,7 +51,6 @@
 #include "vmw_screen.h"
 #include "vmw_buffer.h"
 
-
 struct vmw_gmr_bufmgr;
 
 
@@ -63,6 +62,7 @@ struct vmw_gmr_buffer
    
    struct vmw_region *region;
    void *map;
+   unsigned map_flags;
 };
 
 
@@ -113,6 +113,25 @@ vmw_gmr_buffer_map(struct pb_buffer *_buf,
                    void *flush_ctx)
 {
    struct vmw_gmr_buffer *buf = vmw_gmr_buffer(_buf);
+   int ret;
+
+   if (!buf->map)
+      buf->map = vmw_ioctl_region_map(buf->region);
+
+   if (!buf->map)
+      return NULL;
+
+
+   if ((_buf->usage & VMW_BUFFER_USAGE_SYNC) &&
+       !(flags & PB_USAGE_UNSYNCHRONIZED)) {
+      ret = vmw_ioctl_syncforcpu(buf->region,
+                                 !!(flags & PB_USAGE_DONTBLOCK),
+                                 !(flags & PB_USAGE_CPU_WRITE),
+                                 FALSE);
+      if (ret)
+         return NULL;
+   }
+
    return buf->map;
 }
 
@@ -120,8 +139,15 @@ vmw_gmr_buffer_map(struct pb_buffer *_buf,
 static void
 vmw_gmr_buffer_unmap(struct pb_buffer *_buf)
 {
-   /* Do nothing */
-   (void)_buf;
+   struct vmw_gmr_buffer *buf = vmw_gmr_buffer(_buf);
+   unsigned flags = buf->map_flags;
+
+   if ((_buf->usage & VMW_BUFFER_USAGE_SYNC) &&
+       !(flags & PB_USAGE_UNSYNCHRONIZED)) {
+      vmw_ioctl_releasefromcpu(buf->region,
+                               !(flags & PB_USAGE_CPU_WRITE),
+                               FALSE);
+   }
 }
 
 
@@ -167,35 +193,33 @@ const struct pb_vtbl vmw_gmr_buffer_vtbl = {
 static struct pb_buffer *
 vmw_gmr_bufmgr_create_buffer(struct pb_manager *_mgr,
                          pb_size size,
-                         const struct pb_desc *desc) 
+                         const struct pb_desc *pb_desc) 
 {
    struct vmw_gmr_bufmgr *mgr = vmw_gmr_bufmgr(_mgr);
    struct vmw_winsys_screen *vws = mgr->vws;
    struct vmw_gmr_buffer *buf;
+   const struct vmw_buffer_desc *desc =
+      (const struct vmw_buffer_desc *) pb_desc;
    
    buf = CALLOC_STRUCT(vmw_gmr_buffer);
    if(!buf)
       goto error1;
 
    pipe_reference_init(&buf->base.reference, 1);
-   buf->base.alignment = desc->alignment;
-   buf->base.usage = desc->usage;
-   buf->base.size = size;
+   buf->base.alignment = pb_desc->alignment;
+   buf->base.usage = pb_desc->usage & ~VMW_BUFFER_USAGE_SHARED;
    buf->base.vtbl = &vmw_gmr_buffer_vtbl;
    buf->mgr = mgr;
-
-   buf->region = vmw_ioctl_region_create(vws, size);
-   if(!buf->region)
-      goto error2;
+   buf->base.size = size;
+   if ((pb_desc->usage & VMW_BUFFER_USAGE_SHARED) && desc->region) {
+      buf->region = desc->region;
+   } else {
+      buf->region = vmw_ioctl_region_create(vws, size);
+      if(!buf->region)
+	 goto error2;
+   }
 	 
-   buf->map = vmw_ioctl_region_map(buf->region);
-   if(!buf->map)
-      goto error3;
-
    return &buf->base;
-
-error3:
-   vmw_ioctl_region_destroy(buf->region);
 error2:
    FREE(buf);
 error1:
@@ -257,3 +281,91 @@ vmw_gmr_bufmgr_region_ptr(struct pb_buffer *buf,
    
    return TRUE;
 }
+
+#ifdef DEBUG
+struct svga_winsys_buffer {
+   struct pb_buffer *pb_buf;
+   struct debug_flush_buf *fbuf;
+};
+
+struct pb_buffer *
+vmw_pb_buffer(struct svga_winsys_buffer *buffer)
+{
+   assert(buffer);
+   return buffer->pb_buf;
+}
+
+struct svga_winsys_buffer *
+vmw_svga_winsys_buffer_wrap(struct pb_buffer *buffer)
+{
+   struct svga_winsys_buffer *buf;
+
+   if (!buffer)
+      return NULL;
+
+   buf = CALLOC_STRUCT(svga_winsys_buffer);
+   if (!buf) {
+      pb_reference(&buffer, NULL);
+      return NULL;
+   }
+
+   buf->pb_buf = buffer;
+   buf->fbuf = debug_flush_buf_create(TRUE, VMW_DEBUG_FLUSH_STACK);
+   return buf;
+}
+
+struct debug_flush_buf *
+vmw_debug_flush_buf(struct svga_winsys_buffer *buffer)
+{
+   return buffer->fbuf;
+}
+
+#endif
+
+void
+vmw_svga_winsys_buffer_destroy(struct svga_winsys_screen *sws,
+                               struct svga_winsys_buffer *buf)
+{
+   struct pb_buffer *pbuf = vmw_pb_buffer(buf);
+   (void)sws;
+   pb_reference(&pbuf, NULL);
+#ifdef DEBUG
+   debug_flush_buf_reference(&buf->fbuf, NULL);
+   FREE(buf);
+#endif
+}
+
+void *
+vmw_svga_winsys_buffer_map(struct svga_winsys_screen *sws,
+                           struct svga_winsys_buffer *buf,
+                           unsigned flags)
+{
+   void *map;
+
+   (void)sws;
+   if (flags & PIPE_TRANSFER_UNSYNCHRONIZED)
+      flags &= ~PIPE_TRANSFER_DONTBLOCK;
+
+   map = pb_map(vmw_pb_buffer(buf), flags, NULL);
+
+#ifdef DEBUG
+   if (map != NULL)
+      debug_flush_map(buf->fbuf, flags);
+#endif
+
+   return map;
+}
+
+
+void
+vmw_svga_winsys_buffer_unmap(struct svga_winsys_screen *sws,
+                             struct svga_winsys_buffer *buf)
+{
+   (void)sws;
+
+#ifdef DEBUG
+   debug_flush_unmap(buf->fbuf);
+#endif
+
+   pb_unmap(vmw_pb_buffer(buf));
+}
diff --git a/src/gallium/winsys/svga/drm/vmw_buffer.h b/src/gallium/winsys/svga/drm/vmw_buffer.h
index 41fb4476da5..e0bb8085a48 100644
--- a/src/gallium/winsys/svga/drm/vmw_buffer.h
+++ b/src/gallium/winsys/svga/drm/vmw_buffer.h
@@ -29,6 +29,12 @@
 
 #include <assert.h>
 #include "pipe/p_compiler.h"
+#include "pipebuffer/pb_bufmgr.h"
+#include "util/u_debug_flush.h"
+
+
+#define VMW_BUFFER_USAGE_SHARED    (1 << 20)
+#define VMW_BUFFER_USAGE_SYNC      (1 << 21)
 
 struct SVGAGuestPtr;
 struct pb_buffer;
@@ -37,7 +43,22 @@ struct svga_winsys_buffer;
 struct svga_winsys_surface;
 struct vmw_winsys_screen;
 
+struct vmw_buffer_desc {
+   struct pb_desc pb_desc;
+   struct vmw_region *region;
+};
+
+
+#ifdef DEBUG
 
+struct pb_buffer *
+vmw_pb_buffer(struct svga_winsys_buffer *buffer);
+struct svga_winsys_buffer *
+vmw_svga_winsys_buffer_wrap(struct pb_buffer *buffer);
+struct debug_flush_buf *
+vmw_debug_flush_buf(struct svga_winsys_buffer *buffer);
+
+#else
 static INLINE struct pb_buffer *
 vmw_pb_buffer(struct svga_winsys_buffer *buffer)
 {
@@ -47,12 +68,23 @@ vmw_pb_buffer(struct svga_winsys_buffer *buffer)
 
 
 static INLINE struct svga_winsys_buffer *
-vmw_svga_winsys_buffer(struct pb_buffer *buffer)
+vmw_svga_winsys_buffer_wrap(struct pb_buffer *buffer)
 {
-   assert(buffer);
    return (struct svga_winsys_buffer *)buffer;
 }
+#endif
+
+void
+vmw_svga_winsys_buffer_destroy(struct svga_winsys_screen *sws,
+                               struct svga_winsys_buffer *buf);
+void *
+vmw_svga_winsys_buffer_map(struct svga_winsys_screen *sws,
+                           struct svga_winsys_buffer *buf,
+                           unsigned flags);
 
+void
+vmw_svga_winsys_buffer_unmap(struct svga_winsys_screen *sws,
+                             struct svga_winsys_buffer *buf);
 
 struct pb_manager *
 vmw_gmr_bufmgr_create(struct vmw_winsys_screen *vws);
diff --git a/src/gallium/winsys/svga/drm/vmw_context.c b/src/gallium/winsys/svga/drm/vmw_context.c
index eb89272a5e6..1234a5edce3 100644
--- a/src/gallium/winsys/svga/drm/vmw_context.c
+++ b/src/gallium/winsys/svga/drm/vmw_context.c
@@ -29,6 +29,8 @@
 #include "util/u_debug.h"
 #include "util/u_memory.h"
 #include "util/u_debug_stack.h"
+#include "util/u_debug_flush.h"
+#include "util/u_hash_table.h"
 #include "pipebuffer/pb_buffer.h"
 #include "pipebuffer/pb_validate.h"
 
@@ -38,19 +40,38 @@
 #include "vmw_buffer.h"
 #include "vmw_surface.h"
 #include "vmw_fence.h"
+#include "vmw_shader.h"
 
 #define VMW_COMMAND_SIZE (64*1024)
 #define VMW_SURFACE_RELOCS (1024)
+#define VMW_SHADER_RELOCS (1024)
 #define VMW_REGION_RELOCS (512)
 
 #define VMW_MUST_FLUSH_STACK 8
 
-struct vmw_region_relocation
+struct vmw_buffer_relocation
 {
-   struct SVGAGuestPtr *where;
    struct pb_buffer *buffer;
-   /* TODO: put offset info inside where */
+   boolean is_mob;
    uint32 offset;
+
+   union {
+      struct {
+	 struct SVGAGuestPtr *where;
+      } region;
+      struct {
+	 SVGAMobId *id;
+	 uint32 *offset_into_mob;
+      } mob;
+   };
+};
+
+struct vmw_ctx_validate_item {
+   union {
+      struct vmw_svga_winsys_surface *vsurf;
+      struct vmw_svga_winsys_shader *vshader;
+   };
+   boolean referenced;
 };
 
 struct vmw_svga_winsys_context
@@ -58,10 +79,12 @@ struct vmw_svga_winsys_context
    struct svga_winsys_context base;
 
    struct vmw_winsys_screen *vws;
+   struct util_hash_table *hash;
 
 #ifdef DEBUG
    boolean must_flush;
    struct debug_stack_frame must_flush_stack[VMW_MUST_FLUSH_STACK];
+   struct debug_flush_ctx *fctx;
 #endif
 
    struct {
@@ -72,7 +95,7 @@ struct vmw_svga_winsys_context
    } command;
 
    struct {
-      struct vmw_svga_winsys_surface *handles[VMW_SURFACE_RELOCS];
+      struct vmw_ctx_validate_item items[VMW_SURFACE_RELOCS];
       uint32_t size;
       uint32_t used;
       uint32_t staged;
@@ -80,20 +103,30 @@ struct vmw_svga_winsys_context
    } surface;
    
    struct {
-      struct vmw_region_relocation relocs[VMW_REGION_RELOCS];
+      struct vmw_buffer_relocation relocs[VMW_REGION_RELOCS];
       uint32_t size;
       uint32_t used;
       uint32_t staged;
       uint32_t reserved;
    } region;
 
+   struct {
+      struct vmw_ctx_validate_item items[VMW_SHADER_RELOCS];
+      uint32_t size;
+      uint32_t used;
+      uint32_t staged;
+      uint32_t reserved;
+   } shader;
+
    struct pb_validate *validate;
 
    /**
-    * The amount of GMR that is referred by the commands currently batched
-    * in the context.
+    * The amount of surface, GMR or MOB memory that is referred by the commands
+    * currently batched in the context command buffer.
     */
-   uint32_t seen_regions;
+   uint64_t seen_surfaces;
+   uint64_t seen_regions;
+   uint64_t seen_mobs;
 
    /**
     * Whether this context should fail to reserve more commands, not because it
@@ -140,7 +173,7 @@ vmw_swc_flush(struct svga_winsys_context *swc,
    
       /* Apply relocations */
       for(i = 0; i < vswc->region.used; ++i) {
-         struct vmw_region_relocation *reloc = &vswc->region.relocs[i];
+         struct vmw_buffer_relocation *reloc = &vswc->region.relocs[i];
          struct SVGAGuestPtr ptr;
 
          if(!vmw_gmr_bufmgr_region_ptr(reloc->buffer, &ptr))
@@ -148,7 +181,16 @@ vmw_swc_flush(struct svga_winsys_context *swc,
 
          ptr.offset += reloc->offset;
 
-         *reloc->where = ptr;
+	 if (reloc->is_mob) {
+	    if (reloc->mob.id)
+	       *reloc->mob.id = ptr.gmrId;
+	    if (reloc->mob.offset_into_mob)
+	       *reloc->mob.offset_into_mob = ptr.offset;
+	    else {
+	       assert(ptr.offset == 0);
+	    }
+	 } else
+	    *reloc->region.where = ptr;
       }
 
       if (vswc->command.used || pfence != NULL)
@@ -166,27 +208,37 @@ vmw_swc_flush(struct svga_winsys_context *swc,
    vswc->command.reserved = 0;
 
    for(i = 0; i < vswc->surface.used + vswc->surface.staged; ++i) {
-      struct vmw_svga_winsys_surface *vsurf =
-	 vswc->surface.handles[i];
-      p_atomic_dec(&vsurf->validated);
-      vmw_svga_winsys_surface_reference(&vswc->surface.handles[i], NULL);
+      struct vmw_ctx_validate_item *isurf = &vswc->surface.items[i];
+      if (isurf->referenced)
+         p_atomic_dec(&isurf->vsurf->validated);
+      vmw_svga_winsys_surface_reference(&isurf->vsurf, NULL);
    }
 
+   util_hash_table_clear(vswc->hash);
    vswc->surface.used = 0;
    vswc->surface.reserved = 0;
 
-   for(i = 0; i < vswc->region.used + vswc->region.staged; ++i) {
-      pb_reference(&vswc->region.relocs[i].buffer, NULL);
+   for(i = 0; i < vswc->shader.used + vswc->shader.staged; ++i) {
+      struct vmw_ctx_validate_item *ishader = &vswc->shader.items[i];
+      if (ishader->referenced)
+         p_atomic_dec(&ishader->vshader->validated);
+      vmw_svga_winsys_shader_reference(&ishader->vshader, NULL);
    }
 
+   vswc->shader.used = 0;
+   vswc->shader.reserved = 0;
+
    vswc->region.used = 0;
    vswc->region.reserved = 0;
 
 #ifdef DEBUG
    vswc->must_flush = FALSE;
+   debug_flush_flush(vswc->fctx);
 #endif
    vswc->preemptive_flush = FALSE;
+   vswc->seen_surfaces = 0;
    vswc->seen_regions = 0;
+   vswc->seen_mobs = 0;
 
    if(pfence)
       vmw_fence_reference(vswc->vws, pfence, fence);
@@ -210,6 +262,7 @@ vmw_swc_reserve(struct svga_winsys_context *swc,
       debug_backtrace_dump(vswc->must_flush_stack, VMW_MUST_FLUSH_STACK);
       assert(!vswc->must_flush);
    }
+   debug_flush_might_flush(vswc->fctx);
 #endif
 
    assert(nr_bytes <= vswc->command.size);
@@ -219,6 +272,7 @@ vmw_swc_reserve(struct svga_winsys_context *swc,
    if(vswc->preemptive_flush ||
       vswc->command.used + nr_bytes > vswc->command.size ||
       vswc->surface.used + nr_relocs > vswc->surface.size ||
+      vswc->shader.used + nr_relocs > vswc->shader.size ||
       vswc->region.used + nr_relocs > vswc->region.size) {
 #ifdef DEBUG
       vswc->must_flush = TRUE;
@@ -230,17 +284,190 @@ vmw_swc_reserve(struct svga_winsys_context *swc,
 
    assert(vswc->command.used + nr_bytes <= vswc->command.size);
    assert(vswc->surface.used + nr_relocs <= vswc->surface.size);
+   assert(vswc->shader.used + nr_relocs <= vswc->shader.size);
    assert(vswc->region.used + nr_relocs <= vswc->region.size);
    
    vswc->command.reserved = nr_bytes;
    vswc->surface.reserved = nr_relocs;
    vswc->surface.staged = 0;
+   vswc->shader.reserved = nr_relocs;
+   vswc->shader.staged = 0;
    vswc->region.reserved = nr_relocs;
    vswc->region.staged = 0;
    
    return vswc->command.buffer + vswc->command.used;
 }
 
+static void
+vmw_swc_context_relocation(struct svga_winsys_context *swc,
+			   uint32 *cid)
+{
+   *cid = swc->cid;
+}
+
+static boolean
+vmw_swc_add_validate_buffer(struct vmw_svga_winsys_context *vswc,
+			    struct pb_buffer *pb_buf,
+			    unsigned flags)
+{
+   enum pipe_error ret;
+   unsigned translated_flags;
+
+   /*
+    * TODO: Update pb_validate to provide a similar functionality
+    * (Check buffer already present before adding)
+    */
+   if (util_hash_table_get(vswc->hash, pb_buf) != pb_buf) {
+      translated_flags = vmw_translate_to_pb_flags(flags);
+      ret = pb_validate_add_buffer(vswc->validate, pb_buf, translated_flags);
+      /* TODO: Update pipebuffer to reserve buffers and not fail here */
+      assert(ret == PIPE_OK);
+      (void)ret;
+      (void)util_hash_table_set(vswc->hash, pb_buf, pb_buf);
+      return TRUE;
+   }
+
+   return FALSE;
+}
+
+static void
+vmw_swc_region_relocation(struct svga_winsys_context *swc,
+                          struct SVGAGuestPtr *where,
+                          struct svga_winsys_buffer *buffer,
+                          uint32 offset,
+                          unsigned flags)
+{
+   struct vmw_svga_winsys_context *vswc = vmw_svga_winsys_context(swc);
+   struct vmw_buffer_relocation *reloc;
+
+   assert(vswc->region.staged < vswc->region.reserved);
+
+   reloc = &vswc->region.relocs[vswc->region.used + vswc->region.staged];
+   reloc->region.where = where;
+
+   /*
+    * pb_validate holds a refcount to the buffer, so no need to
+    * refcount it again in the relocation.
+    */
+   reloc->buffer = vmw_pb_buffer(buffer);
+   reloc->offset = offset;
+   reloc->is_mob = FALSE;
+   ++vswc->region.staged;
+
+   if (vmw_swc_add_validate_buffer(vswc, reloc->buffer, flags)) {
+      vswc->seen_regions += reloc->buffer->size;
+      if(vswc->seen_regions >= VMW_GMR_POOL_SIZE/5)
+         vswc->preemptive_flush = TRUE;
+   }
+
+#ifdef DEBUG
+   if (!(flags & SVGA_RELOC_INTERNAL))
+      debug_flush_cb_reference(vswc->fctx, vmw_debug_flush_buf(buffer));
+#endif
+}
+
+static void
+vmw_swc_mob_relocation(struct svga_winsys_context *swc,
+		       SVGAMobId *id,
+		       uint32 *offset_into_mob,
+		       struct svga_winsys_buffer *buffer,
+		       uint32 offset,
+		       unsigned flags)
+{
+   struct vmw_svga_winsys_context *vswc = vmw_svga_winsys_context(swc);
+   struct vmw_buffer_relocation *reloc;
+
+   assert(vswc->region.staged < vswc->region.reserved);
+
+   reloc = &vswc->region.relocs[vswc->region.used + vswc->region.staged];
+   reloc->mob.id = id;
+   reloc->mob.offset_into_mob = offset_into_mob;
+
+   /*
+    * pb_validate holds a refcount to the buffer, so no need to
+    * refcount it again in the relocation.
+    */
+   reloc->buffer = vmw_pb_buffer(buffer);
+   reloc->offset = offset;
+   reloc->is_mob = TRUE;
+   ++vswc->region.staged;
+
+   if (vmw_swc_add_validate_buffer(vswc, reloc->buffer, flags)) {
+      vswc->seen_mobs += reloc->buffer->size;
+      /* divide by 5, tested for best performance */
+      if (vswc->seen_mobs >= vswc->vws->ioctl.max_mob_memory / 5)
+         vswc->preemptive_flush = TRUE;
+   }
+
+#ifdef DEBUG
+   if (!(flags & SVGA_RELOC_INTERNAL))
+      debug_flush_cb_reference(vswc->fctx, vmw_debug_flush_buf(buffer));
+#endif
+}
+
+
+/**
+ * vmw_swc_surface_clear_reference - Clear referenced info for a surface
+ *
+ * @swc:   Pointer to an svga_winsys_context
+ * @vsurf: Pointer to a vmw_svga_winsys_surface, the referenced info of which
+ *         we want to clear
+ *
+ * This is primarily used by a discard surface map to indicate that the
+ * surface data is no longer referenced by a draw call, and mapping it
+ * should therefore no longer cause a flush.
+ */
+void
+vmw_swc_surface_clear_reference(struct svga_winsys_context *swc,
+                                struct vmw_svga_winsys_surface *vsurf)
+{
+   struct vmw_svga_winsys_context *vswc = vmw_svga_winsys_context(swc);
+   struct vmw_ctx_validate_item *isrf =
+      util_hash_table_get(vswc->hash, vsurf);
+
+   if (isrf && isrf->referenced) {
+      isrf->referenced = FALSE;
+      p_atomic_dec(&vsurf->validated);
+   }
+}
+
+static void
+vmw_swc_surface_only_relocation(struct svga_winsys_context *swc,
+				uint32 *where,
+				struct vmw_svga_winsys_surface *vsurf,
+				unsigned flags)
+{
+   struct vmw_svga_winsys_context *vswc = vmw_svga_winsys_context(swc);
+   struct vmw_ctx_validate_item *isrf;
+
+   assert(vswc->surface.staged < vswc->surface.reserved);
+   isrf = util_hash_table_get(vswc->hash, vsurf);
+
+   if (isrf == NULL) {
+      isrf = &vswc->surface.items[vswc->surface.used + vswc->surface.staged];
+      vmw_svga_winsys_surface_reference(&isrf->vsurf, vsurf);
+      isrf->referenced = FALSE;
+      /*
+       * Note that a failure here may just fall back to unhashed behavior
+       * and potentially cause unnecessary flushing, so ignore the
+       * return code.
+       */
+      (void) util_hash_table_set(vswc->hash, vsurf, isrf);
+      ++vswc->surface.staged;
+
+      vswc->seen_surfaces += vsurf->size;
+      /* divide by 5 not well tuned for performance */
+      if (vswc->seen_surfaces >= vswc->vws->ioctl.max_surface_memory / 5)
+         vswc->preemptive_flush = TRUE;
+   }
+
+   if (!(flags & SVGA_RELOC_INTERNAL) && !isrf->referenced) {
+      isrf->referenced = TRUE;
+      p_atomic_inc(&vsurf->validated);
+   }
+
+   *where = vsurf->sid;
+}
 
 static void
 vmw_swc_surface_relocation(struct svga_winsys_context *swc,
@@ -249,75 +476,79 @@ vmw_swc_surface_relocation(struct svga_winsys_context *swc,
                            struct svga_winsys_surface *surface,
                            unsigned flags)
 {
-   struct vmw_svga_winsys_context *vswc = vmw_svga_winsys_context(swc);
    struct vmw_svga_winsys_surface *vsurf;
 
+   assert(swc->have_gb_objects || mobid == NULL);
+
    if(!surface) {
       *where = SVGA3D_INVALID_ID;
+      if (mobid)
+         *mobid = SVGA3D_INVALID_ID;
       return;
    }
 
-   assert(vswc->surface.staged < vswc->surface.reserved);
-
    vsurf = vmw_svga_winsys_surface(surface);
+   vmw_swc_surface_only_relocation(swc, where, vsurf, flags);
 
-   *where = vsurf->sid;
+   if (swc->have_gb_objects && vsurf->buf != NULL) {
 
-   vmw_svga_winsys_surface_reference(&vswc->surface.handles[vswc->surface.used + vswc->surface.staged], vsurf);
-   p_atomic_inc(&vsurf->validated);
-   ++vswc->surface.staged;
-}
+      /*
+       * Make sure backup buffer ends up fenced.
+       */
 
+      pipe_mutex_lock(vsurf->mutex);
+      assert(vsurf->buf != NULL);
+      
+      vmw_swc_mob_relocation(swc, mobid, NULL, (struct svga_winsys_buffer *)
+                             vsurf->buf, 0, flags);
+      pipe_mutex_unlock(vsurf->mutex);
+   }
+}
 
 static void
-vmw_swc_region_relocation(struct svga_winsys_context *swc,
-                          struct SVGAGuestPtr *where,
-                          struct svga_winsys_buffer *buffer,
-                          uint32 offset,
-                          unsigned flags)
+vmw_swc_shader_relocation(struct svga_winsys_context *swc,
+			  uint32 *shid,
+			  uint32 *mobid,
+			  uint32 *offset,
+			  struct svga_winsys_gb_shader *shader)
 {
    struct vmw_svga_winsys_context *vswc = vmw_svga_winsys_context(swc);
-   struct vmw_region_relocation *reloc;
-   unsigned translated_flags;
-   enum pipe_error ret;
-   
-   assert(vswc->region.staged < vswc->region.reserved);
+   struct vmw_svga_winsys_shader *vshader;
+   struct vmw_ctx_validate_item *ishader;
+   if(!shader) {
+      *shid = SVGA3D_INVALID_ID;
+      return;
+   }
 
-   reloc = &vswc->region.relocs[vswc->region.used + vswc->region.staged];
-   reloc->where = where;
-   pb_reference(&reloc->buffer, vmw_pb_buffer(buffer));
-   reloc->offset = offset;
+   assert(vswc->shader.staged < vswc->shader.reserved);
+   vshader = vmw_svga_winsys_shader(shader);
+   ishader = util_hash_table_get(vswc->hash, vshader);
+
+   if (ishader == NULL) {
+      ishader = &vswc->shader.items[vswc->shader.used + vswc->shader.staged];
+      vmw_svga_winsys_shader_reference(&ishader->vshader, vshader);
+      ishader->referenced = FALSE;
+      /*
+       * Note that a failure here may just fall back to unhashed behavior
+       * and potentially cause unnecessary flushing, so ignore the
+       * return code.
+       */
+      (void) util_hash_table_set(vswc->hash, vshader, ishader);
+      ++vswc->shader.staged;
+   }
 
-   ++vswc->region.staged;
+   if (!ishader->referenced) {
+      ishader->referenced = TRUE;
+      p_atomic_inc(&vshader->validated);
+   }
 
-   translated_flags = vmw_translate_to_pb_flags(flags);
-   ret = pb_validate_add_buffer(vswc->validate, reloc->buffer, translated_flags);
-   /* TODO: Update pipebuffer to reserve buffers and not fail here */
-   assert(ret == PIPE_OK);
-   (void)ret;
+   *shid = vshader->shid;
 
-   /*
-    * Flush preemptively the FIFO commands to keep the GMR working set within
-    * the GMR pool size.
-    *
-    * This is necessary for applications like SPECviewperf that generate huge
-    * amounts of immediate vertex data, so that we don't pile up too much of
-    * that vertex data neither in the guest nor in the host.
-    *
-    * Note that in the current implementation if a region is referred twice in
-    * a command stream, it will be accounted twice. We could detect repeated
-    * regions and count only once, but there is no incentive to do that, since
-    * regions are typically short-lived; always referred in a single command;
-    * and at the worst we just flush the commands a bit sooner, which for the
-    * SVGA virtual device it's not a performance issue since flushing commands
-    * to the FIFO won't cause flushing in the host.
-    */
-   vswc->seen_regions += reloc->buffer->size;
-   if(vswc->seen_regions >= VMW_GMR_POOL_SIZE/3)
-      vswc->preemptive_flush = TRUE;
+   if (mobid != NULL && vshader->buf)
+      vmw_swc_mob_relocation(swc, mobid, offset, vshader->buf,
+			     0, SVGA_RELOC_READ);
 }
 
-
 static void
 vmw_swc_commit(struct svga_winsys_context *swc)
 {
@@ -334,6 +565,12 @@ vmw_swc_commit(struct svga_winsys_context *swc)
    vswc->surface.staged = 0;
    vswc->surface.reserved = 0;
 
+   assert(vswc->shader.staged <= vswc->shader.reserved);
+   assert(vswc->shader.used + vswc->shader.staged <= vswc->shader.size);
+   vswc->shader.used += vswc->shader.staged;
+   vswc->shader.staged = 0;
+   vswc->shader.reserved = 0;
+
    assert(vswc->region.staged <= vswc->region.reserved);
    assert(vswc->region.used + vswc->region.staged <= vswc->region.size);
    vswc->region.used += vswc->region.staged;
@@ -348,19 +585,38 @@ vmw_swc_destroy(struct svga_winsys_context *swc)
    struct vmw_svga_winsys_context *vswc = vmw_svga_winsys_context(swc);
    unsigned i;
 
-   for(i = 0; i < vswc->region.used; ++i) {
-      pb_reference(&vswc->region.relocs[i].buffer, NULL);
+   for(i = 0; i < vswc->surface.used; ++i) {
+      struct vmw_ctx_validate_item *isurf = &vswc->surface.items[i];
+      if (isurf->referenced)
+         p_atomic_dec(&isurf->vsurf->validated);
+      vmw_svga_winsys_surface_reference(&isurf->vsurf, NULL);
    }
 
-   for(i = 0; i < vswc->surface.used; ++i) {
-      p_atomic_dec(&vswc->surface.handles[i]->validated);
-      vmw_svga_winsys_surface_reference(&vswc->surface.handles[i], NULL);
+   for(i = 0; i < vswc->shader.used; ++i) {
+      struct vmw_ctx_validate_item *ishader = &vswc->shader.items[i];
+      if (ishader->referenced)
+         p_atomic_dec(&ishader->vshader->validated);
+      vmw_svga_winsys_shader_reference(&ishader->vshader, NULL);
    }
+
+   util_hash_table_destroy(vswc->hash);
    pb_validate_destroy(vswc->validate);
    vmw_ioctl_context_destroy(vswc->vws, swc->cid);
+#ifdef DEBUG
+   debug_flush_ctx_destroy(vswc->fctx);
+#endif
    FREE(vswc);
 }
 
+static unsigned vmw_hash_ptr(void *p)
+{
+   return (unsigned)(unsigned long)p;
+}
+
+static int vmw_ptr_compare(void *key1, void *key2)
+{
+   return (key1 == key2) ? 0 : 1;
+}
 
 struct svga_winsys_context *
 vmw_svga_winsys_context_create(struct svga_winsys_screen *sws)
@@ -376,22 +632,41 @@ vmw_svga_winsys_context_create(struct svga_winsys_screen *sws)
    vswc->base.reserve = vmw_swc_reserve;
    vswc->base.surface_relocation = vmw_swc_surface_relocation;
    vswc->base.region_relocation = vmw_swc_region_relocation;
+   vswc->base.mob_relocation = vmw_swc_mob_relocation;
+   vswc->base.context_relocation = vmw_swc_context_relocation;
+   vswc->base.shader_relocation = vmw_swc_shader_relocation;
    vswc->base.commit = vmw_swc_commit;
    vswc->base.flush = vmw_swc_flush;
+   vswc->base.surface_map = vmw_svga_winsys_surface_map;
+   vswc->base.surface_unmap = vmw_svga_winsys_surface_unmap;
 
    vswc->base.cid = vmw_ioctl_context_create(vws);
+   vswc->base.have_gb_objects = sws->have_gb_objects;
 
    vswc->vws = vws;
 
    vswc->command.size = VMW_COMMAND_SIZE;
    vswc->surface.size = VMW_SURFACE_RELOCS;
+   vswc->shader.size = VMW_SHADER_RELOCS;
    vswc->region.size = VMW_REGION_RELOCS;
 
    vswc->validate = pb_validate_create();
-   if(!vswc->validate) {
-      FREE(vswc);
-      return NULL;
-   }
+   if(!vswc->validate)
+      goto out_no_validate;
+
+   vswc->hash = util_hash_table_create(vmw_hash_ptr, vmw_ptr_compare);
+   if (!vswc->hash)
+      goto out_no_hash;
+
+#ifdef DEBUG
+   vswc->fctx = debug_flush_ctx_create(TRUE, VMW_DEBUG_FLUSH_STACK);
+#endif
 
    return &vswc->base;
+
+out_no_hash:
+   pb_validate_destroy(vswc->validate);
+out_no_validate:
+   FREE(vswc);
+   return NULL;
 }
diff --git a/src/gallium/winsys/svga/drm/vmw_context.h b/src/gallium/winsys/svga/drm/vmw_context.h
index 3d383ccb1ed..2c2fb415782 100644
--- a/src/gallium/winsys/svga/drm/vmw_context.h
+++ b/src/gallium/winsys/svga/drm/vmw_context.h
@@ -62,4 +62,12 @@ struct pipe_screen;
 struct svga_winsys_context *
 vmw_svga_winsys_context_create(struct svga_winsys_screen *sws);
 
+struct vmw_svga_winsys_surface;
+
+
+void
+vmw_swc_surface_clear_reference(struct svga_winsys_context *swc,
+                                struct vmw_svga_winsys_surface *vsurf);
+
+
 #endif /* VMW_CONTEXT_H_ */
diff --git a/src/gallium/winsys/svga/drm/vmw_fence.c b/src/gallium/winsys/svga/drm/vmw_fence.c
index 754f8a666df..8af2250b7ea 100644
--- a/src/gallium/winsys/svga/drm/vmw_fence.c
+++ b/src/gallium/winsys/svga/drm/vmw_fence.c
@@ -22,20 +22,10 @@
  * SOFTWARE.
  *
  **********************************************************/
-/*
- * TODO:
- *
- * Fencing is currently a bit inefficient, since we need to call the
- * kernel do determine a fence object signaled status if the fence is not
- * signaled. This can be greatly improved upon by using the fact that the
- * execbuf ioctl returns the last signaled fence seqno, as does the
- * fence signaled ioctl. We should set up a ring of fence objects and
- * walk through them checking for signaled status each time we receive a
- * new passed fence seqno.
- */
-
 #include "util/u_memory.h"
 #include "util/u_atomic.h"
+#include "util/u_double_list.h"
+#include "os/os_thread.h"
 
 #include "pipebuffer/pb_buffer_fenced.h"
 
@@ -44,20 +34,129 @@
 
 struct vmw_fence_ops 
 {
+   /*
+    * Immutable members.
+    */
    struct pb_fence_ops base;
-
    struct vmw_winsys_screen *vws;
+
+   pipe_mutex mutex;
+
+   /*
+    * Protected by mutex;
+    */
+   struct list_head not_signaled;
+   uint32_t last_signaled;
+   uint32_t last_emitted;
 };
 
 struct vmw_fence
 {
+   struct list_head ops_list;
    int32_t refcount;
    uint32_t handle;
    uint32_t mask;
    int32_t signalled;
+   uint32_t seqno;
 };
 
 /**
+ * vmw_fence_seq_is_signaled - Check whether a fence seqno is
+ * signaled.
+ *
+ * @ops: Pointer to a struct pb_fence_ops.
+ *
+ */
+static INLINE boolean
+vmw_fence_seq_is_signaled(uint32_t seq, uint32_t last, uint32_t cur)
+{
+   return (cur - last <= cur - seq);
+}
+
+
+/**
+ * vmw_fence_ops - Return the vmw_fence_ops structure backing a
+ * struct pb_fence_ops pointer.
+ *
+ * @ops: Pointer to a struct pb_fence_ops.
+ *
+ */
+static INLINE struct vmw_fence_ops *
+vmw_fence_ops(struct pb_fence_ops *ops)
+{
+   assert(ops);
+   return (struct vmw_fence_ops *)ops;
+}
+
+
+/**
+ * vmw_fences_release - Release all fences from the not_signaled
+ * list.
+ *
+ * @ops: Pointer to a struct vmw_fence_ops.
+ *
+ */
+static void
+vmw_fences_release(struct vmw_fence_ops *ops)
+{
+   struct vmw_fence *fence, *n;
+
+   pipe_mutex_lock(ops->mutex);
+   LIST_FOR_EACH_ENTRY_SAFE(fence, n, &ops->not_signaled, ops_list)
+      LIST_DELINIT(&fence->ops_list);
+   pipe_mutex_unlock(ops->mutex);
+}
+
+/**
+ * vmw_fences_signal - Traverse the not_signaled list and try to
+ * signal unsignaled fences.
+ *
+ * @ops: Pointer to a struct pb_fence_ops.
+ * @signaled: Seqno that has signaled.
+ * @emitted: Last seqno emitted by the kernel.
+ * @has_emitted: Whether we provide the emitted value.
+ *
+ */
+void
+vmw_fences_signal(struct pb_fence_ops *fence_ops,
+                  uint32_t signaled,
+                  uint32_t emitted,
+                  boolean has_emitted)
+{
+   struct vmw_fence_ops *ops = NULL;
+   struct vmw_fence *fence, *n;
+
+   if (fence_ops == NULL)
+      return;
+
+   ops = vmw_fence_ops(fence_ops);
+   pipe_mutex_lock(ops->mutex);
+
+   if (!has_emitted) {
+      emitted = ops->last_emitted;
+      if (emitted - signaled > (1 << 30))
+	emitted = signaled;
+   }
+
+   if (signaled == ops->last_signaled && emitted == ops->last_emitted)
+      goto out_unlock;
+
+   LIST_FOR_EACH_ENTRY_SAFE(fence, n, &ops->not_signaled, ops_list) {
+      if (!vmw_fence_seq_is_signaled(fence->seqno, signaled, emitted))
+         break;
+
+      p_atomic_set(&fence->signalled, 1);
+      LIST_DELINIT(&fence->ops_list);
+   }
+   ops->last_signaled = signaled;
+   ops->last_emitted = emitted;
+
+out_unlock:
+   pipe_mutex_unlock(ops->mutex);
+}
+
+
+/**
  * vmw_fence - return the vmw_fence object identified by a
  * struct pipe_fence_handle *
  *
@@ -69,18 +168,22 @@ vmw_fence(struct pipe_fence_handle *fence)
    return (struct vmw_fence *) fence;
 }
 
+
 /**
  * vmw_fence_create - Create a user-space fence object.
  *
+ * @fence_ops: The fence_ops manager to register with.
  * @handle: Handle identifying the kernel fence object.
  * @mask: Mask of flags that this fence object may signal.
  *
  * Returns NULL on failure.
  */
 struct pipe_fence_handle *
-vmw_fence_create(uint32_t handle, uint32_t mask)
+vmw_fence_create(struct pb_fence_ops *fence_ops, uint32_t handle,
+                 uint32_t seqno, uint32_t mask)
 {
    struct vmw_fence *fence = CALLOC_STRUCT(vmw_fence);
+   struct vmw_fence_ops *ops = vmw_fence_ops(fence_ops);
 
    if (!fence)
       return NULL;
@@ -88,27 +191,24 @@ vmw_fence_create(uint32_t handle, uint32_t mask)
    p_atomic_set(&fence->refcount, 1);
    fence->handle = handle;
    fence->mask = mask;
+   fence->seqno = seqno;
    p_atomic_set(&fence->signalled, 0);
+   pipe_mutex_lock(ops->mutex);
+
+   if (vmw_fence_seq_is_signaled(seqno, ops->last_signaled, seqno)) {
+      p_atomic_set(&fence->signalled, 1);
+      LIST_INITHEAD(&fence->ops_list);
+   } else {
+      p_atomic_set(&fence->signalled, 0);
+      LIST_ADDTAIL(&fence->ops_list, &ops->not_signaled);
+   }
 
-   return (struct pipe_fence_handle *) fence;
-}
+   pipe_mutex_unlock(ops->mutex);
 
-/**
- * vmw_fence_ops - Return the vmw_fence_ops structure backing a
- * struct pb_fence_ops pointer.
- *
- * @ops: Pointer to a struct pb_fence_ops.
- *
- */
-static INLINE struct vmw_fence_ops *
-vmw_fence_ops(struct pb_fence_ops *ops)
-{
-   assert(ops);
-   return (struct vmw_fence_ops *)ops;
+   return (struct pipe_fence_handle *) fence;
 }
 
 
-
 /**
  * vmw_fence_reference - Reference / unreference a vmw fence object.
  *
@@ -125,7 +225,14 @@ vmw_fence_reference(struct vmw_winsys_screen *vws,
       struct vmw_fence *vfence = vmw_fence(*ptr);
 
       if (p_atomic_dec_zero(&vfence->refcount)) {
+         struct vmw_fence_ops *ops = vmw_fence_ops(vws->fence_ops);
+
 	 vmw_ioctl_fence_unref(vws, vfence->handle);
+
+         pipe_mutex_lock(ops->mutex);
+         LIST_DELINIT(&vfence->ops_list);
+         pipe_mutex_unlock(ops->mutex);
+
 	 FREE(vfence);
       }
    }
@@ -171,18 +278,21 @@ vmw_fence_signalled(struct vmw_winsys_screen *vws,
    if ((old & vflags) == vflags)
       return 0;
 
+   /*
+    * Currently we update signaled fences on each execbuf call.
+    * That should really be sufficient, and we can avoid
+    * a lot of kernel calls this way.
+    */
+#if 1
    ret = vmw_ioctl_fence_signalled(vws, vfence->handle, vflags);
 
-   if (ret == 0) {
-      int32_t prev = old;
-
-      do {
-	 old = prev;
-	 prev = p_atomic_cmpxchg(&vfence->signalled, old, old | vflags);
-      } while (prev != old);
-   }
-
+   if (ret == 0)
+      p_atomic_set(&vfence->signalled, 1);
    return ret;
+#else
+   (void) ret;
+   return -1;
+#endif
 }
 
 /**
@@ -287,6 +397,7 @@ vmw_fence_ops_fence_finish(struct pb_fence_ops *ops,
 static void
 vmw_fence_ops_destroy(struct pb_fence_ops *ops)
 {
+   vmw_fences_release(vmw_fence_ops(ops));
    FREE(ops);
 }
 
@@ -310,6 +421,8 @@ vmw_fence_ops_create(struct vmw_winsys_screen *vws)
    if(!ops)
       return NULL;
 
+   pipe_mutex_init(ops->mutex);
+   LIST_INITHEAD(&ops->not_signaled);
    ops->base.destroy = &vmw_fence_ops_destroy;
    ops->base.fence_reference = &vmw_fence_ops_fence_reference;
    ops->base.fence_signalled = &vmw_fence_ops_fence_signalled;
@@ -319,5 +432,3 @@ vmw_fence_ops_create(struct vmw_winsys_screen *vws)
 
    return &ops->base;
 }
-
-
diff --git a/src/gallium/winsys/svga/drm/vmw_fence.h b/src/gallium/winsys/svga/drm/vmw_fence.h
index 403ae266d5e..56f1a0ab05d 100644
--- a/src/gallium/winsys/svga/drm/vmw_fence.h
+++ b/src/gallium/winsys/svga/drm/vmw_fence.h
@@ -29,7 +29,7 @@
 
 
 #include "pipe/p_compiler.h"
-
+#include "pipebuffer/pb_buffer_fenced.h"
 
 struct pipe_fence_handle;
 struct pb_fence_ops;
@@ -37,7 +37,8 @@ struct vmw_winsys_screen;
 
 
 struct pipe_fence_handle *
-vmw_fence_create(uint32_t handle, uint32_t mask);
+vmw_fence_create(struct pb_fence_ops *fence_ops,
+		 uint32_t handle, uint32_t seqno, uint32_t mask);
 
 int
 vmw_fence_finish(struct vmw_winsys_screen *vws,
diff --git a/src/gallium/winsys/svga/drm/vmw_screen.c b/src/gallium/winsys/svga/drm/vmw_screen.c
index f7799cec32d..0c343cc7bdd 100644
--- a/src/gallium/winsys/svga/drm/vmw_screen.c
+++ b/src/gallium/winsys/svga/drm/vmw_screen.c
@@ -25,42 +25,94 @@
 
 
 #include "vmw_screen.h"
-
+#include "vmw_fence.h"
 #include "vmw_context.h"
 
 #include "util/u_memory.h"
 #include "pipe/p_compiler.h"
+#include "util/u_hash_table.h"
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <unistd.h>
 
+static struct util_hash_table *dev_hash = NULL;
+
+static int vmw_dev_compare(void *key1, void *key2)
+{
+   return (major(*(dev_t *)key1) == major(*(dev_t *)key2) &&
+           minor(*(dev_t *)key1) == minor(*(dev_t *)key2)) ? 0 : 1;
+}
+
+static unsigned vmw_dev_hash(void *key)
+{
+   return (major(*(dev_t *) key) << 16) | minor(*(dev_t *) key);
+}
 
 /* Called from vmw_drm_create_screen(), creates and initializes the
  * vmw_winsys_screen structure, which is the main entity in this
  * module.
+ * First, check whether a vmw_winsys_screen object already exists for
+ * this device, and in that case return that one, making sure that we
+ * have our own file descriptor open to DRM.
  */
+
 struct vmw_winsys_screen *
 vmw_winsys_create( int fd, boolean use_old_scanout_flag )
 {
-   struct vmw_winsys_screen *vws = CALLOC_STRUCT(vmw_winsys_screen);
+   struct vmw_winsys_screen *vws;
+   struct stat stat_buf;
+
+   if (dev_hash == NULL) {
+      dev_hash = util_hash_table_create(vmw_dev_hash, vmw_dev_compare);
+      if (dev_hash == NULL)
+         return NULL;
+   }
+
+   if (fstat(fd, &stat_buf))
+      return NULL;
+
+   vws = util_hash_table_get(dev_hash, &stat_buf.st_rdev);
+   if (vws) {
+      vws->open_count++;
+      return vws;
+   }
+
+   vws = CALLOC_STRUCT(vmw_winsys_screen);
    if (!vws)
       goto out_no_vws;
 
-   vws->ioctl.drm_fd = fd;
+   vws->device = stat_buf.st_rdev;
+   vws->open_count = 1;
+   vws->ioctl.drm_fd = dup(fd);
    vws->use_old_scanout_flag = use_old_scanout_flag;
+   vws->base.have_gb_dma = TRUE;
 
    if (!vmw_ioctl_init(vws))
       goto out_no_ioctl;
 
+   vws->fence_ops = vmw_fence_ops_create(vws);
+   if (!vws->fence_ops)
+      goto out_no_fence_ops;
+
    if(!vmw_pools_init(vws))
       goto out_no_pools;
 
    if (!vmw_winsys_screen_init_svga(vws))
       goto out_no_svga;
 
+   if (util_hash_table_set(dev_hash, &vws->device, vws) != PIPE_OK)
+      goto out_no_hash_insert;
+
    return vws;
+out_no_hash_insert:
 out_no_svga:
    vmw_pools_cleanup(vws);
 out_no_pools:
+   vws->fence_ops->destroy(vws->fence_ops);
+out_no_fence_ops:
    vmw_ioctl_cleanup(vws);
 out_no_ioctl:
+   close(vws->ioctl.drm_fd);
    FREE(vws);
 out_no_vws:
    return NULL;
@@ -69,7 +121,12 @@ out_no_vws:
 void
 vmw_winsys_destroy(struct vmw_winsys_screen *vws)
 {
-   vmw_pools_cleanup(vws);
-   vmw_ioctl_cleanup(vws);
-   FREE(vws);
+   if (--vws->open_count == 0) {
+      util_hash_table_remove(dev_hash, &vws->device);
+      vmw_pools_cleanup(vws);
+      vws->fence_ops->destroy(vws->fence_ops);
+      vmw_ioctl_cleanup(vws);
+      close(vws->ioctl.drm_fd);
+      FREE(vws);
+   }
 }
diff --git a/src/gallium/winsys/svga/drm/vmw_screen.h b/src/gallium/winsys/svga/drm/vmw_screen.h
index bdc1c7ba3de..f9da3693c58 100644
--- a/src/gallium/winsys/svga/drm/vmw_screen.h
+++ b/src/gallium/winsys/svga/drm/vmw_screen.h
@@ -39,10 +39,13 @@
 #include "pipe/p_state.h"
 
 #include "svga_winsys.h"
+#include "pipebuffer/pb_buffer_fenced.h"
 
 
 #define VMW_GMR_POOL_SIZE (16*1024*1024)
 #define VMW_QUERY_POOL_SIZE (8192)
+#define VMW_DEBUG_FLUSH_STACK 10
+
 /*
  * Something big, but arbitrary. The kernel reports an error if it can't
  * handle this, and the svga driver will resort to multiple partial
@@ -53,6 +56,10 @@
 struct pb_manager;
 struct vmw_region;
 
+struct vmw_cap_3d {
+   boolean has_cap;
+   SVGA3dDevCapResult result;
+};
 
 struct vmw_winsys_screen
 {
@@ -63,7 +70,10 @@ struct vmw_winsys_screen
    struct {
       int drm_fd;
       uint32_t hwversion;
-      uint32_t *buffer;
+      uint32_t num_cap_3d;
+      struct vmw_cap_3d *cap_3d;
+      uint64_t max_mob_memory;
+      uint64_t max_surface_memory;
    } ioctl;
 
    struct {
@@ -72,10 +82,21 @@ struct vmw_winsys_screen
       struct pb_manager *gmr_fenced;
       struct pb_manager *gmr_slab;
       struct pb_manager *gmr_slab_fenced;
-      struct pb_manager *query;
       struct pb_manager *query_mm;
       struct pb_manager *query_fenced;
+      struct pb_manager *mob_fenced;
+      struct pb_manager *mob_cache;
+      struct pb_manager *mob_shader_slab;
+      struct pb_manager *mob_shader_slab_fenced;
    } pools;
+
+   struct pb_fence_ops *fence_ops;
+
+   /*
+    * Screen instances
+    */
+   dev_t device;
+   int open_count;
 };
 
 
@@ -86,6 +107,9 @@ vmw_winsys_screen(struct svga_winsys_screen *base)
 }
 
 /*  */
+uint32_t
+vmw_region_size(struct vmw_region *region);
+
 uint32
 vmw_ioctl_context_create(struct vmw_winsys_screen *vws);
 
@@ -100,6 +124,23 @@ vmw_ioctl_surface_create(struct vmw_winsys_screen *vws,
                               SVGA3dSize size,
                               uint32 numFaces,
                               uint32 numMipLevels);
+uint32
+vmw_ioctl_gb_surface_create(struct vmw_winsys_screen *vws,
+			    SVGA3dSurfaceFlags flags,
+			    SVGA3dSurfaceFormat format,
+			    SVGA3dSize size,
+			    uint32 numFaces,
+			    uint32 numMipLevels,
+                            uint32 buffer_handle,
+			    struct vmw_region **p_region);
+
+int
+vmw_ioctl_gb_surface_ref(struct vmw_winsys_screen *vws,
+                         uint32_t handle,
+                         SVGA3dSurfaceFlags *flags,
+                         SVGA3dSurfaceFormat *format,
+                         uint32_t *numMipLevels,
+                         struct vmw_region **p_region);
 
 void
 vmw_ioctl_surface_destroy(struct vmw_winsys_screen *vws,
@@ -140,12 +181,28 @@ void
 vmw_ioctl_fence_unref(struct vmw_winsys_screen *vws,
 		      uint32_t handle);
 
+uint32
+vmw_ioctl_shader_create(struct vmw_winsys_screen *vws,
+			SVGA3dShaderType type,
+			uint32 code_len);
+void
+vmw_ioctl_shader_destroy(struct vmw_winsys_screen *vws, uint32 shid);
 
+int
+vmw_ioctl_syncforcpu(struct vmw_region *region,
+                     boolean dont_block,
+                     boolean readonly,
+                     boolean allow_cs);
+void
+vmw_ioctl_releasefromcpu(struct vmw_region *region,
+                         boolean readonly,
+                         boolean allow_cs);
 /* Initialize parts of vmw_winsys_screen at startup:
  */
 boolean vmw_ioctl_init(struct vmw_winsys_screen *vws);
 boolean vmw_pools_init(struct vmw_winsys_screen *vws);
 boolean vmw_query_pools_init(struct vmw_winsys_screen *vws);
+boolean vmw_mob_pools_init(struct vmw_winsys_screen *vws);
 boolean vmw_winsys_screen_init_svga(struct vmw_winsys_screen *vws);
 
 void vmw_ioctl_cleanup(struct vmw_winsys_screen *vws);
@@ -156,4 +213,13 @@ void vmw_winsys_destroy(struct vmw_winsys_screen *sws);
 void vmw_winsys_screen_set_throttling(struct pipe_screen *screen,
 				      uint32_t throttle_us);
 
+struct pb_manager *
+simple_fenced_bufmgr_create(struct pb_manager *provider,
+			    struct pb_fence_ops *ops);
+void
+vmw_fences_signal(struct pb_fence_ops *fence_ops,
+                  uint32_t signaled,
+                  uint32_t emitted,
+                  boolean has_emitted);
+
 #endif /* VMW_SCREEN_H_ */
diff --git a/src/gallium/winsys/svga/drm/vmw_screen_dri.c b/src/gallium/winsys/svga/drm/vmw_screen_dri.c
index a17cdf7cf92..0f5a9552d85 100644
--- a/src/gallium/winsys/svga/drm/vmw_screen_dri.c
+++ b/src/gallium/winsys/svga/drm/vmw_screen_dri.c
@@ -32,7 +32,9 @@
 #include "vmw_context.h"
 #include "vmw_screen.h"
 #include "vmw_surface.h"
+#include "vmw_buffer.h"
 #include "svga_drm_public.h"
+#include "svga3d_surfacedefs.h"
 
 #include "state_tracker/drm_driver.h"
 
@@ -52,6 +54,11 @@ static struct svga_winsys_surface *
 vmw_drm_surface_from_handle(struct svga_winsys_screen *sws,
 			    struct winsys_handle *whandle,
 			    SVGA3dSurfaceFormat *format);
+
+static struct svga_winsys_surface *
+vmw_drm_gb_surface_from_handle(struct svga_winsys_screen *sws,
+                               struct winsys_handle *whandle,
+                               SVGA3dSurfaceFormat *format);
 static boolean
 vmw_drm_surface_get_handle(struct svga_winsys_screen *sws,
 			   struct svga_winsys_surface *surface,
@@ -109,7 +116,8 @@ svga_drm_winsys_screen_create(int fd)
       goto out_no_vws;
 
    /* XXX do this properly */
-   vws->base.surface_from_handle = vmw_drm_surface_from_handle;
+   vws->base.surface_from_handle = vws->base.have_gb_objects ?
+      vmw_drm_gb_surface_from_handle : vmw_drm_surface_from_handle;
    vws->base.surface_get_handle = vmw_drm_surface_get_handle;
 
    return &vws->base;
@@ -150,6 +158,83 @@ vmw_dri1_intersect_src_bbox(struct drm_clip_rect *dst,
    return TRUE;
 }
 
+/**
+ * vmw_drm_gb_surface_from_handle - Create a shared surface
+ *
+ * @sws: Screen to register the surface with.
+ * @whandle: struct winsys_handle identifying the kernel surface object
+ * @format: On successful return points to a value describing the
+ * surface format.
+ *
+ * Returns a refcounted pointer to a struct svga_winsys_surface
+ * embedded in a struct vmw_svga_winsys_surface on success or NULL
+ * on failure.
+ */
+static struct svga_winsys_surface *
+vmw_drm_gb_surface_from_handle(struct svga_winsys_screen *sws,
+                               struct winsys_handle *whandle,
+                               SVGA3dSurfaceFormat *format)
+{
+    struct vmw_svga_winsys_surface *vsrf;
+    struct svga_winsys_surface *ssrf;
+    struct vmw_winsys_screen *vws = vmw_winsys_screen(sws);
+    SVGA3dSurfaceFlags flags;
+    uint32_t mip_levels;
+    struct vmw_buffer_desc desc;
+    struct pb_manager *provider = vws->pools.gmr;
+    struct pb_buffer *pb_buf;
+    int ret;
+
+    ret = vmw_ioctl_gb_surface_ref(vws, whandle->handle, &flags, format,
+                                   &mip_levels, &desc.region);
+
+    if (ret) {
+	fprintf(stderr, "Failed referencing shared surface. SID %d.\n"
+		"Error %d (%s).\n",
+		whandle->handle, ret, strerror(-ret));
+	return NULL;
+    }
+
+    if (mip_levels != 1) {
+       fprintf(stderr, "Incorrect number of mipmap levels on shared surface."
+               " SID %d, levels %d\n",
+               whandle->handle, mip_levels);
+       goto out_mip;
+    }
+
+    vsrf = CALLOC_STRUCT(vmw_svga_winsys_surface);
+    if (!vsrf)
+	goto out_mip;
+
+    pipe_reference_init(&vsrf->refcnt, 1);
+    p_atomic_set(&vsrf->validated, 0);
+    vsrf->screen = vws;
+    vsrf->sid = whandle->handle;
+    vsrf->size = vmw_region_size(desc.region);
+
+    /*
+     * Synchronize backing buffers of shared surfaces using the
+     * kernel, since we don't pass fence objects around between
+     * processes.
+     */
+    desc.pb_desc.alignment = 4096;
+    desc.pb_desc.usage = VMW_BUFFER_USAGE_SHARED | VMW_BUFFER_USAGE_SYNC;
+    pb_buf = provider->create_buffer(provider, vsrf->size, &desc.pb_desc);
+    vsrf->buf = vmw_svga_winsys_buffer_wrap(pb_buf);
+    if (!vsrf->buf)
+       goto out_no_buf;
+    ssrf = svga_winsys_surface(vsrf);
+
+    return ssrf;
+
+out_no_buf:
+    FREE(vsrf);
+out_mip:
+    vmw_ioctl_region_destroy(desc.region);
+    vmw_ioctl_surface_destroy(vws, whandle->handle);
+    return NULL;
+}
+
 static struct svga_winsys_surface *
 vmw_drm_surface_from_handle(struct svga_winsys_screen *sws,
 			    struct winsys_handle *whandle,
@@ -162,6 +247,7 @@ vmw_drm_surface_from_handle(struct svga_winsys_screen *sws,
     struct drm_vmw_surface_arg *req = &arg.req;
     struct drm_vmw_surface_create_req *rep = &arg.rep;
     uint32_t handle = 0;
+    SVGA3dSize size;
     int ret;
     int i;
 
@@ -187,6 +273,7 @@ vmw_drm_surface_from_handle(struct svga_winsys_screen *sws,
 
     memset(&arg, 0, sizeof(arg));
     req->sid = handle;
+    rep->size_addr = (size_t)&size;
 
     ret = drmCommandWriteRead(vws->ioctl.drm_fd, DRM_VMW_REF_SURFACE,
 			      &arg, sizeof(arg));
@@ -235,6 +322,11 @@ vmw_drm_surface_from_handle(struct svga_winsys_screen *sws,
     ssrf = svga_winsys_surface(vsrf);
     *format = rep->format;
 
+    /* Estimate usage, for early flushing. */
+    vsrf->size = svga3dsurface_get_serialized_size(rep->format, size,
+                                                   rep->mip_levels[0],
+                                                   FALSE);
+
     return ssrf;
 
 out_mip:
diff --git a/src/gallium/winsys/svga/drm/vmw_screen_ioctl.c b/src/gallium/winsys/svga/drm/vmw_screen_ioctl.c
index 36888dc2cce..501c047c32a 100644
--- a/src/gallium/winsys/svga/drm/vmw_screen_ioctl.c
+++ b/src/gallium/winsys/svga/drm/vmw_screen_ioctl.c
@@ -43,6 +43,7 @@
 #include "xf86drm.h"
 #include "vmwgfx_drm.h"
 #include "svga3d_caps.h"
+#include "svga3d_reg.h"
 
 #include "os/os_mman.h"
 
@@ -51,7 +52,6 @@
 
 struct vmw_region
 {
-   SVGAGuestPtr ptr;
    uint32_t handle;
    uint64_t map_handle;
    void *data;
@@ -66,6 +66,13 @@ struct vmw_region
  */
 #define SVGA3D_SURFACE_HINT_SCANOUT (1 << 9)
 
+
+uint32_t
+vmw_region_size(struct vmw_region *region)
+{
+   return region->size;
+}
+
 uint32
 vmw_ioctl_context_create(struct vmw_winsys_screen *vws)
 {
@@ -168,6 +175,139 @@ vmw_ioctl_surface_create(struct vmw_winsys_screen *vws,
    return rep->sid;
 }
 
+
+uint32
+vmw_ioctl_gb_surface_create(struct vmw_winsys_screen *vws,
+			    SVGA3dSurfaceFlags flags,
+			    SVGA3dSurfaceFormat format,
+			    SVGA3dSize size,
+			    uint32_t numFaces,
+			    uint32_t numMipLevels,
+                            uint32_t buffer_handle,
+			    struct vmw_region **p_region)
+{
+   union drm_vmw_gb_surface_create_arg s_arg;
+   struct drm_vmw_gb_surface_create_req *req = &s_arg.req;
+   struct drm_vmw_gb_surface_create_rep *rep = &s_arg.rep;
+   struct vmw_region *region = NULL;
+   int ret;
+
+   vmw_printf("%s flags %d format %d\n", __FUNCTION__, flags, format);
+
+   if (p_region) {
+      region = CALLOC_STRUCT(vmw_region);
+      if (!region)
+         return SVGA3D_INVALID_ID;
+   }
+
+   memset(&s_arg, 0, sizeof(s_arg));
+   if (flags & SVGA3D_SURFACE_HINT_SCANOUT) {
+      req->svga3d_flags = (uint32_t) (flags & ~SVGA3D_SURFACE_HINT_SCANOUT);
+      req->drm_surface_flags = drm_vmw_surface_flag_scanout;
+   } else {
+      req->svga3d_flags = (uint32_t) flags;
+   }
+   req->format = (uint32_t) format;
+   req->drm_surface_flags |= drm_vmw_surface_flag_shareable; 
+   req->drm_surface_flags |= drm_vmw_surface_flag_create_buffer; 
+
+   assert(numFaces * numMipLevels < DRM_VMW_MAX_SURFACE_FACES*
+	  DRM_VMW_MAX_MIP_LEVELS);
+   req->base_size.width = size.width;
+   req->base_size.height = size.height;
+   req->base_size.depth = size.depth;
+   req->mip_levels = numMipLevels;
+   req->multisample_count = 0;
+   req->autogen_filter = SVGA3D_TEX_FILTER_NONE;
+   if (buffer_handle)
+      req->buffer_handle = buffer_handle;
+   else
+      req->buffer_handle = SVGA3D_INVALID_ID;
+
+   ret = drmCommandWriteRead(vws->ioctl.drm_fd, DRM_VMW_GB_SURFACE_CREATE,
+			     &s_arg, sizeof(s_arg));
+
+   if (ret)
+      goto out_fail_create;
+
+   if (p_region) {
+      region->handle = rep->buffer_handle;
+      region->map_handle = rep->buffer_map_handle;
+      region->drm_fd = vws->ioctl.drm_fd;
+      region->size = rep->backup_size;
+      *p_region = region;
+   }
+
+   vmw_printf("Surface id is %d\n", rep->sid);
+   return rep->handle;
+
+out_fail_create:
+   if (region)
+      FREE(region);
+   return SVGA3D_INVALID_ID;
+}
+
+/**
+ * vmw_ioctl_gb_surface_ref - Put a reference on a guest-backed surface and
+ * get surface information
+ *
+ * @vws: Screen to register the reference on
+ * @handle: Kernel handle of the guest-backed surface
+ * @flags: flags used when the surface was created
+ * @format: Format used when the surface was created
+ * @numMipLevels: Number of mipmap levels of the surface
+ * @p_region: On successful return points to a newly allocated
+ * struct vmw_region holding a reference to the surface backup buffer.
+ *
+ * Returns 0 on success, a system error on failure.
+ */
+int
+vmw_ioctl_gb_surface_ref(struct vmw_winsys_screen *vws,
+                         uint32_t handle,
+                         SVGA3dSurfaceFlags *flags,
+                         SVGA3dSurfaceFormat *format,
+                         uint32_t *numMipLevels,
+                         struct vmw_region **p_region)
+{
+   union drm_vmw_gb_surface_reference_arg s_arg;
+   struct drm_vmw_surface_arg *req = &s_arg.req;
+   struct drm_vmw_gb_surface_ref_rep *rep = &s_arg.rep;
+   struct vmw_region *region = NULL;
+   int ret;
+
+   vmw_printf("%s flags %d format %d\n", __FUNCTION__, flags, format);
+
+   assert(p_region != NULL);
+   region = CALLOC_STRUCT(vmw_region);
+   if (!region)
+      return -ENOMEM;
+
+   memset(&s_arg, 0, sizeof(s_arg));
+   req->sid = handle;
+
+   ret = drmCommandWriteRead(vws->ioctl.drm_fd, DRM_VMW_GB_SURFACE_REF,
+			     &s_arg, sizeof(s_arg));
+
+   if (ret)
+      goto out_fail_ref;
+
+   region->handle = rep->crep.buffer_handle;
+   region->map_handle = rep->crep.buffer_map_handle;
+   region->drm_fd = vws->ioctl.drm_fd;
+   region->size = rep->crep.backup_size;
+   *p_region = region;
+
+   *flags = rep->creq.svga3d_flags;
+   *format = rep->creq.format;
+   *numMipLevels = rep->creq.mip_levels;
+
+   return 0;
+out_fail_ref:
+   if (region)
+      FREE(region);
+   return ret;
+}
+
 void
 vmw_ioctl_surface_destroy(struct vmw_winsys_screen *vws, uint32 sid)
 {
@@ -238,8 +378,11 @@ vmw_ioctl_command(struct vmw_winsys_screen *vws, int32_t cid,
 	 *pfence = NULL;
    } else {
       if (pfence) {
-	 *pfence = vmw_fence_create(rep.handle, rep.mask);
+         vmw_fences_signal(vws->fence_ops, rep.passed_seqno, rep.seqno,
+                           TRUE);
 
+	 *pfence = vmw_fence_create(vws->fence_ops, rep.handle,
+				    rep.seqno, rep.mask);
 	 if (*pfence == NULL) {
 	    /*
 	     * Fence creation failed. Need to sync.
@@ -279,8 +422,6 @@ vmw_ioctl_region_create(struct vmw_winsys_screen *vws, uint32_t size)
       goto out_err1;
    }
 
-   region->ptr.gmrId = rep->cur_gmr_id;
-   region->ptr.offset = rep->cur_gmr_offset;
    region->data = NULL;
    region->handle = rep->handle;
    region->map_handle = rep->map_handle;
@@ -321,7 +462,8 @@ vmw_ioctl_region_destroy(struct vmw_region *region)
 SVGAGuestPtr
 vmw_ioctl_region_ptr(struct vmw_region *region)
 {
-   return region->ptr;
+   SVGAGuestPtr ptr = {region->handle, 0};
+   return ptr;
 }
 
 void *
@@ -356,6 +498,69 @@ vmw_ioctl_region_unmap(struct vmw_region *region)
    --region->map_count;
 }
 
+/**
+ * vmw_ioctl_syncforcpu - Synchronize a buffer object for CPU usage
+ *
+ * @region: Pointer to a struct vmw_region representing the buffer object.
+ * @dont_block: Dont wait for GPU idle, but rather return -EBUSY if the
+ * GPU is busy with the buffer object.
+ * @readonly: Hint that the CPU access is read-only.
+ * @allow_cs: Allow concurrent command submission while the buffer is
+ * synchronized for CPU. If FALSE command submissions referencing the
+ * buffer will block until a corresponding call to vmw_ioctl_releasefromcpu.
+ *
+ * This function idles any GPU activities touching the buffer and blocks
+ * command submission of commands referencing the buffer, even from
+ * other processes.
+ */
+int
+vmw_ioctl_syncforcpu(struct vmw_region *region,
+                     boolean dont_block,
+                     boolean readonly,
+                     boolean allow_cs)
+{
+   struct drm_vmw_synccpu_arg arg;
+
+   memset(&arg, 0, sizeof(arg));
+   arg.op = drm_vmw_synccpu_grab;
+   arg.handle = region->handle;
+   arg.flags = drm_vmw_synccpu_read;
+   if (!readonly)
+      arg.flags |= drm_vmw_synccpu_write;
+   if (dont_block)
+      arg.flags |= drm_vmw_synccpu_dontblock;
+   if (allow_cs)
+      arg.flags |= drm_vmw_synccpu_allow_cs;
+
+   return drmCommandWrite(region->drm_fd, DRM_VMW_SYNCCPU, &arg, sizeof(arg));
+}
+
+/**
+ * vmw_ioctl_releasefromcpu - Undo a previous syncforcpu.
+ *
+ * @region: Pointer to a struct vmw_region representing the buffer object.
+ * @readonly: Should hold the same value as the matching syncforcpu call.
+ * @allow_cs: Should hold the same value as the matching syncforcpu call.
+ */
+void
+vmw_ioctl_releasefromcpu(struct vmw_region *region,
+                         boolean readonly,
+                         boolean allow_cs)
+{
+   struct drm_vmw_synccpu_arg arg;
+
+   memset(&arg, 0, sizeof(arg));
+   arg.op = drm_vmw_synccpu_release;
+   arg.handle = region->handle;
+   arg.flags = drm_vmw_synccpu_read;
+   if (!readonly)
+      arg.flags |= drm_vmw_synccpu_write;
+   if (allow_cs)
+      arg.flags |= drm_vmw_synccpu_allow_cs;
+
+   (void) drmCommandWrite(region->drm_fd, DRM_VMW_SYNCCPU, &arg, sizeof(arg));
+}
+
 void
 vmw_ioctl_fence_unref(struct vmw_winsys_screen *vws,
 		      uint32_t handle)
@@ -405,6 +610,8 @@ vmw_ioctl_fence_signalled(struct vmw_winsys_screen *vws,
    if (ret != 0)
       return ret;
 
+   vmw_fences_signal(vws->fence_ops, arg.passed_seqno, 0, FALSE);
+
    return (arg.signaled) ? 0 : -1;
 }
 
@@ -435,6 +642,113 @@ vmw_ioctl_fence_finish(struct vmw_winsys_screen *vws,
    return 0;
 }
 
+uint32
+vmw_ioctl_shader_create(struct vmw_winsys_screen *vws,
+			SVGA3dShaderType type,
+			uint32 code_len)
+{
+   struct drm_vmw_shader_create_arg sh_arg;
+   int ret;
+
+   VMW_FUNC;
+
+   memset(&sh_arg, 0, sizeof(sh_arg));
+
+   sh_arg.size = code_len;
+   sh_arg.buffer_handle = SVGA3D_INVALID_ID;
+   sh_arg.shader_handle = SVGA3D_INVALID_ID;
+   switch (type) {
+   case SVGA3D_SHADERTYPE_VS:
+      sh_arg.shader_type = drm_vmw_shader_type_vs;
+      break;
+   case SVGA3D_SHADERTYPE_PS:
+      sh_arg.shader_type = drm_vmw_shader_type_ps;
+      break;
+   default:
+      assert(!"Invalid shader type.");
+      break;
+   }
+
+   ret = drmCommandWriteRead(vws->ioctl.drm_fd, DRM_VMW_CREATE_SHADER,
+			     &sh_arg, sizeof(sh_arg));
+
+   if (ret)
+      return SVGA3D_INVALID_ID;
+
+   return sh_arg.shader_handle;
+}
+
+void
+vmw_ioctl_shader_destroy(struct vmw_winsys_screen *vws, uint32 shid)
+{
+   struct drm_vmw_shader_arg sh_arg;
+
+   VMW_FUNC;
+
+   memset(&sh_arg, 0, sizeof(sh_arg));
+   sh_arg.handle = shid;
+
+   (void)drmCommandWrite(vws->ioctl.drm_fd, DRM_VMW_UNREF_SHADER,
+			 &sh_arg, sizeof(sh_arg));
+
+}
+
+static int
+vmw_ioctl_parse_caps(struct vmw_winsys_screen *vws,
+		     const uint32_t *cap_buffer)
+{
+   int i;
+
+   if (vws->base.have_gb_objects) {
+      for (i = 0; i < vws->ioctl.num_cap_3d; ++i) {
+	 vws->ioctl.cap_3d[i].has_cap = TRUE;
+	 vws->ioctl.cap_3d[i].result.u = cap_buffer[i];
+      }
+      return 0;
+   } else {
+      const uint32 *capsBlock;
+      const SVGA3dCapsRecord *capsRecord = NULL;
+      uint32 offset;
+      const SVGA3dCapPair *capArray;
+      int numCaps, index;
+
+      /*
+       * Search linearly through the caps block records for the specified type.
+       */
+      capsBlock = cap_buffer;
+      for (offset = 0; capsBlock[offset] != 0; offset += capsBlock[offset]) {
+	 const SVGA3dCapsRecord *record;
+	 assert(offset < SVGA_FIFO_3D_CAPS_SIZE);
+	 record = (const SVGA3dCapsRecord *) (capsBlock + offset);
+	 if ((record->header.type >= SVGA3DCAPS_RECORD_DEVCAPS_MIN) &&
+	     (record->header.type <= SVGA3DCAPS_RECORD_DEVCAPS_MAX) &&
+	     (!capsRecord || (record->header.type > capsRecord->header.type))) {
+	    capsRecord = record;
+	 }
+      }
+
+      if(!capsRecord)
+	 return -1;
+
+      /*
+       * Calculate the number of caps from the size of the record.
+       */
+      capArray = (const SVGA3dCapPair *) capsRecord->data;
+      numCaps = (int) ((capsRecord->header.length * sizeof(uint32) -
+			sizeof capsRecord->header) / (2 * sizeof(uint32)));
+
+      for (i = 0; i < numCaps; i++) {
+	 index = capArray[i][0];
+	 if (index < vws->ioctl.num_cap_3d) {
+	    vws->ioctl.cap_3d[index].has_cap = TRUE;
+	    vws->ioctl.cap_3d[index].result.u = capArray[i][1];
+	 } else {
+	    debug_printf("Unknown devcaps seen: %d\n", index);
+	 }
+      }
+   }
+   return 0;
+}
 
 boolean
 vmw_ioctl_init(struct vmw_winsys_screen *vws)
@@ -443,9 +757,19 @@ vmw_ioctl_init(struct vmw_winsys_screen *vws)
    struct drm_vmw_get_3d_cap_arg cap_arg;
    unsigned int size;
    int ret;
+   uint32_t *cap_buffer;
+   drmVersionPtr version;
+   boolean drm_gb_capable;
 
    VMW_FUNC;
 
+   version = drmGetVersion(vws->ioctl.drm_fd);
+   if (!version)
+      goto out_no_version;
+
+   drm_gb_capable = version->version_major > 2 ||
+      (version->version_major == 2 && version->version_minor > 4);
+
    memset(&gp_arg, 0, sizeof(gp_arg));
    gp_arg.param = DRM_VMW_PARAM_3D;
    ret = drmCommandWriteRead(vws->ioctl.drm_fd, DRM_VMW_GET_PARAM,
@@ -466,15 +790,78 @@ vmw_ioctl_init(struct vmw_winsys_screen *vws)
    }
    vws->ioctl.hwversion = gp_arg.value;
 
-   size = SVGA_FIFO_3D_CAPS_SIZE * sizeof(uint32_t);
-   vws->ioctl.buffer = calloc(1, size);
-   if (!vws->ioctl.buffer) {
+   memset(&gp_arg, 0, sizeof(gp_arg));
+   gp_arg.param = DRM_VMW_PARAM_HW_CAPS;
+   ret = drmCommandWriteRead(vws->ioctl.drm_fd, DRM_VMW_GET_PARAM,
+                             &gp_arg, sizeof(gp_arg));
+   if (ret)
+      vws->base.have_gb_objects = FALSE;
+   else
+      vws->base.have_gb_objects =
+         !!(gp_arg.value & (uint64_t) SVGA_CAP_GBOBJECTS);
+   
+   if (vws->base.have_gb_objects && !drm_gb_capable)
+      goto out_no_3d;
+
+   if (vws->base.have_gb_objects) {
+      memset(&gp_arg, 0, sizeof(gp_arg));
+      gp_arg.param = DRM_VMW_PARAM_3D_CAPS_SIZE;
+      ret = drmCommandWriteRead(vws->ioctl.drm_fd, DRM_VMW_GET_PARAM,
+                                &gp_arg, sizeof(gp_arg));
+      if (ret)
+         size = SVGA_FIFO_3D_CAPS_SIZE * sizeof(uint32_t);
+      else
+         size = gp_arg.value;
+   
+      if (vws->base.have_gb_objects)
+         vws->ioctl.num_cap_3d = size / sizeof(uint32_t);
+      else
+         vws->ioctl.num_cap_3d = SVGA3D_DEVCAP_MAX;
+
+
+      memset(&gp_arg, 0, sizeof(gp_arg));
+      gp_arg.param = DRM_VMW_PARAM_MAX_MOB_MEMORY;
+      ret = drmCommandWriteRead(vws->ioctl.drm_fd, DRM_VMW_GET_PARAM,
+                                &gp_arg, sizeof(gp_arg));
+      if (ret) {
+         /* Just guess a large enough value. */
+         vws->ioctl.max_mob_memory = 256*1024*1024;
+      } else {
+         vws->ioctl.max_mob_memory = gp_arg.value;
+      }
+      /* Never early flush surfaces, mobs do accounting. */
+      vws->ioctl.max_surface_memory = -1;
+   } else {
+      vws->ioctl.num_cap_3d = SVGA3D_DEVCAP_MAX;
+
+      memset(&gp_arg, 0, sizeof(gp_arg));
+      gp_arg.param = DRM_VMW_PARAM_MAX_SURF_MEMORY;
+      ret = drmCommandWriteRead(vws->ioctl.drm_fd, DRM_VMW_GET_PARAM,
+                                &gp_arg, sizeof(gp_arg));
+      if (ret) {
+         /* Just guess a large enough value, around 800mb. */
+         vws->ioctl.max_surface_memory = 0x300000000;
+      } else {
+         vws->ioctl.max_surface_memory = gp_arg.value;
+      }
+      size = SVGA_FIFO_3D_CAPS_SIZE * sizeof(uint32_t);
+   }
+
+   cap_buffer = calloc(1, size);
+   if (!cap_buffer) {
       debug_printf("Failed alloc fifo 3D caps buffer.\n");
       goto out_no_3d;
    }
 
+   vws->ioctl.cap_3d = calloc(vws->ioctl.num_cap_3d, 
+			      sizeof(*vws->ioctl.cap_3d));
+   if (!vws->ioctl.cap_3d) {
+      debug_printf("Failed alloc fifo 3D caps buffer.\n");
+      goto out_no_caparray;
+   }
+      
    memset(&cap_arg, 0, sizeof(cap_arg));
-   cap_arg.buffer = (uint64_t) (unsigned long) (vws->ioctl.buffer);
+   cap_arg.buffer = (uint64_t) (unsigned long) (cap_buffer);
    cap_arg.max_size = size;
 
    ret = drmCommandWrite(vws->ioctl.drm_fd, DRM_VMW_GET_3D_CAP,
@@ -486,11 +873,24 @@ vmw_ioctl_init(struct vmw_winsys_screen *vws)
       goto out_no_caps;
    }
 
+   ret = vmw_ioctl_parse_caps(vws, cap_buffer);
+   if (ret) {
+      debug_printf("Failed to parse 3D capabilities"
+		   " (%i, %s).\n", ret, strerror(-ret));
+      goto out_no_caps;
+   }
+   free(cap_buffer);
+   drmFreeVersion(version);
    vmw_printf("%s OK\n", __FUNCTION__);
    return TRUE;
   out_no_caps:
-   free(vws->ioctl.buffer);
+   free(vws->ioctl.cap_3d);
+  out_no_caparray:
+   free(cap_buffer);
   out_no_3d:
+   drmFreeVersion(version);
+  out_no_version:
+   vws->ioctl.num_cap_3d = 0;
    debug_printf("%s Failed\n", __FUNCTION__);
    return FALSE;
 }
diff --git a/src/gallium/winsys/svga/drm/vmw_screen_pools.c b/src/gallium/winsys/svga/drm/vmw_screen_pools.c
index 2525604f952..7f7b77982af 100644
--- a/src/gallium/winsys/svga/drm/vmw_screen_pools.c
+++ b/src/gallium/winsys/svga/drm/vmw_screen_pools.c
@@ -32,13 +32,6 @@
 #include "pipebuffer/pb_buffer.h"
 #include "pipebuffer/pb_bufmgr.h"
 
-/*
- * TODO: Have the query pool always ask the fence manager for
- * SVGA_FENCE_FLAG_QUERY signaled. Unfortunately, pb_fenced doesn't
- * support that currently, so we'd have to create a separate
- * pb_fence_ops wrapper that does this implicitly.
- */
-
 /**
  * vmw_pools_cleanup - Destroy the buffer pools.
  *
@@ -47,20 +40,32 @@
 void
 vmw_pools_cleanup(struct vmw_winsys_screen *vws)
 {
-   if(vws->pools.gmr_fenced)
-      vws->pools.gmr_fenced->destroy(vws->pools.gmr_fenced);
+   if (vws->pools.mob_shader_slab_fenced)
+      vws->pools.mob_shader_slab_fenced->destroy
+         (vws->pools.mob_shader_slab_fenced);
+   if (vws->pools.mob_shader_slab)
+      vws->pools.mob_shader_slab->destroy(vws->pools.mob_shader_slab);
+   if (vws->pools.mob_fenced)
+      vws->pools.mob_fenced->destroy(vws->pools.mob_fenced);
+   if (vws->pools.mob_cache)
+      vws->pools.mob_cache->destroy(vws->pools.mob_cache);
+
    if (vws->pools.query_fenced)
       vws->pools.query_fenced->destroy(vws->pools.query_fenced);
+   if (vws->pools.query_mm)
+      vws->pools.query_mm->destroy(vws->pools.query_mm);
 
-   /* gmr_mm pool is already destroyed above */
-
+   if(vws->pools.gmr_fenced)
+      vws->pools.gmr_fenced->destroy(vws->pools.gmr_fenced);
+   if (vws->pools.gmr_mm)
+      vws->pools.gmr_mm->destroy(vws->pools.gmr_mm);
    if (vws->pools.gmr_slab_fenced)
       vws->pools.gmr_slab_fenced->destroy(vws->pools.gmr_slab_fenced);
+   if (vws->pools.gmr_slab)
+      vws->pools.gmr_slab->destroy(vws->pools.gmr_slab);
 
    if(vws->pools.gmr)
       vws->pools.gmr->destroy(vws->pools.gmr);
-   if(vws->pools.query)
-      vws->pools.query->destroy(vws->pools.query);
 }
 
 
@@ -77,21 +82,14 @@ vmw_pools_cleanup(struct vmw_winsys_screen *vws)
 boolean
 vmw_query_pools_init(struct vmw_winsys_screen *vws)
 {
-   vws->pools.query = vmw_gmr_bufmgr_create(vws);
-   if(!vws->pools.query)
-      return FALSE;
-
-   vws->pools.query_mm = mm_bufmgr_create(vws->pools.query,
+   vws->pools.query_mm = mm_bufmgr_create(vws->pools.gmr,
 					  VMW_QUERY_POOL_SIZE,
 					  3 /* 8 alignment */);
-   if(!vws->pools.query_mm)
-      goto out_no_query_mm;
+   if (!vws->pools.query_mm)
+      return FALSE;
 
-   vws->pools.query_fenced = fenced_bufmgr_create(
-      vws->pools.query_mm,
-      vmw_fence_ops_create(vws),
-      VMW_QUERY_POOL_SIZE,
-      ~0);
+   vws->pools.query_fenced = simple_fenced_bufmgr_create(
+      vws->pools.query_mm, vws->fence_ops);
 
    if(!vws->pools.query_fenced)
       goto out_no_query_fenced;
@@ -100,8 +98,60 @@ vmw_query_pools_init(struct vmw_winsys_screen *vws)
 
   out_no_query_fenced:
    vws->pools.query_mm->destroy(vws->pools.query_mm);
-  out_no_query_mm:
-   vws->pools.query->destroy(vws->pools.query);
+   return FALSE;
+}
+
+/**
+ * vmw_mob_pool_init - Create a pool of fenced kernel buffers.
+ *
+ * @vws: Pointer to a struct vmw_winsys_screen.
+ *
+ * Typically this pool should be created on demand when we
+ * detect that the app will be using MOB buffers.
+ */
+boolean
+vmw_mob_pools_init(struct vmw_winsys_screen *vws)
+{
+   struct pb_desc desc;
+
+   vws->pools.mob_cache = 
+      pb_cache_manager_create(vws->pools.gmr, 100000, 2,
+                              VMW_BUFFER_USAGE_SHARED);
+   if (!vws->pools.mob_cache)
+      return FALSE;
+
+   vws->pools.mob_fenced = 
+      simple_fenced_bufmgr_create(vws->pools.mob_cache,
+                                  vws->fence_ops);
+   if(!vws->pools.mob_fenced)
+      goto out_no_mob_fenced;
+   
+   desc.alignment = 64;
+   desc.usage = ~(SVGA_BUFFER_USAGE_PINNED | VMW_BUFFER_USAGE_SHARED |
+                  VMW_BUFFER_USAGE_SYNC);
+   vws->pools.mob_shader_slab =
+      pb_slab_range_manager_create(vws->pools.mob_cache,
+                                   64,
+                                   8192,
+                                   16384,
+                                   &desc);
+   if(!vws->pools.mob_shader_slab)
+      goto out_no_mob_shader_slab;
+
+   vws->pools.mob_shader_slab_fenced =
+      simple_fenced_bufmgr_create(vws->pools.mob_shader_slab,
+				  vws->fence_ops);
+   if(!vws->pools.mob_fenced)
+      goto out_no_mob_shader_slab_fenced;
+
+   return TRUE;
+
+ out_no_mob_shader_slab_fenced:
+   vws->pools.mob_shader_slab->destroy(vws->pools.mob_shader_slab);
+ out_no_mob_shader_slab:
+   vws->pools.mob_fenced->destroy(vws->pools.mob_fenced);
+ out_no_mob_fenced:
+   vws->pools.mob_cache->destroy(vws->pools.mob_cache);
    return FALSE;
 }
 
@@ -119,33 +169,27 @@ vmw_pools_init(struct vmw_winsys_screen *vws)
    if(!vws->pools.gmr)
       goto error;
 
-   vws->pools.gmr_mm = mm_bufmgr_create(vws->pools.gmr,
-                                        VMW_GMR_POOL_SIZE,
-                                        12 /* 4096 alignment */);
-   if(!vws->pools.gmr_mm)
-      goto error;
+   if ((vws->base.have_gb_objects && vws->base.have_gb_dma) ||
+       !vws->base.have_gb_objects) {
+      /*
+       * A managed pool for DMA buffers.
+       */
+      vws->pools.gmr_mm = mm_bufmgr_create(vws->pools.gmr,
+                                           VMW_GMR_POOL_SIZE,
+                                           12 /* 4096 alignment */);
+      if(!vws->pools.gmr_mm)
+         goto error;
 
-   /*
-    * We disallow "CPU" buffers to be created by the fenced_bufmgr_create,
-    * because that defers "GPU" buffer creation to buffer validation,
-    * and at buffer validation we have no means of handling failures
-    * due to pools space shortage or fragmentation. Effectively this
-    * makes sure all failures are reported immediately on buffer allocation,
-    * and we can revert to allocating directly from the kernel.
-    */
-   vws->pools.gmr_fenced = fenced_bufmgr_create(
-      vws->pools.gmr_mm,
-      vmw_fence_ops_create(vws),
-      VMW_GMR_POOL_SIZE,
-      0);
+      vws->pools.gmr_fenced = simple_fenced_bufmgr_create
+         (vws->pools.gmr_mm, vws->fence_ops);
 
 #ifdef DEBUG
-   vws->pools.gmr_fenced = pb_debug_manager_create(vws->pools.gmr_fenced,
-						   4096,
-						   4096);
+      vws->pools.gmr_fenced = pb_debug_manager_create(vws->pools.gmr_fenced,
+                                                      4096,
+                                                      4096);
 #endif
-   if(!vws->pools.gmr_fenced)
-      goto error;
+      if(!vws->pools.gmr_fenced)
+         goto error;
 
    /*
     * The slab pool allocates buffers directly from the kernel except
@@ -153,30 +197,33 @@ vmw_pools_init(struct vmw_winsys_screen *vws)
     * not to waste memory, since a kernel buffer is a minimum 4096 bytes.
     *
     * Here we use it only for emergency in the case our pre-allocated
-    * buffer pool runs out of memory.
+    * managed buffer pool runs out of memory.
     */
-   desc.alignment = 64;
-   desc.usage = ~0;
-   vws->pools.gmr_slab = pb_slab_range_manager_create(vws->pools.gmr,
-						      64,
-						      8192,
-						      16384,
-						      &desc);
-   if (!vws->pools.gmr_slab)
-       goto error;
-
-   vws->pools.gmr_slab_fenced =
-       fenced_bufmgr_create(vws->pools.gmr_slab,
-			    vmw_fence_ops_create(vws),
-			    VMW_MAX_BUFFER_SIZE,
-			    0);
-
-   if (!vws->pools.gmr_slab_fenced)
-       goto error;
+
+      desc.alignment = 64;
+      desc.usage = ~(SVGA_BUFFER_USAGE_PINNED | SVGA_BUFFER_USAGE_SHADER |
+                     VMW_BUFFER_USAGE_SHARED | VMW_BUFFER_USAGE_SYNC);
+      vws->pools.gmr_slab = pb_slab_range_manager_create(vws->pools.gmr,
+                                                         64,
+                                                         8192,
+                                                         16384,
+                                                         &desc);
+      if (!vws->pools.gmr_slab)
+         goto error;
+
+      vws->pools.gmr_slab_fenced =
+         simple_fenced_bufmgr_create(vws->pools.gmr_slab, vws->fence_ops);
+
+      if (!vws->pools.gmr_slab_fenced)
+         goto error;
+   }
 
    vws->pools.query_fenced = NULL;
    vws->pools.query_mm = NULL;
-   vws->pools.query = NULL;
+   vws->pools.mob_cache = NULL;
+
+   if (vws->base.have_gb_objects && !vmw_mob_pools_init(vws))
+      goto error;
 
    return TRUE;
 
@@ -184,4 +231,3 @@ error:
    vmw_pools_cleanup(vws);
    return FALSE;
 }
-
diff --git a/src/gallium/winsys/svga/drm/vmw_screen_svga.c b/src/gallium/winsys/svga/drm/vmw_screen_svga.c
index bf817ca88d8..bfba2038559 100644
--- a/src/gallium/winsys/svga/drm/vmw_screen_svga.c
+++ b/src/gallium/winsys/svga/drm/vmw_screen_svga.c
@@ -47,7 +47,14 @@
 #include "vmw_surface.h"
 #include "vmw_buffer.h"
 #include "vmw_fence.h"
+#include "vmw_shader.h"
+#include "svga3d_surfacedefs.h"
 
+/**
+ * Try to get a surface backing buffer from the cache
+ * if it's this size or smaller.
+ */
+#define VMW_TRY_CACHED_SIZE (2*1024*1024)
 
 static struct svga_winsys_buffer *
 vmw_svga_winsys_buffer_create(struct svga_winsys_screen *sws,
@@ -56,64 +63,37 @@ vmw_svga_winsys_buffer_create(struct svga_winsys_screen *sws,
                               unsigned size)
 {
    struct vmw_winsys_screen *vws = vmw_winsys_screen(sws);
-   struct pb_desc desc;
+   struct vmw_buffer_desc desc;
    struct pb_manager *provider;
    struct pb_buffer *buffer;
 
    memset(&desc, 0, sizeof desc);
-   desc.alignment = alignment;
-   desc.usage = usage;
+   desc.pb_desc.alignment = alignment;
+   desc.pb_desc.usage = usage;
 
    if (usage == SVGA_BUFFER_USAGE_PINNED) {
       if (vws->pools.query_fenced == NULL && !vmw_query_pools_init(vws))
 	 return NULL;
       provider = vws->pools.query_fenced;
+   } else if (usage == SVGA_BUFFER_USAGE_SHADER) {
+      provider = vws->pools.mob_shader_slab_fenced;
    } else
       provider = vws->pools.gmr_fenced;
 
    assert(provider);
-   buffer = provider->create_buffer(provider, size, &desc);
+   buffer = provider->create_buffer(provider, size, &desc.pb_desc);
 
    if(!buffer && provider == vws->pools.gmr_fenced) {
 
       assert(provider);
       provider = vws->pools.gmr_slab_fenced;
-      buffer = provider->create_buffer(provider, size, &desc);
+      buffer = provider->create_buffer(provider, size, &desc.pb_desc);
    }
 
    if (!buffer)
       return NULL;
 
-   return vmw_svga_winsys_buffer(buffer);
-}
-
-
-static void *
-vmw_svga_winsys_buffer_map(struct svga_winsys_screen *sws,
-                           struct svga_winsys_buffer *buf,
-                           unsigned flags)
-{
-   (void)sws;
-   return pb_map(vmw_pb_buffer(buf), flags, NULL);
-}
-
-
-static void
-vmw_svga_winsys_buffer_unmap(struct svga_winsys_screen *sws,
-                             struct svga_winsys_buffer *buf)
-{
-   (void)sws;
-   pb_unmap(vmw_pb_buffer(buf));
-}
-
-
-static void
-vmw_svga_winsys_buffer_destroy(struct svga_winsys_screen *sws,
-                               struct svga_winsys_buffer *buf)
-{
-   struct pb_buffer *pbuf = vmw_pb_buffer(buf);
-   (void)sws;
-   pb_reference(&pbuf, NULL);
+   return vmw_svga_winsys_buffer_wrap(buffer);
 }
 
 
@@ -161,7 +141,12 @@ vmw_svga_winsys_surface_create(struct svga_winsys_screen *sws,
 {
    struct vmw_winsys_screen *vws = vmw_winsys_screen(sws);
    struct vmw_svga_winsys_surface *surface;
+   struct vmw_buffer_desc desc;
+   struct pb_manager *provider = vws->pools.mob_fenced;
+   uint32_t buffer_size;
+
 
+   memset(&desc, 0, sizeof(desc));
    surface = CALLOC_STRUCT(vmw_svga_winsys_surface);
    if(!surface)
       goto no_surface;
@@ -169,15 +154,96 @@ vmw_svga_winsys_surface_create(struct svga_winsys_screen *sws,
    pipe_reference_init(&surface->refcnt, 1);
    p_atomic_set(&surface->validated, 0);
    surface->screen = vws;
-   surface->sid = vmw_ioctl_surface_create(vws,
-                                           flags, format, size,
-                                           numFaces, numMipLevels);
-   if(surface->sid == SVGA3D_INVALID_ID)
-      goto no_sid;
+   pipe_mutex_init(surface->mutex);
+
+   /*
+    * Used for the backing buffer GB surfaces, and to approximate
+    * when to flush on non-GB hosts.
+    */
+   buffer_size = svga3dsurface_get_serialized_size(format, size, numMipLevels, (numFaces == 6));
+   if (sws->have_gb_objects) {
+      SVGAGuestPtr ptr = {0,0};
+
+      /*
+       * If the backing buffer size is small enough, try to allocate a
+       * buffer out of the buffer cache. Otherwise, let the kernel allocate
+       * a suitable buffer for us.
+       */
+      if (buffer_size < VMW_TRY_CACHED_SIZE) {
+         struct pb_buffer *pb_buf;
+
+         surface->size = buffer_size;
+         desc.pb_desc.alignment = 4096;
+         desc.pb_desc.usage = 0;
+         pb_buf = provider->create_buffer(provider, buffer_size, &desc.pb_desc);
+         surface->buf = vmw_svga_winsys_buffer_wrap(pb_buf);
+         if (surface->buf && !vmw_gmr_bufmgr_region_ptr(pb_buf, &ptr))
+            assert(0);
+      }
+
+      surface->sid = vmw_ioctl_gb_surface_create(vws,
+						 flags, format, size,
+						 numFaces, numMipLevels,
+                                                 ptr.gmrId,
+                                                 surface->buf ? NULL :
+						 &desc.region);
+
+      if (surface->sid == SVGA3D_INVALID_ID && surface->buf) {
+
+         /*
+          * Kernel refused to allocate a surface for us.
+          * Perhaps something was wrong with our buffer?
+          * This is really a guard against future new size requirements
+          * on the backing buffers.
+          */
+         vmw_svga_winsys_buffer_destroy(sws, surface->buf);
+         surface->buf = NULL;
+         surface->sid = vmw_ioctl_gb_surface_create(vws,
+                                                    flags, format, size,
+                                                    numFaces, numMipLevels,
+                                                    0,
+                                                    &desc.region);
+         if (surface->sid == SVGA3D_INVALID_ID)
+            goto no_sid;
+      }
+
+      /*
+       * If the kernel created the buffer for us, wrap it into a
+       * vmw_svga_winsys_buffer.
+       */
+      if (surface->buf == NULL) {
+         struct pb_buffer *pb_buf;
+
+         surface->size = vmw_region_size(desc.region);
+         desc.pb_desc.alignment = 4096;
+         desc.pb_desc.usage = VMW_BUFFER_USAGE_SHARED;
+         pb_buf = provider->create_buffer(provider, surface->size,
+                                          &desc.pb_desc);
+         surface->buf = vmw_svga_winsys_buffer_wrap(pb_buf);
+         if (surface->buf == NULL) {
+            vmw_ioctl_region_destroy(desc.region);
+            vmw_ioctl_surface_destroy(vws, surface->sid);
+            goto no_sid;
+         }
+      }
+   } else {
+      surface->sid = vmw_ioctl_surface_create(vws,
+                                              flags, format, size,
+                                              numFaces, numMipLevels);
+      if(surface->sid == SVGA3D_INVALID_ID)
+         goto no_sid;
+
+      /* Best estimate for surface size, used for early flushing. */
+      surface->size = buffer_size;
+      surface->buf = NULL; 
+   }      
 
    return svga_winsys_surface(surface);
 
 no_sid:
+   if (surface->buf)
+      vmw_svga_winsys_buffer_destroy(sws, surface->buf);
+
    FREE(surface);
 no_surface:
    return NULL;
@@ -220,6 +286,9 @@ vmw_svga_winsys_get_hw_version(struct svga_winsys_screen *sws)
 {
    struct vmw_winsys_screen *vws = vmw_winsys_screen(sws);
 
+   if (sws->have_gb_objects)
+      return SVGA3D_HWVERSION_WS8_B1;
+
    return (SVGA3dHardwareVersion) vws->ioctl.hwversion;
 }
 
@@ -228,70 +297,70 @@ static boolean
 vmw_svga_winsys_get_cap(struct svga_winsys_screen *sws,
                         SVGA3dDevCapIndex index,
                         SVGA3dDevCapResult *result)
-{
+{   
    struct vmw_winsys_screen *vws = vmw_winsys_screen(sws);
-   const uint32 *capsBlock;
-   const SVGA3dCapsRecord *capsRecord = NULL;
-   uint32 offset;
-   const SVGA3dCapPair *capArray;
-   int numCaps, first, last;
-
-   if(vws->ioctl.hwversion < SVGA3D_HWVERSION_WS6_B1)
-      return FALSE;
-
-   /*
-    * Search linearly through the caps block records for the specified type.
-    */
-   capsBlock = (const uint32 *)vws->ioctl.buffer;
-   for (offset = 0; capsBlock[offset] != 0; offset += capsBlock[offset]) {
-      const SVGA3dCapsRecord *record;
-      assert(offset < SVGA_FIFO_3D_CAPS_SIZE);
-      record = (const SVGA3dCapsRecord *) (capsBlock + offset);
-      if ((record->header.type >= SVGA3DCAPS_RECORD_DEVCAPS_MIN) &&
-          (record->header.type <= SVGA3DCAPS_RECORD_DEVCAPS_MAX) &&
-          (!capsRecord || (record->header.type > capsRecord->header.type))) {
-         capsRecord = record;
-      }
-   }
 
-   if(!capsRecord)
+   if (index > vws->ioctl.num_cap_3d || !vws->ioctl.cap_3d[index].has_cap)      
       return FALSE;
 
-   /*
-    * Calculate the number of caps from the size of the record.
-    */
-   capArray = (const SVGA3dCapPair *) capsRecord->data;
-   numCaps = (int) ((capsRecord->header.length * sizeof(uint32) -
-                     sizeof capsRecord->header) / (2 * sizeof(uint32)));
-
-   /*
-    * Binary-search for the cap with the specified index.
-    */
-   for (first = 0, last = numCaps - 1; first <= last; ) {
-      int mid = (first + last) / 2;
+   *result = vws->ioctl.cap_3d[index].result;
+   return TRUE;
+}
 
-      if ((SVGA3dDevCapIndex) capArray[mid][0] == index) {
-         /*
-          * Found it.
-          */
-         result->u = capArray[mid][1];
-         return TRUE;
-      }
+static struct svga_winsys_gb_shader *
+vmw_svga_winsys_shader_create(struct svga_winsys_screen *sws,
+			      SVGA3dShaderType type,
+			      const uint32 *bytecode,
+			      uint32 bytecodeLen)
+{
+   struct vmw_winsys_screen *vws = vmw_winsys_screen(sws);
+   struct vmw_svga_winsys_shader *shader;
+   void *code;
+
+   shader = CALLOC_STRUCT(vmw_svga_winsys_shader);
+   if(!shader)
+      goto out_no_shader;
+
+   pipe_reference_init(&shader->refcnt, 1);
+   p_atomic_set(&shader->validated, 0);
+   shader->screen = vws;
+   shader->buf = vmw_svga_winsys_buffer_create(sws, 64,
+					       SVGA_BUFFER_USAGE_SHADER,
+					       bytecodeLen);
+   if (!shader->buf)
+      goto out_no_buf;
+
+   code = vmw_svga_winsys_buffer_map(sws, shader->buf, PIPE_TRANSFER_WRITE);
+   if (!code)
+      goto out_no_buf;
+
+   memcpy(code, bytecode, bytecodeLen);
+   vmw_svga_winsys_buffer_unmap(sws, shader->buf);
+
+   shader->shid = vmw_ioctl_shader_create(vws, type, bytecodeLen);
+   if(shader->shid == SVGA3D_INVALID_ID)
+      goto out_no_shid;
+
+   return svga_winsys_shader(shader);
+
+out_no_shid:
+   vmw_svga_winsys_buffer_destroy(sws, shader->buf);
+out_no_buf:
+   FREE(shader);
+out_no_shader:
+   return NULL;
+}
 
-      /*
-       * Divide and conquer.
-       */
-      if ((SVGA3dDevCapIndex) capArray[mid][0] > index) {
-         last = mid - 1;
-      } else {
-         first = mid + 1;
-      }
-   }
+static void
+vmw_svga_winsys_shader_destroy(struct svga_winsys_screen *sws,
+			       struct svga_winsys_gb_shader *shader)
+{
+   struct vmw_svga_winsys_shader *d_shader =
+      vmw_svga_winsys_shader(shader);
 
-   return FALSE;
+   vmw_svga_winsys_shader_reference(&d_shader, NULL);
 }
 
-
 boolean
 vmw_winsys_screen_init_svga(struct vmw_winsys_screen *vws)
 {
@@ -308,6 +377,8 @@ vmw_winsys_screen_init_svga(struct vmw_winsys_screen *vws)
    vws->base.buffer_destroy = vmw_svga_winsys_buffer_destroy;
    vws->base.fence_reference = vmw_svga_winsys_fence_reference;
    vws->base.fence_signalled = vmw_svga_winsys_fence_signalled;
+   vws->base.shader_create = vmw_svga_winsys_shader_create;
+   vws->base.shader_destroy = vmw_svga_winsys_shader_destroy;
    vws->base.fence_finish = vmw_svga_winsys_fence_finish;
 
    return TRUE;
diff --git a/src/gallium/winsys/svga/drm/vmw_shader.c b/src/gallium/winsys/svga/drm/vmw_shader.c
new file mode 100644
index 00000000000..e82486ab608
--- /dev/null
+++ b/src/gallium/winsys/svga/drm/vmw_shader.c
@@ -0,0 +1,64 @@
+/**********************************************************
+ * Copyright 2009-2012 VMware, Inc.  All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person
+ * obtaining a copy of this software and associated documentation
+ * files (the "Software"), to deal in the Software without
+ * restriction, including without limitation the rights to use, copy,
+ * modify, merge, publish, distribute, sublicense, and/or sell copies
+ * of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be
+ * included in all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ **********************************************************/
+
+
+#include "svga_cmd.h"
+#include "util/u_debug.h"
+#include "util/u_memory.h"
+
+#include "vmw_shader.h"
+#include "vmw_screen.h"
+
+void
+vmw_svga_winsys_shader_reference(struct vmw_svga_winsys_shader **pdst,
+                                 struct vmw_svga_winsys_shader *src)
+{
+   struct pipe_reference *src_ref;
+   struct pipe_reference *dst_ref;
+   struct vmw_svga_winsys_shader *dst;
+
+   if(pdst == NULL || *pdst == src)
+      return;
+
+   dst = *pdst;
+
+   src_ref = src ? &src->refcnt : NULL;
+   dst_ref = dst ? &dst->refcnt : NULL;
+
+   if (pipe_reference(dst_ref, src_ref)) {
+      struct svga_winsys_screen *sws = &dst->screen->base;
+
+      vmw_ioctl_shader_destroy(dst->screen, dst->shid);
+#ifdef DEBUG
+      /* to detect dangling pointers */
+      assert(p_atomic_read(&dst->validated) == 0);
+      dst->shid = SVGA3D_INVALID_ID;
+#endif
+      sws->buffer_destroy(sws, dst->buf);
+      FREE(dst);
+   }
+
+   *pdst = src;
+}
diff --git a/src/gallium/winsys/svga/drm/vmw_shader.h b/src/gallium/winsys/svga/drm/vmw_shader.h
new file mode 100644
index 00000000000..1fd8c3311f9
--- /dev/null
+++ b/src/gallium/winsys/svga/drm/vmw_shader.h
@@ -0,0 +1,67 @@
+/**********************************************************
+ * Copyright 2009-2012 VMware, Inc.  All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person
+ * obtaining a copy of this software and associated documentation
+ * files (the "Software"), to deal in the Software without
+ * restriction, including without limitation the rights to use, copy,
+ * modify, merge, publish, distribute, sublicense, and/or sell copies
+ * of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be
+ * included in all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ **********************************************************/
+
+/**
+ * @file
+ * Shaders for VMware SVGA winsys.
+ *
+ * @author Jose Fonseca <jfonseca@vmware.com>
+ * @author Thomas Hellstrom <thellstrom@vmware.com>
+ */
+
+#ifndef VMW_SHADER_H_
+
+#include "pipe/p_compiler.h"
+#include "util/u_atomic.h"
+#include "util/u_inlines.h"
+
+struct vmw_svga_winsys_shader
+{
+   int32_t validated;
+   struct pipe_reference refcnt;
+
+   struct vmw_winsys_screen *screen;
+   struct svga_winsys_buffer *buf;
+   uint32_t shid;
+};
+
+static INLINE struct svga_winsys_gb_shader *
+svga_winsys_shader(struct vmw_svga_winsys_shader *shader)
+{
+   assert(!shader || shader->shid != SVGA3D_INVALID_ID);
+   return (struct svga_winsys_gb_shader *)shader;
+}
+
+static INLINE struct vmw_svga_winsys_shader *
+vmw_svga_winsys_shader(struct svga_winsys_gb_shader *shader)
+{
+   return (struct vmw_svga_winsys_shader *)shader;
+}
+
+void
+vmw_svga_winsys_shader_reference(struct vmw_svga_winsys_shader **pdst,
+                                  struct vmw_svga_winsys_shader *src);
+
+#endif /* VMW_SHADER_H_ */
diff --git a/src/gallium/winsys/svga/drm/vmw_surface.c b/src/gallium/winsys/svga/drm/vmw_surface.c
index 5f1b9ad5770..cf648b4dd93 100644
--- a/src/gallium/winsys/svga/drm/vmw_surface.c
+++ b/src/gallium/winsys/svga/drm/vmw_surface.c
@@ -27,9 +27,152 @@
 #include "svga_cmd.h"
 #include "util/u_debug.h"
 #include "util/u_memory.h"
-
+#include "pipe/p_defines.h"
 #include "vmw_surface.h"
 #include "vmw_screen.h"
+#include "vmw_buffer.h"
+#include "vmw_context.h"
+#include "pipebuffer/pb_bufmgr.h"
+
+
+void *
+vmw_svga_winsys_surface_map(struct svga_winsys_context *swc,
+                            struct svga_winsys_surface *srf,
+                            unsigned flags, boolean *retry)
+{
+   struct vmw_svga_winsys_surface *vsrf = vmw_svga_winsys_surface(srf);
+   void *data = NULL;
+   struct pb_buffer *pb_buf;
+   uint32_t pb_flags;
+   struct vmw_winsys_screen *vws = vsrf->screen;
+   
+   *retry = FALSE;
+   assert((flags & (PIPE_TRANSFER_READ | PIPE_TRANSFER_WRITE)) != 0);
+   pipe_mutex_lock(vsrf->mutex);
+
+   if (vsrf->mapcount) {
+      /*
+       * Only allow multiple readers to map.
+       */
+      if ((flags & PIPE_TRANSFER_WRITE) ||
+          (vsrf->map_mode & PIPE_TRANSFER_WRITE))
+         goto out_unlock;
+      
+      data = vsrf->data;
+      goto out_mapped;
+   }
+
+   vsrf->rebind = FALSE;
+
+   /*
+    * If we intend to read, there's no point discarding the
+    * data if busy.
+    */
+   if (flags & PIPE_TRANSFER_READ || vsrf->shared)
+      flags &= ~PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE;
+
+   /*
+    * Discard is a hint to a synchronized map.
+    */
+   if (flags & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE)
+      flags &= ~PIPE_TRANSFER_UNSYNCHRONIZED;
+
+   /*
+    * The surface is allowed to be referenced on the command stream iff
+    * we're mapping unsynchronized or discard. This is an early check.
+    * We need to recheck after a failing discard map.
+    */
+   if (!(flags & (PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE |
+                  PIPE_TRANSFER_UNSYNCHRONIZED)) &&
+       p_atomic_read(&vsrf->validated)) {
+      *retry = TRUE;
+      goto out_unlock;
+   }
+
+   pb_flags = flags & (PIPE_TRANSFER_READ_WRITE | PIPE_TRANSFER_UNSYNCHRONIZED);
+
+   if (flags & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
+      struct pb_manager *provider;
+      struct pb_desc desc;
+
+      /*
+       * First, if possible, try to map existing storage with DONTBLOCK.
+       */
+      if (!p_atomic_read(&vsrf->validated)) {
+         data = vmw_svga_winsys_buffer_map(&vws->base, vsrf->buf,
+                                           PIPE_TRANSFER_DONTBLOCK | pb_flags);
+         if (data)
+            goto out_mapped;
+      } 
+
+      /*
+       * Attempt to get a new buffer.
+       */
+      provider = vws->pools.mob_fenced;
+      memset(&desc, 0, sizeof(desc));
+      desc.alignment = 4096;
+      pb_buf = provider->create_buffer(provider, vsrf->size, &desc);
+      if (pb_buf != NULL) {
+         struct svga_winsys_buffer *vbuf =
+            vmw_svga_winsys_buffer_wrap(pb_buf);
+
+         data = vmw_svga_winsys_buffer_map(&vws->base, vbuf, pb_flags);
+         if (data) {
+            vsrf->rebind = TRUE;
+            /*
+             * We've discarded data on this surface and thus
+             * it's data is no longer consider referenced.
+             */
+            vmw_swc_surface_clear_reference(swc, vsrf);
+            if (vsrf->buf)
+               vmw_svga_winsys_buffer_destroy(&vws->base, vsrf->buf);
+            vsrf->buf = vbuf;
+            goto out_mapped;
+         } else
+            vmw_svga_winsys_buffer_destroy(&vws->base, vbuf);
+      }
+      /*
+       * We couldn't get and map a new buffer for some reason.
+       * Fall through to an ordinary map.
+       * But tell pipe driver to flush now if already on validate list,
+       * Otherwise we'll overwrite previous contents.
+       */
+      if (!(flags & PIPE_TRANSFER_UNSYNCHRONIZED) && 
+          p_atomic_read(&vsrf->validated)) {
+         *retry = TRUE;
+         goto out_unlock;
+      }
+   }
+
+   pb_flags |= (flags & PIPE_TRANSFER_DONTBLOCK);
+   data = vmw_svga_winsys_buffer_map(&vws->base, vsrf->buf, pb_flags);
+   if (data == NULL)
+      goto out_unlock;
+
+out_mapped:
+   ++vsrf->mapcount;
+   vsrf->data = data;
+   vsrf->map_mode = flags & (PIPE_TRANSFER_READ | PIPE_TRANSFER_WRITE);
+out_unlock:
+   pipe_mutex_unlock(vsrf->mutex);
+   return data;
+}
+
+
+void
+vmw_svga_winsys_surface_unmap(struct svga_winsys_context *swc,
+                              struct svga_winsys_surface *srf,
+                              boolean *rebind)
+{
+   struct vmw_svga_winsys_surface *vsrf = vmw_svga_winsys_surface(srf);
+   pipe_mutex_lock(vsrf->mutex);
+   if (--vsrf->mapcount == 0) {
+      *rebind = vsrf->rebind;
+      vsrf->rebind = FALSE;
+      vmw_svga_winsys_buffer_unmap(&vsrf->screen->base, vsrf->buf);
+   }
+   pipe_mutex_unlock(vsrf->mutex);
+}
 
 void
 vmw_svga_winsys_surface_reference(struct vmw_svga_winsys_surface **pdst,
@@ -48,12 +191,15 @@ vmw_svga_winsys_surface_reference(struct vmw_svga_winsys_surface **pdst,
    dst_ref = dst ? &dst->refcnt : NULL;
 
    if (pipe_reference(dst_ref, src_ref)) {
+      if (dst->buf)
+         vmw_svga_winsys_buffer_destroy(&dst->screen->base, dst->buf);
       vmw_ioctl_surface_destroy(dst->screen, dst->sid);
 #ifdef DEBUG
       /* to detect dangling pointers */
       assert(p_atomic_read(&dst->validated) == 0);
       dst->sid = SVGA3D_INVALID_ID;
 #endif
+      pipe_mutex_destroy(dst->mutex);
       FREE(dst);
    }
 
diff --git a/src/gallium/winsys/svga/drm/vmw_surface.h b/src/gallium/winsys/svga/drm/vmw_surface.h
index 3d61595c288..e44d0554fbc 100644
--- a/src/gallium/winsys/svga/drm/vmw_surface.h
+++ b/src/gallium/winsys/svga/drm/vmw_surface.h
@@ -38,6 +38,8 @@
 #include "pipe/p_compiler.h"
 #include "util/u_atomic.h"
 #include "util/u_inlines.h"
+#include "os/os_thread.h"
+#include "pipebuffer/pb_buffer.h"
 
 #define VMW_MAX_PRESENTS 3
 
@@ -54,6 +56,15 @@ struct vmw_svga_winsys_surface
    /* FIXME: make this thread safe */
    unsigned next_present_no;
    uint32_t present_fences[VMW_MAX_PRESENTS];
+
+   pipe_mutex mutex;
+   struct svga_winsys_buffer *buf; /* Current backing guest buffer */
+   uint32_t mapcount; /* Number of mappers */
+   uint32_t map_mode; /* PIPE_TRANSFER_[READ|WRITE] */
+   void *data; /* Pointer to data if mapcount != 0*/
+   boolean shared; /* Shared surface. Never discard */
+   uint32_t size; /* Size of backing buffer */
+   boolean rebind; /* Surface needs a rebind after next unmap */
 };
 
 
@@ -75,5 +86,13 @@ vmw_svga_winsys_surface(struct svga_winsys_surface *surf)
 void
 vmw_svga_winsys_surface_reference(struct vmw_svga_winsys_surface **pdst,
                                   struct vmw_svga_winsys_surface *src);
+void *
+vmw_svga_winsys_surface_map(struct svga_winsys_context *swc,
+			    struct svga_winsys_surface *srf,
+			    unsigned flags, boolean *retry);
+void
+vmw_svga_winsys_surface_unmap(struct svga_winsys_context *swc,
+                              struct svga_winsys_surface *srf,
+                              boolean *rebind);
 
 #endif /* VMW_SURFACE_H_ */
diff --git a/src/gallium/winsys/svga/drm/vmwgfx_drm.h b/src/gallium/winsys/svga/drm/vmwgfx_drm.h
index bad47a702f2..e42b3f5bf1d 100644
--- a/src/gallium/winsys/svga/drm/vmwgfx_drm.h
+++ b/src/gallium/winsys/svga/drm/vmwgfx_drm.h
@@ -28,6 +28,10 @@
 #ifndef __VMWGFX_DRM_H__
 #define __VMWGFX_DRM_H__
 
+#ifndef __KERNEL__
+#include <drm.h>
+#endif
+
 #define DRM_VMW_MAX_SURFACE_FACES 6
 #define DRM_VMW_MAX_MIP_LEVELS 24
 
@@ -54,7 +58,12 @@
 #define DRM_VMW_FENCE_EVENT          17
 #define DRM_VMW_PRESENT              18
 #define DRM_VMW_PRESENT_READBACK     19
-
+#define DRM_VMW_UPDATE_LAYOUT        20
+#define DRM_VMW_CREATE_SHADER        21
+#define DRM_VMW_UNREF_SHADER         22
+#define DRM_VMW_GB_SURFACE_CREATE    23
+#define DRM_VMW_GB_SURFACE_REF       24
+#define DRM_VMW_SYNCCPU              25
 
 /*************************************************************************/
 /**
@@ -75,6 +84,9 @@
 #define DRM_VMW_PARAM_FIFO_CAPS        4
 #define DRM_VMW_PARAM_MAX_FB_SIZE      5
 #define DRM_VMW_PARAM_FIFO_HW_VERSION  6
+#define DRM_VMW_PARAM_MAX_SURF_MEMORY  7
+#define DRM_VMW_PARAM_3D_CAPS_SIZE     8
+#define DRM_VMW_PARAM_MAX_MOB_MEMORY   9
 
 /**
  * struct drm_vmw_getparam_arg
@@ -661,6 +673,51 @@ struct drm_vmw_fence_arg {
 
 /*************************************************************************/
 /**
+ * DRM_VMW_FENCE_EVENT
+ *
+ * Queues an event on a fence to be delivered on the drm character device
+ * when the fence has signaled the DRM_VMW_FENCE_FLAG_EXEC flag.
+ * Optionally the approximate time when the fence signaled is
+ * given by the event.
+ */
+
+/*
+ * The event type
+ */
+#define DRM_VMW_EVENT_FENCE_SIGNALED 0x80000000
+
+struct drm_vmw_event_fence {
+	struct drm_event base;
+	uint64_t user_data;
+	uint32_t tv_sec;
+	uint32_t tv_usec;
+};
+
+/*
+ * Flags that may be given to the command.
+ */
+/* Request fence signaled time on the event. */
+#define DRM_VMW_FE_FLAG_REQ_TIME (1 << 0)
+
+/**
+ * struct drm_vmw_fence_event_arg
+ *
+ * @fence_rep: Pointer to fence_rep structure cast to uint64_t or 0 if
+ * the fence is not supposed to be referenced by user-space.
+ * @user_info: Info to be delivered with the event.
+ * @handle: Attach the event to this fence only.
+ * @flags: A set of flags as defined above.
+ */
+struct drm_vmw_fence_event_arg {
+	uint64_t fence_rep;
+	uint64_t user_data;
+	uint32_t handle;
+	uint32_t flags;
+};
+
+
+/*************************************************************************/
+/**
  * DRM_VMW_PRESENT
  *
  * Executes an SVGA present on a given fb for a given surface. The surface
@@ -720,5 +777,276 @@ struct drm_vmw_present_readback_arg {
 	 uint64_t fence_rep;
 };
 
+/*************************************************************************/
+/**
+ * DRM_VMW_UPDATE_LAYOUT - Update layout
+ *
+ * Updates the preferred modes and connection status for connectors. The
+ * command consists of one drm_vmw_update_layout_arg pointing to an array
+ * of num_outputs drm_vmw_rect's.
+ */
+
+/**
+ * struct drm_vmw_update_layout_arg
+ *
+ * @num_outputs: number of active connectors
+ * @rects: pointer to array of drm_vmw_rect cast to an uint64_t
+ *
+ * Input argument to the DRM_VMW_UPDATE_LAYOUT Ioctl.
+ */
+struct drm_vmw_update_layout_arg {
+	uint32_t num_outputs;
+	uint32_t pad64;
+	uint64_t rects;
+};
+
+
+/*************************************************************************/
+/**
+ * DRM_VMW_CREATE_SHADER - Create shader
+ *
+ * Creates a shader and optionally binds it to a dma buffer containing
+ * the shader byte-code.
+ */
+
+/**
+ * enum drm_vmw_shader_type - Shader types
+ */
+enum drm_vmw_shader_type {
+	drm_vmw_shader_type_vs = 0,
+	drm_vmw_shader_type_ps,
+	drm_vmw_shader_type_gs
+};
+
+
+/**
+ * struct drm_vmw_shader_create_arg
+ *
+ * @shader_type: Shader type of the shader to create.
+ * @size: Size of the byte-code in bytes.
+ * where the shader byte-code starts
+ * @buffer_handle: Buffer handle identifying the buffer containing the
+ * shader byte-code
+ * @shader_handle: On successful completion contains a handle that
+ * can be used to subsequently identify the shader.
+ * @offset: Offset in bytes into the buffer given by @buffer_handle,
+ *
+ * Input / Output argument to the DRM_VMW_CREATE_SHADER Ioctl.
+ */
+struct drm_vmw_shader_create_arg {
+	enum drm_vmw_shader_type shader_type;
+	uint32_t size;
+	uint32_t buffer_handle;
+	uint32_t shader_handle;
+	uint64_t offset;
+};
+
+/*************************************************************************/
+/**
+ * DRM_VMW_UNREF_SHADER - Unreferences a shader
+ *
+ * Destroys a user-space reference to a shader, optionally destroying
+ * it.
+ */
+
+/**
+ * struct drm_vmw_shader_arg
+ *
+ * @handle: Handle identifying the shader to destroy.
+ *
+ * Input argument to the DRM_VMW_UNREF_SHADER ioctl.
+ */
+struct drm_vmw_shader_arg {
+	uint32_t handle;
+	uint32_t pad64;
+};
+
+/*************************************************************************/
+/**
+ * DRM_VMW_GB_SURFACE_CREATE - Create a host guest-backed surface.
+ *
+ * Allocates a surface handle and queues a create surface command
+ * for the host on the first use of the surface. The surface ID can
+ * be used as the surface ID in commands referencing the surface.
+ */
+
+/**
+ * enum drm_vmw_surface_flags
+ *
+ * @drm_vmw_surface_flag_shareable:     Whether the surface is shareable
+ * @drm_vmw_surface_flag_scanout:       Whether the surface is a scanout
+ *                                      surface.
+ * @drm_vmw_surface_flag_create_buffer: Create a backup buffer if none is
+ *                                      given.
+ */
+enum drm_vmw_surface_flags {
+	drm_vmw_surface_flag_shareable = (1 << 0),
+	drm_vmw_surface_flag_scanout = (1 << 1),
+	drm_vmw_surface_flag_create_buffer = (1 << 2)
+};
+
+/**
+ * struct drm_vmw_gb_surface_create_req
+ *
+ * @svga3d_flags:     SVGA3d surface flags for the device.
+ * @format:           SVGA3d format.
+ * @mip_level:        Number of mip levels for all faces.
+ * @drm_surface_flags Flags as described above.
+ * @multisample_count Future use. Set to 0.
+ * @autogen_filter    Future use. Set to 0.
+ * @buffer_handle     Buffer handle of backup buffer. SVGA3D_INVALID_ID
+ *                    if none.
+ * @base_size         Size of the base mip level for all faces.
+ *
+ * Input argument to the  DRM_VMW_GB_SURFACE_CREATE Ioctl.
+ * Part of output argument for the DRM_VMW_GB_SURFACE_REF Ioctl.
+ */
+struct drm_vmw_gb_surface_create_req {
+	uint32_t svga3d_flags;
+	uint32_t format;
+	uint32_t mip_levels;
+	enum drm_vmw_surface_flags drm_surface_flags;
+	uint32_t multisample_count;
+	uint32_t autogen_filter;
+	uint32_t buffer_handle;
+	uint32_t pad64;
+	struct drm_vmw_size base_size;
+};
+
+/**
+ * struct drm_vmw_gb_surface_create_rep
+ *
+ * @handle:            Surface handle.
+ * @backup_size:       Size of backup buffers for this surface.
+ * @buffer_handle:     Handle of backup buffer. SVGA3D_INVALID_ID if none.
+ * @buffer_size:       Actual size of the buffer identified by
+ *                     @buffer_handle
+ * @buffer_map_handle: Offset into device address space for the buffer
+ *                     identified by @buffer_handle.
+ *
+ * Part of output argument for the DRM_VMW_GB_SURFACE_REF ioctl.
+ * Output argument for the DRM_VMW_GB_SURFACE_CREATE ioctl.
+ */
+struct drm_vmw_gb_surface_create_rep {
+	uint32_t handle;
+	uint32_t backup_size;
+	uint32_t buffer_handle;
+	uint32_t buffer_size;
+	uint64_t buffer_map_handle;
+};
+
+/**
+ * union drm_vmw_gb_surface_create_arg
+ *
+ * @req: Input argument as described above.
+ * @rep: Output argument as described above.
+ *
+ * Argument to the DRM_VMW_GB_SURFACE_CREATE ioctl.
+ */
+union drm_vmw_gb_surface_create_arg {
+	struct drm_vmw_gb_surface_create_rep rep;
+	struct drm_vmw_gb_surface_create_req req;
+};
+
+/*************************************************************************/
+/**
+ * DRM_VMW_GB_SURFACE_REF - Reference a host surface.
+ *
+ * Puts a reference on a host surface with a given handle, as previously
+ * returned by the DRM_VMW_GB_SURFACE_CREATE ioctl.
+ * A reference will make sure the surface isn't destroyed while we hold
+ * it and will allow the calling client to use the surface handle in
+ * the command stream.
+ *
+ * On successful return, the Ioctl returns the surface information given
+ * to and returned from the DRM_VMW_GB_SURFACE_CREATE ioctl.
+ */
+
+/**
+ * struct drm_vmw_gb_surface_reference_arg
+ *
+ * @creq: The data used as input when the surface was created, as described
+ *        above at "struct drm_vmw_gb_surface_create_req"
+ * @crep: Additional data output when the surface was created, as described
+ *        above at "struct drm_vmw_gb_surface_create_rep"
+ *
+ * Output Argument to the DRM_VMW_GB_SURFACE_REF ioctl.
+ */
+struct drm_vmw_gb_surface_ref_rep {
+	struct drm_vmw_gb_surface_create_req creq;
+	struct drm_vmw_gb_surface_create_rep crep;
+};
+
+/**
+ * union drm_vmw_gb_surface_reference_arg
+ *
+ * @req: Input data as described above at "struct drm_vmw_surface_arg"
+ * @rep: Output data as described above at "struct drm_vmw_gb_surface_ref_rep"
+ *
+ * Argument to the DRM_VMW_GB_SURFACE_REF Ioctl.
+ */
+union drm_vmw_gb_surface_reference_arg {
+	struct drm_vmw_gb_surface_ref_rep rep;
+	struct drm_vmw_surface_arg req;
+};
+
+
+/*************************************************************************/
+/**
+ * DRM_VMW_SYNCCPU - Sync a DMA buffer / MOB for CPU access.
+ *
+ * Idles any previously submitted GPU operations on the buffer and
+ * by default blocks command submissions that reference the buffer.
+ * If the file descriptor used to grab a blocking CPU sync is closed, the
+ * cpu sync is released.
+ * The flags argument indicates how the grab / release operation should be
+ * performed:
+ */
+
+/**
+ * enum drm_vmw_synccpu_flags - Synccpu flags:
+ *
+ * @drm_vmw_synccpu_read: Sync for read. If sync is done for read only, it's a
+ * hint to the kernel to allow command submissions that references the buffer
+ * for read-only.
+ * @drm_vmw_synccpu_write: Sync for write. Block all command submissions
+ * referencing this buffer.
+ * @drm_vmw_synccpu_dontblock: Dont wait for GPU idle, but rather return
+ * -EBUSY should the buffer be busy.
+ * @drm_vmw_synccpu_allow_cs: Allow command submission that touches the buffer
+ * while the buffer is synced for CPU. This is similar to the GEM bo idle
+ * behavior.
+ */
+enum drm_vmw_synccpu_flags {
+	drm_vmw_synccpu_read = (1 << 0),
+	drm_vmw_synccpu_write = (1 << 1),
+	drm_vmw_synccpu_dontblock = (1 << 2),
+	drm_vmw_synccpu_allow_cs = (1 << 3)
+};
+
+/**
+ * enum drm_vmw_synccpu_op - Synccpu operations:
+ *
+ * @drm_vmw_synccpu_grab:    Grab the buffer for CPU operations
+ * @drm_vmw_synccpu_release: Release a previous grab.
+ */
+enum drm_vmw_synccpu_op {
+	drm_vmw_synccpu_grab,
+	drm_vmw_synccpu_release
+};
+
+/**
+ * struct drm_vmw_synccpu_arg
+ *
+ * @op:			     The synccpu operation as described above.
+ * @handle:		     Handle identifying the buffer object.
+ * @flags:		     Flags as described above.
+ */
+struct drm_vmw_synccpu_arg {
+	enum drm_vmw_synccpu_op op;
+	enum drm_vmw_synccpu_flags flags;
+	uint32_t handle;
+	uint32_t pad64;
+};
 
 #endif