i915: Pulled in changes from i915tex

23 files changed, 3712 insertions, 726 deletions

diff --git a/src/gallium/winsys/dri/intel/Makefile b/src/gallium/winsys/dri/intel/Makefile
index 40654bb2ac3..f1409392484 100644
--- a/src/gallium/winsys/dri/intel/Makefile
+++ b/src/gallium/winsys/dri/intel/Makefile
@@ -19,11 +19,14 @@ DRIVER_SOURCES = \
 	intel_context.c \
 	intel_lock.c \
 	intel_screen.c \
-	intel_batchpool.c
+	ws_dri_bufmgr.c \
+	ws_dri_drmpool.c \
+	ws_dri_fencemgr.c \
+	ws_dri_mallocpool.c \
+	ws_dri_slabpool.c
 
 C_SOURCES = \
 	$(COMMON_GALLIUM_SOURCES) \
-	$(COMMON_BM_SOURCES) \
 	$(DRIVER_SOURCES)
 
 ASM_SOURCES = 
diff --git a/src/gallium/winsys/dri/intel/intel_batchbuffer.c b/src/gallium/winsys/dri/intel/intel_batchbuffer.c
index 5830b88b37e..05223bf700d 100644
--- a/src/gallium/winsys/dri/intel/intel_batchbuffer.c
+++ b/src/gallium/winsys/dri/intel/intel_batchbuffer.c
@@ -25,108 +25,95 @@
  * 
  **************************************************************************/
 
-#include <errno.h>
-#include <stdio.h>
 #include "intel_batchbuffer.h"
-#include "intel_context.h"
-#include "intel_screen.h"
-#include "intel_reg.h"
-#include "drm.h"
-
-/* Relocations in kernel space:
- *    - pass dma buffer seperately
- *    - memory manager knows how to patch
- *    - pass list of dependent buffers
- *    - pass relocation list
- *
- * Either:
- *    - get back an offset for buffer to fire
- *    - memory manager knows how to fire buffer
- *
- * Really want the buffer to be AGP and pinned.
- *
- */
-
-/* Cliprect fence: The highest fence protecting a dma buffer
- * containing explicit cliprect information.  Like the old drawable
- * lock but irq-driven.  X server must wait for this fence to expire
- * before changing cliprects [and then doing sw rendering?].  For
- * other dma buffers, the scheduler will grab current cliprect info
- * and mix into buffer.  X server must hold the lock while changing
- * cliprects???  Make per-drawable.  Need cliprects in shared memory
- * -- beats storing them with every cmd buffer in the queue.
- *
- * ==> X server must wait for this fence to expire before touching the
- * framebuffer with new cliprects.
- *
- * ==> Cliprect-dependent buffers associated with a
- * cliprect-timestamp.  All of the buffers associated with a timestamp
- * must go to hardware before any buffer with a newer timestamp.
- *
- * ==> Dma should be queued per-drawable for correct X/GL
- * synchronization.  Or can fences be used for this?
- *
- * Applies to: Blit operations, metaops, X server operations -- X
- * server automatically waits on its own dma to complete before
- * modifying cliprects ???
- */
+#include "intel_ioctl.h"
+#include <errno.h>
 
+#if 0
 static void
-intel_dump_batchbuffer(uint offset, uint * ptr, uint count)
+intel_dump_batchbuffer(GLuint offset, GLuint * ptr, GLuint count)
 {
    int i;
-   printf("\n\n\nSTART BATCH (%d dwords):\n", count / 4);
-   for (i = 0; i < count / 4; i += 1)
-      printf("\t0x%08x\n", ptr[i]);
-   printf("END BATCH\n\n\n");
+   fprintf(stderr, "\n\n\nSTART BATCH (%d dwords):\n", count / 4);
+   for (i = 0; i < count / 4; i += 4)
+      fprintf(stderr, "0x%x:\t0x%08x 0x%08x 0x%08x 0x%08x\n",
+              offset + i * 4, ptr[i], ptr[i + 1], ptr[i + 2], ptr[i + 3]);
+   fprintf(stderr, "END BATCH\n\n\n");
+}
+#endif
+
+static void 
+intel_realloc_relocs(struct intel_batchbuffer *batch, int num_relocs)
+{
+    unsigned long size = num_relocs * I915_RELOC0_STRIDE + I915_RELOC_HEADER;
+    
+    size *= sizeof(uint32_t);
+    batch->reloc = realloc(batch->reloc, size);
+    batch->reloc_size = num_relocs;
 }
 
 
 void
 intel_batchbuffer_reset(struct intel_batchbuffer *batch)
 {
-   int i;
-
-   if (batch->map) {
-      driBOUnmap(batch->buffer);
-      batch->map = NULL;
-   }
-
    /*
     * Get a new, free batchbuffer.
     */
-   batch->size =  BATCH_SZ;
-   driBOData(batch->buffer, batch->size, NULL, 0);
+    drmBO *bo;
+    struct drm_bo_info_req *req;
+    
+   driBOUnrefUserList(batch->list);
+   driBOResetList(batch->list);
 
-   driBOResetList(&batch->list);
+   batch->size = 4096; // ZZZ JB batch->intel->intelScreen->maxBatchSize;
+   driBOData(batch->buffer, batch->size, NULL, NULL, 0);
 
    /*
-    * Unreference buffers previously on the relocation list.
+    * Add the batchbuffer to the validate list.
     */
-   for (i = 0; i < batch->nr_relocs; i++) {
-      struct buffer_reloc *r = &batch->reloc[i];
-      driBOUnReference(r->buf);
-   }
 
-   batch->list_count = 0;
-   batch->nr_relocs = 0;
-   batch->flags = 0;
+   driBOAddListItem(batch->list, batch->buffer,
+		    DRM_BO_FLAG_EXE | DRM_BO_FLAG_MEM_TT,
+		    DRM_BO_FLAG_EXE | DRM_BO_MASK_MEM,
+		    &batch->dest_location, &batch->node);
+
+   req = &batch->node->bo_arg.d.req.bo_req;
 
    /*
-    * We don't refcount the batchbuffer itself since we can't destroy it
-    * while it's on the list.
+    * Set up information needed for us to make relocations
+    * relative to the underlying drm buffer objects.
     */
 
-   driBOAddListItem(&batch->list, batch->buffer,
-                    DRM_BO_FLAG_MEM_TT | DRM_BO_FLAG_EXE,
-                    DRM_BO_MASK_MEM | DRM_BO_FLAG_EXE);
+   driReadLockKernelBO();
+   bo = driBOKernel(batch->buffer);
+   req->presumed_offset = (uint64_t) bo->offset;
+   req->hint = DRM_BO_HINT_PRESUMED_OFFSET;
+   batch->drmBOVirtual = (uint8_t *) bo->virtual;
+   driReadUnlockKernelBO();
 
+   /*
+    * Adjust the relocation buffer size.
+    */
+
+   if (batch->reloc_size > INTEL_MAX_RELOCS ||
+       batch->reloc == NULL) 
+     intel_realloc_relocs(batch, INTEL_DEFAULT_RELOCS);
+   
+   assert(batch->reloc != NULL);
+   batch->reloc[0] = 0; /* No relocs yet. */
+   batch->reloc[1] = 1; /* Reloc type 1 */
+   batch->reloc[2] = 0; /* Only a single relocation list. */
+   batch->reloc[3] = 0; /* Only a single relocation list. */
 
    batch->map = driBOMap(batch->buffer, DRM_BO_FLAG_WRITE, 0);
+   batch->poolOffset = driBOPoolOffset(batch->buffer);
    batch->ptr = batch->map;
+   batch->dirty_state = ~0;
+   batch->nr_relocs = 0;
+   batch->flags = 0;
+   batch->id = 0;//batch->intel->intelScreen->batch_id++;
 }
 
-
 /*======================================================================
  * Public functions
  */
@@ -141,121 +128,253 @@ intel_batchbuffer_alloc(struct intel_context *intel)
                  &batch->buffer, 4096,
                  DRM_BO_FLAG_MEM_TT | DRM_BO_FLAG_EXE, 0);
    batch->last_fence = NULL;
-   driBOCreateList(20, &batch->list);
+   batch->list = driBOCreateList(20);
+   batch->reloc = NULL;
    intel_batchbuffer_reset(batch);
    return batch;
 }
 
-
 void
 intel_batchbuffer_free(struct intel_batchbuffer *batch)
 {
    if (batch->last_fence) {
       driFenceFinish(batch->last_fence,
-                     DRM_FENCE_TYPE_EXE | DRM_I915_FENCE_TYPE_RW,
-                     GL_FALSE);
-      driFenceUnReference(batch->last_fence);
-      batch->last_fence = NULL;
+		     DRM_FENCE_TYPE_EXE, GL_FALSE);
+      driFenceUnReference(&batch->last_fence);
    }
    if (batch->map) {
       driBOUnmap(batch->buffer);
       batch->map = NULL;
    }
    driBOUnReference(batch->buffer);
+   driBOFreeList(batch->list);
+   if (batch->reloc)
+       free(batch->reloc);
    batch->buffer = NULL;
    free(batch);
 }
 
+void
+intel_offset_relocation(struct intel_batchbuffer *batch,
+			unsigned pre_add,
+			struct _DriBufferObject *driBO,
+			uint64_t val_flags,
+			uint64_t val_mask)
+{
+    int itemLoc;
+    struct _drmBONode *node;
+    uint32_t *reloc;
+    struct drm_bo_info_req *req;
+    
+    driBOAddListItem(batch->list, driBO, val_flags, val_mask,
+		     &itemLoc, &node);
+    req = &node->bo_arg.d.req.bo_req;
+
+    if (!(req->hint &  DRM_BO_HINT_PRESUMED_OFFSET)) {
+
+	/*
+	 * Stop other threads from tampering with the underlying
+	 * drmBO while we're reading its offset.
+	 */
+
+	driReadLockKernelBO();
+	req->presumed_offset = (uint64_t) driBOKernel(driBO)->offset;
+	driReadUnlockKernelBO();
+	req->hint = DRM_BO_HINT_PRESUMED_OFFSET;
+    }
+    
+    pre_add += driBOPoolOffset(driBO);
+
+    if (batch->nr_relocs == batch->reloc_size)
+	intel_realloc_relocs(batch, batch->reloc_size * 2);
+
+    reloc = batch->reloc + 
+	(I915_RELOC_HEADER + batch->nr_relocs * I915_RELOC0_STRIDE);
+
+    reloc[0] = ((uint8_t *)batch->ptr - batch->drmBOVirtual);
+    intel_batchbuffer_emit_dword(batch, req->presumed_offset + pre_add);
+    reloc[1] = pre_add;
+    reloc[2] = itemLoc;
+    reloc[3] = batch->dest_location;
+    batch->nr_relocs++;
+}
 
 static void
-intel_batch_ioctl(struct intel_context *intel,
-                  uint start_offset, uint used, boolean allow_unlock)
+i915_drm_copy_reply(const struct drm_bo_info_rep * rep, drmBO * buf)
 {
-   drmI830BatchBuffer batch;
-
-   batch.start = start_offset;
-   batch.used = used;
-   batch.cliprects = NULL; /* unused */
-   batch.num_cliprects = 0;
-   batch.DR1 = 0;
-   batch.DR4 = 0; /* still need this ? */
-
-   DBG(IOCTL, "%s: 0x%x..0x%x DR4: %x cliprects: %d\n",
-       __FUNCTION__,
-       batch.start,
-       batch.start + batch.used * 4, batch.DR4, batch.num_cliprects);
-
-   if (drmCommandWrite(intel->driFd, DRM_I830_BATCHBUFFER, &batch,
-                       sizeof(batch))) {
-      printf("DRM_I830_BATCHBUFFER: %d\n", -errno);
-      UNLOCK_HARDWARE(intel);
-      exit(1);
-   }
+    buf->handle = rep->handle;
+    buf->flags = rep->flags;
+    buf->size = rep->size;
+    buf->offset = rep->offset;
+    buf->mapHandle = rep->arg_handle;
+    buf->proposedFlags = rep->proposed_flags;
+    buf->start = rep->buffer_start;
+    buf->fenceFlags = rep->fence_flags;
+    buf->replyFlags = rep->rep_flags;
+    buf->pageAlignment = rep->page_alignment;
 }
 
+static int 
+i915_execbuf(struct intel_batchbuffer *batch,
+	     GLuint used,
+	     GLboolean ignore_cliprects,
+	     drmBOList *list,
+	     struct drm_i915_execbuffer *ea)
+{
+   struct intel_context *intel = batch->intel;
+   drmBONode *node;
+   drmMMListHead *l;
+   struct drm_i915_op_arg *arg, *first;
+   struct drm_bo_op_req *req;
+   struct drm_bo_info_rep *rep;
+   uint64_t *prevNext = NULL;
+   drmBO *buf;
+   int ret = 0;
+   uint32_t count = 0;
+
+   first = NULL;
+   for (l = list->list.next; l != &list->list; l = l->next) {
+      node = DRMLISTENTRY(drmBONode, l, head);
+      
+      arg = &node->bo_arg;
+      req = &arg->d.req;
+      
+      if (!first)
+	 first = arg;
+      
+      if (prevNext)
+	 *prevNext = (unsigned long)arg;
+      
+      prevNext = &arg->next;
+      req->bo_req.handle = node->buf->handle;
+      req->op = drm_bo_validate;
+      req->bo_req.flags = node->arg0;
+      req->bo_req.mask = node->arg1;
+      req->bo_req.hint |= 0;
+      count++;
+   }
+
+   memset(ea, 0, sizeof(*ea));
+   ea->num_buffers = count;
+   ea->batch.start = batch->poolOffset;
+   ea->batch.used = used;
+#if 0 /* ZZZ JB: no cliprects used */
+   ea->batch.cliprects = intel->pClipRects;
+   ea->batch.num_cliprects = ignore_cliprects ? 0 : intel->numClipRects;
+   ea->batch.DR1 = 0;
+   ea->batch.DR4 = 0;((((GLuint) intel->drawX) & 0xffff) |
+		   (((GLuint) intel->drawY) << 16));
+#else
+   ea->batch.cliprects = NULL;
+   ea->batch.num_cliprects = 0;
+   ea->batch.DR1 = 0;
+   ea->batch.DR4 = 0;
+#endif
+   ea->fence_arg.flags = DRM_I915_FENCE_FLAG_FLUSHED;
+   ea->ops_list = (unsigned long) first;
+   first->reloc_ptr = (unsigned long) batch->reloc;
+   batch->reloc[0] = batch->nr_relocs;
+
+   //return -EFAULT;
+   do {
+      ret = drmCommandWriteRead(intel->driFd, DRM_I915_EXECBUFFER, ea,
+				sizeof(*ea));
+   } while (ret == -EAGAIN);
+
+   if (ret != 0)
+      return ret;
+
+   for (l = list->list.next; l != &list->list; l = l->next) {
+      node = DRMLISTENTRY(drmBONode, l, head);
+      arg = &node->bo_arg;
+      rep = &arg->d.rep.bo_info;
+
+      if (!arg->handled) {
+	 return -EFAULT;
+      }
+      if (arg->d.rep.ret)
+	 return arg->d.rep.ret;
+
+      buf = node->buf;
+      i915_drm_copy_reply(rep, buf);
+   }
+   return 0;
+}
 
 /* TODO: Push this whole function into bufmgr.
  */
-static void
+static struct _DriFenceObject *
 do_flush_locked(struct intel_batchbuffer *batch,
-                uint used, boolean allow_unlock)
+                GLuint used,
+                GLboolean ignore_cliprects, GLboolean allow_unlock)
 {
-   uint *ptr;
-   uint i, fenceFlags;
+   struct intel_context *intel = batch->intel;
    struct _DriFenceObject *fo;
+   drmFence fence;
+   drmBOList *boList;
+   struct drm_i915_execbuffer ea;
+   int ret = 0;
+
+   driBOValidateUserList(batch->list);
+   boList = driGetdrmBOList(batch->list);
+
+#if 0 /* ZZZ JB Allways run */
+   if (!(intel->numClipRects == 0 && !ignore_cliprects)) {
+#else
+   if (1) {
+#endif
+      ret = i915_execbuf(batch, used, ignore_cliprects, boList, &ea);
+   } else {
+     driPutdrmBOList(batch->list);
+     fo = NULL;
+     goto out;
+   }
+   driPutdrmBOList(batch->list);
+   if (ret)
+      abort();
 
-   driBOValidateList(batch->intel->driFd, &batch->list);
-
-   /* Apply the relocations.  This nasty map indicates to me that the
-    * whole task should be done internally by the memory manager, and
-    * that dma buffers probably need to be pinned within agp space.
-    */
-   ptr = (uint *) driBOMap(batch->buffer, DRM_BO_FLAG_WRITE,
-                             DRM_BO_HINT_ALLOW_UNFENCED_MAP);
+   if (ea.fence_arg.error != 0) {
 
-   for (i = 0; i < batch->nr_relocs; i++) {
-      struct buffer_reloc *r = &batch->reloc[i];
+     /*
+      * The hardware has been idled by the kernel.
+      * Don't fence the driBOs.
+      */
 
-      ptr[r->offset / 4] = driBOOffset(r->buf) + r->delta;
+       if (batch->last_fence)
+	   driFenceUnReference(&batch->last_fence);
+#if 0 /* ZZZ JB: no _mesa_* funcs in gallium */
+       _mesa_printf("fence error\n");
+#endif
+       batch->last_fence = NULL;
+       fo = NULL;
+       goto out;
    }
 
-   if (0)
-      intel_dump_batchbuffer(0, ptr, used);
+   fence.handle = ea.fence_arg.handle;
+   fence.fence_class = ea.fence_arg.fence_class;
+   fence.type = ea.fence_arg.type;
+   fence.flags = ea.fence_arg.flags;
+   fence.signaled = ea.fence_arg.signaled;
 
-   driBOUnmap(batch->buffer);
-   batch->map = NULL;
-
-   intel_batch_ioctl(batch->intel,
-                     driBOOffset(batch->buffer),
-                     used, allow_unlock);
-
-   /*
-    * Kernel fencing. The flags tells the kernel that we've 
-    * programmed an MI_FLUSH.
-    */
-   fenceFlags = DRM_I915_FENCE_FLAG_FLUSHED;
-   fo = driFenceBuffers(batch->intel->driFd, "Batch fence", fenceFlags);
+   fo = driBOFenceUserList(batch->intel->intelScreen->mgr, batch->list,
+			   "SuperFence", &fence);
 
+   if (driFenceType(fo) & DRM_I915_FENCE_TYPE_RW) {
+       if (batch->last_fence)
+	   driFenceUnReference(&batch->last_fence);
    /*
-    * User space fencing.
-    */
-   driBOFence(batch->buffer, fo);
-
-   if (driFenceType(fo) == DRM_FENCE_TYPE_EXE) {
-     /*
-      * Oops. We only validated a batch buffer. This means we
-      * didn't do any proper rendering. Discard this fence object.
-      */
-      driFenceUnReference(fo);
-   }
-   else {
-      driFenceUnReference(batch->last_fence);
-      batch->last_fence = fo;
-      for (i = 0; i < batch->nr_relocs; i++) {
-	struct buffer_reloc *r = &batch->reloc[i];
-	driBOFence(r->buf, fo);
-      }
-   }
+	* FIXME: Context last fence??
+	*/
+       batch->last_fence = fo;
+       driFenceReference(fo);
+   } 
+ out:
+#if 0 /* ZZZ JB: fix this */
+   intel->vtbl.lost_hardware(intel);
+#else
+   (void)intel;
+#endif
+   return fo;
 }
 
 
@@ -263,29 +382,41 @@ struct _DriFenceObject *
 intel_batchbuffer_flush(struct intel_batchbuffer *batch)
 {
    struct intel_context *intel = batch->intel;
-   uint used = batch->ptr - batch->map;
-   const boolean was_locked = intel->locked;
+   GLuint used = batch->ptr - batch->map;
+   GLboolean was_locked = intel->locked;
+   struct _DriFenceObject *fence;
 
    if (used == 0)
       return batch->last_fence;
 
-#define MI_FLUSH ((0 << 29) | (4 << 23))
-
    /* Add the MI_BATCH_BUFFER_END.  Always add an MI_FLUSH - this is a
     * performance drain that we would like to avoid.
     */
+#if 0 /* ZZZ JB: what should we do here? */
    if (used & 4) {
-      ((int *) batch->ptr)[0] = MI_FLUSH;
+      ((int *) batch->ptr)[0] = intel->vtbl.flush_cmd();
       ((int *) batch->ptr)[1] = 0;
       ((int *) batch->ptr)[2] = MI_BATCH_BUFFER_END;
       used += 12;
    }
    else {
-      ((int *) batch->ptr)[0] = MI_FLUSH;
+      ((int *) batch->ptr)[0] = intel->vtbl.flush_cmd();
       ((int *) batch->ptr)[1] = MI_BATCH_BUFFER_END;
       used += 8;
    }
-
+#else
+   if (used & 4) {
+      ((int *) batch->ptr)[0] = ((0<<29)|(4<<23)); // MI_FLUSH;
+      ((int *) batch->ptr)[1] = 0;
+      ((int *) batch->ptr)[2] = (0xA<<23); // MI_BATCH_BUFFER_END;
+      used += 12;
+   }
+   else {
+      ((int *) batch->ptr)[0] = ((0<<29)|(4<<23)); // MI_FLUSH;
+      ((int *) batch->ptr)[1] = (0xA<<23); // MI_BATCH_BUFFER_END;
+      used += 8;
+   }
+#endif
    driBOUnmap(batch->buffer);
    batch->ptr = NULL;
    batch->map = NULL;
@@ -296,7 +427,8 @@ intel_batchbuffer_flush(struct intel_batchbuffer *batch)
    if (!was_locked)
       LOCK_HARDWARE(intel);
 
-   do_flush_locked(batch, used, GL_FALSE);
+   fence = do_flush_locked(batch, used, !(batch->flags & INTEL_BATCH_CLIPRECTS),
+			   GL_FALSE);
 
    if (!was_locked)
       UNLOCK_HARDWARE(intel);
@@ -304,52 +436,20 @@ intel_batchbuffer_flush(struct intel_batchbuffer *batch)
    /* Reset the buffer:
     */
    intel_batchbuffer_reset(batch);
-   return batch->last_fence;
+   return fence;
 }
 
-
 void
 intel_batchbuffer_finish(struct intel_batchbuffer *batch)
 {
    struct _DriFenceObject *fence = intel_batchbuffer_flush(batch);
-   if (fence) {
-      driFenceReference(fence);
-      driFenceFinish(fence,
-                     DRM_FENCE_TYPE_EXE | DRM_I915_FENCE_TYPE_RW,
-                     GL_FALSE);
-      driFenceUnReference(fence);
-   }
+   driFenceFinish(fence, driFenceType(fence), GL_FALSE);
+   driFenceUnReference(&fence);
 }
 
-
-/*  This is the only way buffers get added to the validate list.
- */
-boolean
-intel_batchbuffer_emit_reloc(struct intel_batchbuffer *batch,
-                             struct _DriBufferObject *buffer,
-                             uint flags, uint mask, uint delta)
-{
-   assert(batch->nr_relocs < MAX_RELOCS);
-
-   driBOAddListItem(&batch->list, buffer, flags, mask);
-
-   {
-      struct buffer_reloc *r = &batch->reloc[batch->nr_relocs++];
-      driBOReference(buffer);
-      r->buf = buffer;
-      r->offset = batch->ptr - batch->map;
-      r->delta = delta;
-      *(uint *) batch->ptr = 0x12345678;
-   }
-
-   batch->ptr += 4;
-   return GL_TRUE;
-}
-
-
 void
 intel_batchbuffer_data(struct intel_batchbuffer *batch,
-                       const void *data, uint bytes, uint flags)
+                       const void *data, GLuint bytes, GLuint flags)
 {
    assert((bytes & 3) == 0);
    intel_batchbuffer_require_space(batch, bytes, flags);
diff --git a/src/gallium/winsys/dri/intel/intel_batchbuffer.h b/src/gallium/winsys/dri/intel/intel_batchbuffer.h
index caf6870a3c0..9e4b8043bfb 100644
--- a/src/gallium/winsys/dri/intel/intel_batchbuffer.h
+++ b/src/gallium/winsys/dri/intel/intel_batchbuffer.h
@@ -1,54 +1,20 @@
-/**************************************************************************
- * 
- * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
- * 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 TUNGSTEN GRAPHICS 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.
- * 
- **************************************************************************/
-
 #ifndef INTEL_BATCHBUFFER_H
 #define INTEL_BATCHBUFFER_H
 
-#include "pipe/p_debug.h"
-#include "pipe/p_compiler.h"
-#include "dri_bufmgr.h"
+#include "mtypes.h"
+#include "ws_dri_bufmgr.h"
 
 struct intel_context;
 
 #define BATCH_SZ 16384
 #define BATCH_RESERVED 16
 
-#define MAX_RELOCS 4096
+#define INTEL_DEFAULT_RELOCS 100
+#define INTEL_MAX_RELOCS 400
 
 #define INTEL_BATCH_NO_CLIPRECTS 0x1
 #define INTEL_BATCH_CLIPRECTS    0x2
 
-struct buffer_reloc
-{
-   struct _DriBufferObject *buf;
-   uint offset;
-   uint delta;                /* not needed? */
-};
-
 struct intel_batchbuffer
 {
    struct bufmgr *bm;
@@ -56,19 +22,30 @@ struct intel_batchbuffer
 
    struct _DriBufferObject *buffer;
    struct _DriFenceObject *last_fence;
-   uint flags;
+   GLuint flags;
 
-   drmBOList list;
-   uint list_count;
-   ubyte *map;
-   ubyte *ptr;
+   struct _DriBufferList *list;
+   GLuint list_count;
+   GLubyte *map;
+   GLubyte *ptr;
 
-   struct buffer_reloc reloc[MAX_RELOCS];
-   uint nr_relocs;
-   uint size;
+   uint32_t *reloc;
+   GLuint reloc_size;
+   GLuint nr_relocs;
+
+   GLuint size;
+
+   GLuint dirty_state;
+   GLuint id;
+
+  uint32_t poolOffset;
+  uint8_t *drmBOVirtual;
+  struct _drmBONode *node; /* Validation list node for this buffer */
+  int dest_location;     /* Validation list sequence for this buffer */
 };
 
-struct intel_batchbuffer *intel_batchbuffer_alloc(struct intel_context *intel);
+struct intel_batchbuffer *intel_batchbuffer_alloc(struct intel_context
+                                                  *intel);
 
 void intel_batchbuffer_free(struct intel_batchbuffer *batch);
 
@@ -82,26 +59,28 @@ void intel_batchbuffer_reset(struct intel_batchbuffer *batch);
 
 
 /* Unlike bmBufferData, this currently requires the buffer be mapped.
- * Consider it a convenience function wrapping multiple
+ * Consider it a convenience function wrapping multple
  * intel_buffer_dword() calls.
  */
 void intel_batchbuffer_data(struct intel_batchbuffer *batch,
-                            const void *data, uint bytes, uint flags);
+                            const void *data, GLuint bytes, GLuint flags);
 
 void intel_batchbuffer_release_space(struct intel_batchbuffer *batch,
-                                     uint bytes);
+                                     GLuint bytes);
 
-boolean intel_batchbuffer_emit_reloc(struct intel_batchbuffer *batch,
-                                     struct _DriBufferObject *buffer,
-                                     uint flags,
-                                     uint mask, uint offset);
+void
+intel_offset_relocation(struct intel_batchbuffer *batch,
+			unsigned pre_add,
+			struct _DriBufferObject *driBO,
+			uint64_t val_flags,
+			uint64_t val_mask);
 
 /* Inline functions - might actually be better off with these
  * non-inlined.  Certainly better off switching all command packets to
  * be passed as structs rather than dwords, but that's a little bit of
  * work...
  */
-static INLINE uint
+static INLINE GLuint
 intel_batchbuffer_space(struct intel_batchbuffer *batch)
 {
    return (batch->size - BATCH_RESERVED) - (batch->ptr - batch->map);
@@ -109,22 +88,26 @@ intel_batchbuffer_space(struct intel_batchbuffer *batch)
 
 
 static INLINE void
-intel_batchbuffer_emit_dword(struct intel_batchbuffer *batch, uint dword)
+intel_batchbuffer_emit_dword(struct intel_batchbuffer *batch, GLuint dword)
 {
    assert(batch->map);
    assert(intel_batchbuffer_space(batch) >= 4);
-   *(uint *) (batch->ptr) = dword;
+   *(GLuint *) (batch->ptr) = dword;
    batch->ptr += 4;
 }
 
 static INLINE void
 intel_batchbuffer_require_space(struct intel_batchbuffer *batch,
-                                uint sz, uint flags)
+                                GLuint sz, GLuint flags)
 {
+   struct _DriFenceObject *fence;
+
    assert(sz < batch->size - 8);
    if (intel_batchbuffer_space(batch) < sz ||
-       (batch->flags != 0 && flags != 0 && batch->flags != flags))
-      intel_batchbuffer_flush(batch);
+       (batch->flags != 0 && flags != 0 && batch->flags != flags)) {
+      fence = intel_batchbuffer_flush(batch);
+      driFenceUnReference(&fence);
+   }
 
    batch->flags |= flags;
 }
@@ -134,14 +117,15 @@ intel_batchbuffer_require_space(struct intel_batchbuffer *batch,
 #define BATCH_LOCALS
 
 #define BEGIN_BATCH(n, flags) do {				\
+   assert(!intel->prim.flush);					\
    intel_batchbuffer_require_space(intel->batch, (n)*4, flags);	\
 } while (0)
 
 #define OUT_BATCH(d)  intel_batchbuffer_emit_dword(intel->batch, d)
 
 #define OUT_RELOC(buf,flags,mask,delta) do { 				\
-   assert((delta) >= 0);						\
-   intel_batchbuffer_emit_reloc(intel->batch, buf, flags, mask, delta);	\
+   assert((delta) >= 0);							\
+   intel_offset_relocation(intel->batch, delta, buf, flags, mask); \
 } while (0)
 
 #define ADVANCE_BATCH() do { } while(0)
diff --git a/src/gallium/winsys/dri/intel/intel_batchpool.c b/src/gallium/winsys/dri/intel/intel_batchpool.c
deleted file mode 100644
index ce154c7b884..00000000000
--- a/src/gallium/winsys/dri/intel/intel_batchpool.c
+++ /dev/null
@@ -1,427 +0,0 @@
-/**************************************************************************
- * 
- * Copyright 2006 Tungsten Graphics, Inc., Bismarck, ND., USA
- * 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 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
- * THE COPYRIGHT HOLDERS, AUTHORS 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.
- *
- * The above copyright notice and this permission notice (including the
- * next paragraph) shall be included in all copies or substantial portions
- * of the Software.
- * 
- * 
- **************************************************************************/
-/*
- * Authors: Thomas Hellstr�m <thomas-at-tungstengraphics-dot-com>
- */
-
-/**
- * XXX NOTE: there are no intel dependencies in this file.
- * Rename to dri_batchpool.c?
- */
-
-#include <xf86drm.h>
-#include <stdlib.h>
-#include <stdio.h>
-#include <errno.h>
-
-#include "pipe/p_compiler.h"
-#include "pipe/p_thread.h"
-
-#include "dri_bufpool.h"
-#include "dri_bufmgr.h"
-#include "intel_batchpool.h"
-
-
-typedef struct
-{
-   drmMMListHead head;
-   struct _BPool *parent;
-   struct _DriFenceObject *fence;
-   unsigned long start;
-   int unfenced;
-   int mapped;
-} BBuf;
-
-typedef struct _BPool
-{
-   _glthread_Mutex mutex;
-   unsigned long bufSize;
-   unsigned poolSize;
-   unsigned numFree;
-   unsigned numTot;
-   unsigned numDelayed;
-   unsigned checkDelayed;
-   drmMMListHead free;
-   drmMMListHead delayed;
-   drmMMListHead head;
-   drmBO kernelBO;
-   void *virtual;
-   BBuf *bufs;
-} BPool;
-
-
-static BPool *
-createBPool(int fd, unsigned long bufSize, unsigned numBufs, unsigned flags,
-            unsigned checkDelayed)
-{
-   BPool *p = (BPool *) malloc(sizeof(*p));
-   BBuf *buf;
-   int i;
-
-   if (!p)
-      return NULL;
-
-   p->bufs = (BBuf *) malloc(numBufs * sizeof(*p->bufs));
-   if (!p->bufs) {
-      free(p);
-      return NULL;
-   }
-
-   DRMINITLISTHEAD(&p->free);
-   DRMINITLISTHEAD(&p->head);
-   DRMINITLISTHEAD(&p->delayed);
-
-   p->numTot = numBufs;
-   p->numFree = numBufs;
-   p->bufSize = bufSize;
-   p->numDelayed = 0;
-   p->checkDelayed = checkDelayed;
-
-   _glthread_INIT_MUTEX(p->mutex);
-
-   if (drmBOCreate(fd, 0, numBufs * bufSize, 0, NULL, drm_bo_type_dc,
-                   flags, DRM_BO_HINT_DONT_FENCE, &p->kernelBO)) {
-      free(p->bufs);
-      free(p);
-      return NULL;
-   }
-   if (drmBOMap(fd, &p->kernelBO, DRM_BO_FLAG_READ | DRM_BO_FLAG_WRITE, 0,
-                &p->virtual)) {
-      drmBODestroy(fd, &p->kernelBO);
-      free(p->bufs);
-      free(p);
-      return NULL;
-   }
-
-   /*
-    * We unmap the buffer so that we can validate it later. Note that this is
-    * just a synchronizing operation. The buffer will have a virtual mapping
-    * until it is destroyed.
-    */
-
-   drmBOUnmap(fd, &p->kernelBO);
-
-   buf = p->bufs;
-   for (i = 0; i < numBufs; ++i) {
-      buf->parent = p;
-      buf->fence = NULL;
-      buf->start = i * bufSize;
-      buf->mapped = 0;
-      buf->unfenced = 0;
-      DRMLISTADDTAIL(&buf->head, &p->free);
-      buf++;
-   }
-
-   return p;
-}
-
-
-static void
-pool_checkFree(BPool * p, int wait)
-{
-   drmMMListHead *list, *prev;
-   BBuf *buf;
-   int signaled = 0;
-   int i;
-
-   list = p->delayed.next;
-
-   if (p->numDelayed > 3) {
-      for (i = 0; i < p->numDelayed; i += 3) {
-         list = list->next;
-      }
-   }
-
-   prev = list->prev;
-   for (; list != &p->delayed; list = prev, prev = list->prev) {
-
-      buf = DRMLISTENTRY(BBuf, list, head);
-
-      if (!signaled) {
-         if (wait) {
-            driFenceFinish(buf->fence, DRM_FENCE_TYPE_EXE, 1);
-            signaled = 1;
-         }
-         else {
-            signaled = driFenceSignaled(buf->fence, DRM_FENCE_TYPE_EXE);
-         }
-      }
-
-      if (!signaled)
-         break;
-
-      driFenceUnReference(buf->fence);
-      buf->fence = NULL;
-      DRMLISTDEL(list);
-      p->numDelayed--;
-      DRMLISTADD(list, &p->free);
-      p->numFree++;
-   }
-}
-
-static void *
-pool_create(struct _DriBufferPool *pool,
-            unsigned long size, unsigned flags, unsigned hint,
-            unsigned alignment)
-{
-   BPool *p = (BPool *) pool->data;
-
-   drmMMListHead *item;
-
-   if (alignment && (alignment != 4096))
-      return NULL;
-
-   _glthread_LOCK_MUTEX(p->mutex);
-
-   if (p->numFree == 0)
-      pool_checkFree(p, TRUE);
-
-   if (p->numFree == 0) {
-      fprintf(stderr, "Out of fixed size buffer objects\n");
-      BM_CKFATAL(-ENOMEM);
-   }
-
-   item = p->free.next;
-
-   if (item == &p->free) {
-      fprintf(stderr, "Fixed size buffer pool corruption\n");
-   }
-
-   DRMLISTDEL(item);
-   --p->numFree;
-
-   _glthread_UNLOCK_MUTEX(p->mutex);
-   return (void *) DRMLISTENTRY(BBuf, item, head);
-}
-
-
-static int
-pool_destroy(struct _DriBufferPool *pool, void *private)
-{
-   BBuf *buf = (BBuf *) private;
-   BPool *p = buf->parent;
-
-   _glthread_LOCK_MUTEX(p->mutex);
-
-   if (buf->fence) {
-      DRMLISTADDTAIL(&buf->head, &p->delayed);
-      p->numDelayed++;
-   }
-   else {
-      buf->unfenced = 0;
-      DRMLISTADD(&buf->head, &p->free);
-      p->numFree++;
-   }
-
-   if ((p->numDelayed % p->checkDelayed) == 0)
-      pool_checkFree(p, 0);
-
-   _glthread_UNLOCK_MUTEX(p->mutex);
-   return 0;
-}
-
-
-static int
-pool_map(struct _DriBufferPool *pool, void *private, unsigned flags,
-         int hint, void **virtual)
-{
-
-   BBuf *buf = (BBuf *) private;
-   BPool *p = buf->parent;
-
-   _glthread_LOCK_MUTEX(p->mutex);
-
-   /*
-    * Currently Mesa doesn't have any condition variables to resolve this
-    * cleanly in a multithreading environment.
-    * We bail out instead.
-    */
-
-   if (buf->mapped) {
-      fprintf(stderr, "Trying to map already mapped buffer object\n");
-      BM_CKFATAL(-EINVAL);
-   }
-
-#if 0
-   if (buf->unfenced && !(hint & DRM_BO_HINT_ALLOW_UNFENCED_MAP)) {
-      fprintf(stderr, "Trying to map an unfenced buffer object 0x%08x"
-              " 0x%08x %d\n", hint, flags, buf->start);
-      BM_CKFATAL(-EINVAL);
-   }
-
-#endif
-
-   if (buf->fence) {
-      _glthread_UNLOCK_MUTEX(p->mutex);
-      return -EBUSY;
-   }
-
-   buf->mapped = TRUE;
-   *virtual = (unsigned char *) p->virtual + buf->start;
-   _glthread_UNLOCK_MUTEX(p->mutex);
-   return 0;
-}
-
-static int
-pool_waitIdle(struct _DriBufferPool *pool, void *private, int lazy)
-{
-   BBuf *buf = (BBuf *) private;
-   driFenceFinish(buf->fence, 0x0, lazy);
-   return 0;
-}
-
-static int
-pool_unmap(struct _DriBufferPool *pool, void *private)
-{
-   BBuf *buf = (BBuf *) private;
-
-   buf->mapped = 0;
-   return 0;
-}
-
-static unsigned long
-pool_offset(struct _DriBufferPool *pool, void *private)
-{
-   BBuf *buf = (BBuf *) private;
-   BPool *p = buf->parent;
-
-   return p->kernelBO.offset + buf->start;
-}
-
-static unsigned
-pool_flags(struct _DriBufferPool *pool, void *private)
-{
-   BPool *p = (BPool *) pool->data;
-
-   return p->kernelBO.flags;
-}
-
-static unsigned long
-pool_size(struct _DriBufferPool *pool, void *private)
-{
-   BPool *p = (BPool *) pool->data;
-
-   return p->bufSize;
-}
-
-
-static int
-pool_fence(struct _DriBufferPool *pool, void *private,
-           struct _DriFenceObject *fence)
-{
-   BBuf *buf = (BBuf *) private;
-   BPool *p = buf->parent;
-
-   _glthread_LOCK_MUTEX(p->mutex);
-   if (buf->fence) {
-      driFenceUnReference(buf->fence);
-   }
-   buf->fence = fence;
-   buf->unfenced = 0;
-   driFenceReference(buf->fence);
-   _glthread_UNLOCK_MUTEX(p->mutex);
-
-   return 0;
-}
-
-static drmBO *
-pool_kernel(struct _DriBufferPool *pool, void *private)
-{
-   BBuf *buf = (BBuf *) private;
-   BPool *p = buf->parent;
-
-   return &p->kernelBO;
-}
-
-static int
-pool_validate(struct _DriBufferPool *pool, void *private)
-{
-   BBuf *buf = (BBuf *) private;
-   BPool *p = buf->parent;
-   _glthread_LOCK_MUTEX(p->mutex);
-   buf->unfenced = TRUE;
-   _glthread_UNLOCK_MUTEX(p->mutex);
-   return 0;
-}
-
-static void
-pool_takedown(struct _DriBufferPool *pool)
-{
-   BPool *p = (BPool *) pool->data;
-
-   /*
-    * Wait on outstanding fences. 
-    */
-
-   _glthread_LOCK_MUTEX(p->mutex);
-   while ((p->numFree < p->numTot) && p->numDelayed) {
-      _glthread_UNLOCK_MUTEX(p->mutex);
-      sched_yield();
-      pool_checkFree(p, TRUE);
-      _glthread_LOCK_MUTEX(p->mutex);
-   }
-
-   drmBODestroy(pool->fd, &p->kernelBO);
-   free(p->bufs);
-   _glthread_UNLOCK_MUTEX(p->mutex);
-   free(p);
-   free(pool);
-}
-
-
-struct _DriBufferPool *
-driBatchPoolInit(int fd, unsigned flags,
-                 unsigned long bufSize,
-                 unsigned numBufs, unsigned checkDelayed)
-{
-   struct _DriBufferPool *pool;
-
-   pool = (struct _DriBufferPool *) malloc(sizeof(*pool));
-   if (!pool)
-      return NULL;
-
-   pool->data = createBPool(fd, bufSize, numBufs, flags, checkDelayed);
-   if (!pool->data)
-      return NULL;
-
-   pool->fd = fd;
-   pool->map = &pool_map;
-   pool->unmap = &pool_unmap;
-   pool->destroy = &pool_destroy;
-   pool->offset = &pool_offset;
-   pool->flags = &pool_flags;
-   pool->size = &pool_size;
-   pool->create = &pool_create;
-   pool->fence = &pool_fence;
-   pool->kernel = &pool_kernel;
-   pool->validate = &pool_validate;
-   pool->waitIdle = &pool_waitIdle;
-   pool->setstatic = NULL;
-   pool->takeDown = &pool_takedown;
-   return pool;
-}
diff --git a/src/gallium/winsys/dri/intel/intel_context.c b/src/gallium/winsys/dri/intel/intel_context.c
index 8eba33c3139..cc2fc11c5e7 100644
--- a/src/gallium/winsys/dri/intel/intel_context.c
+++ b/src/gallium/winsys/dri/intel/intel_context.c
@@ -162,6 +162,7 @@ intelCreateContext(const __GLcontextModes * visual,
     * memory pools
     */
    DRM_LIGHT_LOCK(sPriv->fd, &sPriv->pSAREA->lock, driContextPriv->hHWContext);
+   // ZZZ JB should be per screen and not be done per context
    havePools = intelCreatePools(sPriv);
    DRM_UNLOCK(sPriv->fd, &sPriv->pSAREA->lock, driContextPriv->hHWContext);
    if (!havePools)
@@ -234,12 +235,12 @@ intelDestroyContext(__DRIcontextPrivate * driContextPriv)
 
       if (intel->last_swap_fence) {
 	 driFenceFinish(intel->last_swap_fence, DRM_FENCE_TYPE_EXE, GL_TRUE);
-	 driFenceUnReference(intel->last_swap_fence);
+	 driFenceUnReference(&intel->last_swap_fence);
 	 intel->last_swap_fence = NULL;
       }
       if (intel->first_swap_fence) {
 	 driFenceFinish(intel->first_swap_fence, DRM_FENCE_TYPE_EXE, GL_TRUE);
-	 driFenceUnReference(intel->first_swap_fence);
+	 driFenceUnReference(&intel->first_swap_fence);
 	 intel->first_swap_fence = NULL;
       }
 
diff --git a/src/gallium/winsys/dri/intel/intel_context.h b/src/gallium/winsys/dri/intel/intel_context.h
index 45430984d8f..597dc13830b 100644
--- a/src/gallium/winsys/dri/intel/intel_context.h
+++ b/src/gallium/winsys/dri/intel/intel_context.h
@@ -28,7 +28,7 @@
 #ifndef INTEL_CONTEXT_H
 #define INTEL_CONTEXT_H
 
-
+#include <stdint.h>
 #include "drm.h"
 
 #include "pipe/p_debug.h"
diff --git a/src/gallium/winsys/dri/intel/intel_ioctl.c b/src/gallium/winsys/dri/intel/intel_ioctl.c
new file mode 100644
index 00000000000..3250c6b3a9c
--- /dev/null
+++ b/src/gallium/winsys/dri/intel/intel_ioctl.c
@@ -0,0 +1,138 @@
+/**************************************************************************
+ * 
+ * Copyright 2003 Tungsten Graphics, Inc., Cedar Park, Texas.
+ * 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 TUNGSTEN GRAPHICS 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.
+ * 
+ **************************************************************************/
+
+
+#include <stdio.h>
+#include <unistd.h>
+#include <errno.h>
+#include <sched.h>
+
+#include "mtypes.h"
+#include "context.h"
+#include "swrast/swrast.h"
+
+#include "intel_context.h"
+#include "intel_ioctl.h"
+#include "intel_batchbuffer.h"
+#include "intel_blit.h"
+#include "intel_regions.h"
+#include "drm.h"
+
+#define FILE_DEBUG_FLAG DEBUG_IOCTL
+
+int
+intelEmitIrqLocked(struct intel_context *intel)
+{
+   drmI830IrqEmit ie;
+   int ret, seq;
+
+   assert(((*(int *) intel->driHwLock) & ~DRM_LOCK_CONT) ==
+          (DRM_LOCK_HELD | intel->hHWContext));
+
+   ie.irq_seq = &seq;
+
+   ret = drmCommandWriteRead(intel->driFd, DRM_I830_IRQ_EMIT,
+                             &ie, sizeof(ie));
+   if (ret) {
+      fprintf(stderr, "%s: drmI830IrqEmit: %d\n", __FUNCTION__, ret);
+      exit(1);
+   }
+
+   DBG("%s -->  %d\n", __FUNCTION__, seq);
+
+   return seq;
+}
+
+void
+intelWaitIrq(struct intel_context *intel, int seq)
+{
+   int ret;
+
+   DBG("%s %d\n", __FUNCTION__, seq);
+
+   intel->iw.irq_seq = seq;
+
+   do {
+      ret =
+         drmCommandWrite(intel->driFd, DRM_I830_IRQ_WAIT, &intel->iw,
+                         sizeof(intel->iw));
+   } while (ret == -EAGAIN || ret == -EINTR);
+
+   if (ret) {
+      fprintf(stderr, "%s: drmI830IrqWait: %d\n", __FUNCTION__, ret);
+      exit(1);
+   }
+}
+
+
+void
+intel_batch_ioctl(struct intel_context *intel,
+                  GLuint start_offset,
+                  GLuint used,
+                  GLboolean ignore_cliprects, GLboolean allow_unlock)
+{
+   drmI830BatchBuffer batch;
+
+   assert(intel->locked);
+   assert(used);
+
+   DBG("%s used %d offset %x..%x ignore_cliprects %d\n",
+       __FUNCTION__,
+       used, start_offset, start_offset + used, ignore_cliprects);
+
+   /* Throw away non-effective packets.  Won't work once we have
+    * hardware contexts which would preserve statechanges beyond a
+    * single buffer.
+    */
+
+
+
+   batch.start = start_offset;
+   batch.used = used;
+   batch.cliprects = intel->pClipRects;
+   batch.num_cliprects = ignore_cliprects ? 0 : intel->numClipRects;
+   batch.DR1 = 0;
+   batch.DR4 = ((((GLuint) intel->drawX) & 0xffff) |
+                (((GLuint) intel->drawY) << 16));
+
+   DBG("%s: 0x%x..0x%x DR4: %x cliprects: %d\n",
+       __FUNCTION__,
+       batch.start,
+       batch.start + batch.used * 4, batch.DR4, batch.num_cliprects);
+
+   if (drmCommandWrite(intel->driFd, DRM_I830_BATCHBUFFER, &batch,
+                       sizeof(batch))) {
+      fprintf(stderr, "DRM_I830_BATCHBUFFER: %d\n", -errno);
+      UNLOCK_HARDWARE(intel);
+      exit(1);
+   }
+
+   /* FIXME: use hardware contexts to avoid 'losing' hardware after
+    * each buffer flush.
+    */
+   intel->vtbl.lost_hardware(intel);
+}
diff --git a/src/gallium/winsys/dri/intel/intel_batchpool.h b/src/gallium/winsys/dri/intel/intel_ioctl.h
index f6a95723bcf..e8d07de893e 100644
--- a/src/gallium/winsys/dri/intel/intel_batchpool.h
+++ b/src/gallium/winsys/dri/intel/intel_ioctl.h
@@ -25,13 +25,16 @@
  * 
  **************************************************************************/
 
-#ifndef INTEL_BATCHPOOL_H
-#define INTEL_BATCHPOOL_H
+#ifndef INTEL_IOCTL_H
+#define INTEL_IOCTL_H
 
-extern struct _DriBufferPool *driBatchPoolInit(int fd, unsigned flags,
-                                               unsigned long bufSize,
-                                               unsigned numBufs,
-                                               unsigned checkDelayed);
+#include "intel_context.h"
 
+void intelWaitIrq(struct intel_context *intel, int seq);
+int intelEmitIrqLocked(struct intel_context *intel);
 
-#endif /* INTEL_BATCHPOOL_H */
+void intel_batch_ioctl(struct intel_context *intel,
+                       GLuint start_offset,
+                       GLuint used,
+                       GLboolean ignore_cliprects, GLboolean allow_unlock);
+#endif
diff --git a/src/gallium/winsys/dri/intel/intel_screen.c b/src/gallium/winsys/dri/intel/intel_screen.c
index 9e31c013a95..a3fc6d73444 100644
--- a/src/gallium/winsys/dri/intel/intel_screen.c
+++ b/src/gallium/winsys/dri/intel/intel_screen.c
@@ -32,12 +32,12 @@
 #include "intel_context.h"
 #include "intel_screen.h"
 #include "intel_batchbuffer.h"
-#include "intel_batchpool.h"
+//#include "intel_batchpool.h"
 #include "intel_swapbuffers.h"
 #include "intel_winsys.h"
 
 #include "i830_dri.h"
-#include "dri_bufpool.h"
+#include "ws_dri_bufpool.h"
 
 #include "pipe/p_context.h"
 #include "state_tracker/st_public.h"
@@ -132,6 +132,7 @@ intelUpdateScreenRotation(__DRIscreenPrivate * sPriv, drmI830Sarea * sarea)
    assert( sarea->front_size >=
 	   intelScreen->front.pitch * intelScreen->front.height );
 
+#if 0 /* JB not important */
    if (!sarea->front_handle)
       return;
 
@@ -142,30 +143,41 @@ intelUpdateScreenRotation(__DRIscreenPrivate * sPriv, drmI830Sarea * sarea)
       fprintf(stderr, "drmMap(frontbuffer) failed!\n");
       return;
    }
+#endif
 
+#if 0 /* JB */
    if (intelScreen->staticPool) {
       driGenBuffers(intelScreen->staticPool, "static region", 1,
 		    &intelScreen->front.buffer, 64,
 		    DRM_BO_FLAG_MEM_TT | DRM_BO_FLAG_NO_MOVE |
 		    DRM_BO_FLAG_READ | DRM_BO_FLAG_WRITE, 0);
-      
+
       driBOSetStatic(intelScreen->front.buffer, 
 		     intelScreen->front.offset, 		  
 		     intelScreen->front.pitch * intelScreen->front.height, 
 		     intelScreen->front.map, 0);
    }
+#else
+   if (intelScreen->staticPool) {
+      if (intelScreen->front.buffer)
+	 driBOUnReference(intelScreen->front.buffer);
+      driGenBuffers(intelScreen->staticPool, "front", 1, &intelScreen->front.buffer, 0, 0, 0);
+      driBOSetReferenced(intelScreen->front.buffer, sarea->front_bo_handle);
+   }
+#endif
 }
 
 
 boolean
 intelCreatePools(__DRIscreenPrivate * sPriv)
 {
-   unsigned batchPoolSize = 1024*1024;
+   //unsigned batchPoolSize = 1024*1024;
    struct intel_screen *intelScreen = intel_screen(sPriv);
 
    if (intelScreen->havePools)
       return GL_TRUE;
 
+#if 0 /* ZZZ JB fix this */
    intelScreen->staticPool = driDRMStaticPoolInit(sPriv->fd);
    if (!intelScreen->staticPool)
       return GL_FALSE;
@@ -181,7 +193,17 @@ intelCreatePools(__DRIscreenPrivate * sPriv)
       fprintf(stderr, "Failed to initialize batch pool - possible incorrect agpgart installed\n");
       return GL_FALSE;
    }
-   
+#else
+   intelScreen->staticPool = driDRMPoolInit(sPriv->fd);
+   intelScreen->batchPool = driSlabPoolInit(sPriv->fd,
+						DRM_BO_FLAG_EXE |
+						DRM_BO_FLAG_MEM_TT,
+						DRM_BO_FLAG_EXE |
+						DRM_BO_FLAG_MEM_TT,
+						4096, //intelScreen->maxBatchSize,
+						1, 40, 16*16384, 0,
+						intelScreen->fMan);
+#endif
    intelScreen->havePools = GL_TRUE;
 
    intelUpdateScreenRotation(sPriv, intelScreen->sarea);
@@ -240,7 +262,26 @@ intelInitDriver(__DRIscreenPrivate * sPriv)
       (*glx_enable_extension) (psc, "GLX_SGI_make_current_read");
    }
 
-   intelScreen->winsys = intel_create_pipe_winsys(sPriv->fd);
+
+
+#if 1 // ZZZ JB
+   intelScreen->mgr = driFenceMgrTTMInit(sPriv->fd);
+   if (!intelScreen->mgr) {
+      fprintf(stderr, "Failed to create fence manager.\n");
+      return GL_FALSE;
+   }
+
+   intelScreen->fMan = driInitFreeSlabManager(10, 10);
+   if (!intelScreen->fMan) {
+      fprintf(stderr, "Failed to create free slab manager.\n");
+      return GL_FALSE;
+   }
+
+   if (!intelCreatePools(sPriv))
+      return GL_FALSE;
+#endif
+
+   intelScreen->winsys = intel_create_pipe_winsys(sPriv->fd, intelScreen->fMan);
 
    return GL_TRUE;
 }
diff --git a/src/gallium/winsys/dri/intel/intel_screen.h b/src/gallium/winsys/dri/intel/intel_screen.h
index 3396f9e5643..c17f0202e4f 100644
--- a/src/gallium/winsys/dri/intel/intel_screen.h
+++ b/src/gallium/winsys/dri/intel/intel_screen.h
@@ -31,7 +31,7 @@
 #include "dri_util.h"
 #include "i830_common.h"
 #include "xmlconfig.h"
-#include "dri_bufpool.h"
+#include "ws_dri_bufpool.h"
 
 #include "pipe/p_compiler.h"
 
@@ -74,6 +74,13 @@ struct intel_screen
     */
    struct intel_context *dummyContext;
 
+   /* 
+    * New stuff form the i915tex integration
+    */
+   struct _DriFenceMgr *mgr;
+   struct _DriFreeSlabManager *fMan;
+   unsigned batch_id;
+
    struct pipe_winsys *winsys;
 };
 
diff --git a/src/gallium/winsys/dri/intel/intel_swapbuffers.c b/src/gallium/winsys/dri/intel/intel_swapbuffers.c
index 56b86d6a636..1a42389169d 100644
--- a/src/gallium/winsys/dri/intel/intel_swapbuffers.c
+++ b/src/gallium/winsys/dri/intel/intel_swapbuffers.c
@@ -63,7 +63,7 @@ intelDisplaySurface(__DRIdrawablePrivate *dPriv,
 
    if (intel->last_swap_fence) {
       driFenceFinish(intel->last_swap_fence, DRM_FENCE_TYPE_EXE, TRUE);
-      driFenceUnReference(intel->last_swap_fence);
+      driFenceUnReference(&intel->last_swap_fence);
       intel->last_swap_fence = NULL;
    }
    intel->last_swap_fence = intel->first_swap_fence;
@@ -178,7 +178,7 @@ intelDisplaySurface(__DRIdrawablePrivate *dPriv,
       }
 
       if (intel->first_swap_fence)
-	 driFenceUnReference(intel->first_swap_fence);
+	 driFenceUnReference(&intel->first_swap_fence);
       intel->first_swap_fence = intel_batchbuffer_flush(intel->batch);
       driFenceReference(intel->first_swap_fence);
    }
diff --git a/src/gallium/winsys/dri/intel/intel_winsys.h b/src/gallium/winsys/dri/intel/intel_winsys.h
index ffc40782be7..d0a319f9a4f 100644
--- a/src/gallium/winsys/dri/intel/intel_winsys.h
+++ b/src/gallium/winsys/dri/intel/intel_winsys.h
@@ -37,7 +37,7 @@ struct pipe_buffer;
 struct _DriBufferObject;
 
 struct pipe_winsys *
-intel_create_pipe_winsys( int fd );
+intel_create_pipe_winsys( int fd, struct _DriFreeSlabManager *fMan );
 
 void
 intel_destroy_pipe_winsys( struct pipe_winsys *winsys );
diff --git a/src/gallium/winsys/dri/intel/intel_winsys_i915.c b/src/gallium/winsys/dri/intel/intel_winsys_i915.c
index 4d183db7c37..ae3e9654d32 100644
--- a/src/gallium/winsys/dri/intel/intel_winsys_i915.c
+++ b/src/gallium/winsys/dri/intel/intel_winsys_i915.c
@@ -31,8 +31,8 @@
 
 #include <stdlib.h>
 #include <xf86drm.h>
-#include "dri_bufpool.h"
-#include "dri_bufmgr.h"
+#include "ws_dri_bufpool.h"
+#include "ws_dri_bufmgr.h"
 
 #include "intel_context.h"
 #include "intel_batchbuffer.h"
@@ -106,10 +106,18 @@ static void intel_i915_batch_reloc( struct i915_winsys *sws,
       mask |= DRM_BO_FLAG_READ;
    }
 
-   intel_batchbuffer_emit_reloc( intel->batch, 
+#if 0 /* JB old */
+   intel_batchbuffer_emit_reloc( intel->batch,
 				 dri_bo( buf ),
-				 flags, mask, 
+				 flags, mask,
 				 delta );
+#else /* new */
+   intel_offset_relocation( intel->batch,
+				delta,
+				dri_bo( buf ),
+				flags,
+				mask );
+#endif
 }
 
 
@@ -126,8 +134,18 @@ static void intel_i915_batch_flush( struct i915_winsys *sws,
 
    fu.dri = intel_batchbuffer_flush( intel->batch );
 
-   if (fu.dri)
-      iws->pws->fence_reference(iws->pws, fence, fu.pipe);
+   if (!fu.dri && fence) {
+      *fence = NULL;
+      return;
+   }
+
+   if (fu.dri) {
+      if (fence)
+	 *fence = fu.pipe;
+      else
+	 iws->pws->fence_reference(iws->pws, &fu.dri, NULL);
+   }
+
 
 //   if (0) intel_i915_batch_wait_idle( sws );
 }
diff --git a/src/gallium/winsys/dri/intel/intel_winsys_pipe.c b/src/gallium/winsys/dri/intel/intel_winsys_pipe.c
index 77dec9488df..59729a0245d 100644
--- a/src/gallium/winsys/dri/intel/intel_winsys_pipe.c
+++ b/src/gallium/winsys/dri/intel/intel_winsys_pipe.c
@@ -31,8 +31,8 @@
 
 #include <stdlib.h>
 #include <xf86drm.h>
-#include "dri_bufpool.h"
-#include "dri_bufmgr.h"
+//#include "dri_bufpool.h"
+//#include "dri_bufmgr.h"
 
 #include "intel_context.h"
 #include "intel_winsys.h"
@@ -50,6 +50,7 @@
 struct intel_pipe_winsys {
    struct pipe_winsys winsys;
    struct _DriBufferPool *regionPool;
+   struct _DriFreeSlabManager *fMan;
 };
 
 
@@ -142,7 +143,7 @@ intel_buffer_create(struct pipe_winsys *winsys,
    driGenBuffers( iws->regionPool, 
 		  "pipe buffer", 1, &buffer->driBO, alignment, flags, 0 );
 
-   driBOData( buffer->driBO, size, NULL, 0 );
+   driBOData( buffer->driBO, size, NULL, iws->regionPool, 0 );
 
    return &buffer->base;
 }
@@ -155,7 +156,7 @@ intel_user_buffer_create(struct pipe_winsys *winsys, void *ptr, unsigned bytes)
    struct intel_pipe_winsys *iws = intel_pipe_winsys(winsys);
 
    driGenUserBuffer( iws->regionPool, 
-                     "pipe user buffer", &buffer->driBO, ptr, bytes);
+                     "pipe user buffer", &buffer->driBO, ptr, bytes );
 
    return &buffer->base;
 }
@@ -209,7 +210,7 @@ intel_i915_surface_alloc_storage(struct pipe_winsys *winsys,
                                  unsigned flags)
 {
    const unsigned alignment = 64;
-   int ret;
+   //int ret;
 
    surf->width = width;
    surf->height = height;
@@ -249,9 +250,37 @@ intel_get_name( struct pipe_winsys *winsys )
    return "Intel/DRI/ttm";
 }
 
+static void
+intel_fence_reference( struct pipe_winsys *sws,
+                       struct pipe_fence_handle **ptr,
+                       struct pipe_fence_handle *fence )
+{
+   if (*ptr)
+      driFenceUnReference((struct _DriFenceObject **)ptr);
+
+   if (fence)
+      *ptr = (struct pipe_fence_handle *)driFenceReference((struct _DriFenceObject *)fence);
+}
+
+static int
+intel_fence_signalled( struct pipe_winsys *sws,
+                       struct pipe_fence_handle *fence,
+                       unsigned flag )
+{
+   return driFenceSignaled((struct _DriFenceObject *)fence, flag);
+}
+
+static int
+intel_fence_finish( struct pipe_winsys *sws,
+                    struct pipe_fence_handle *fence,
+                    unsigned flag )
+{
+   /* JB: Lets allways lazy wait */
+   return driFenceFinish((struct _DriFenceObject *)fence, flag, 1);
+}
 
 struct pipe_winsys *
-intel_create_pipe_winsys( int fd )
+intel_create_pipe_winsys( int fd, struct _DriFreeSlabManager *fMan )
 {
    struct intel_pipe_winsys *iws = CALLOC_STRUCT( intel_pipe_winsys );
    
@@ -273,8 +302,20 @@ intel_create_pipe_winsys( int fd )
    iws->winsys.surface_alloc_storage = intel_i915_surface_alloc_storage;
    iws->winsys.surface_release = intel_i915_surface_release;
 
+   iws->winsys.fence_reference = intel_fence_reference;
+   iws->winsys.fence_signalled = intel_fence_signalled;
+   iws->winsys.fence_finish = intel_fence_finish;
+
    if (fd)
-      iws->regionPool = driDRMPoolInit(fd);
+      iws->regionPool = driSlabPoolInit(fd,
+			DRM_BO_FLAG_READ |
+			DRM_BO_FLAG_WRITE |
+			DRM_BO_FLAG_MEM_TT,
+			DRM_BO_FLAG_READ |
+			DRM_BO_FLAG_WRITE |
+			DRM_BO_FLAG_MEM_TT,
+			64, 6, 16, 4096, 0,
+			fMan);
 
    return &iws->winsys;
 }
diff --git a/src/gallium/winsys/dri/intel/server/i830_common.h b/src/gallium/winsys/dri/intel/server/i830_common.h
index d4d58886ce0..f1fd3939ab9 100644
--- a/src/gallium/winsys/dri/intel/server/i830_common.h
+++ b/src/gallium/winsys/dri/intel/server/i830_common.h
@@ -54,6 +54,7 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 #define DRM_I830_DESTROY_HEAP             0x0c
 #define DRM_I830_SET_VBLANK_PIPE          0x0d
 #define DRM_I830_GET_VBLANK_PIPE          0x0e
+#define DRM_I830_MMIO		       	  0x10
 
 typedef struct {
    enum {
@@ -85,6 +86,7 @@ typedef struct {
         int last_enqueue;	/* last time a buffer was enqueued */
 	int last_dispatch;	/* age of the most recently dispatched buffer */
 	int ctxOwner;		/* last context to upload state */
+	/** Last context that used the buffer manager. */
 	int texAge;
         int pf_enabled;		/* is pageflipping allowed? */
         int pf_active;               
@@ -121,20 +123,29 @@ typedef struct {
 	unsigned int rotated_tiled;
 	unsigned int rotated2_tiled;
 
-	int pipeA_x;
-	int pipeA_y;
-	int pipeA_w;
-	int pipeA_h;
-	int pipeB_x;
-	int pipeB_y;
-	int pipeB_w;
-	int pipeB_h;
+	int planeA_x;
+	int planeA_y;
+	int planeA_w;
+	int planeA_h;
+	int planeB_x;
+	int planeB_y;
+	int planeB_w;
+	int planeB_h;
 
 	/* Triple buffering */
 	drm_handle_t third_handle;
 	int third_offset;
 	int third_size;
 	unsigned int third_tiled;
+
+	/* buffer object handles for the static buffers.  May change
+	 * over the lifetime of the client, though it doesn't in our current
+	 * implementation.
+	 */
+	unsigned int front_bo_handle;
+	unsigned int back_bo_handle;
+	unsigned int third_bo_handle;
+	unsigned int depth_bo_handle;
 } drmI830Sarea;
 
 /* Flags for perf_boxes
@@ -223,4 +234,23 @@ typedef struct {
         int pipe;
 } drmI830VBlankPipe;
 
+#define MMIO_READ  0
+#define MMIO_WRITE 1
+
+#define MMIO_REGS_IA_PRIMATIVES_COUNT           0
+#define MMIO_REGS_IA_VERTICES_COUNT             1
+#define MMIO_REGS_VS_INVOCATION_COUNT           2
+#define MMIO_REGS_GS_PRIMITIVES_COUNT           3
+#define MMIO_REGS_GS_INVOCATION_COUNT           4
+#define MMIO_REGS_CL_PRIMITIVES_COUNT           5
+#define MMIO_REGS_CL_INVOCATION_COUNT           6
+#define MMIO_REGS_PS_INVOCATION_COUNT           7
+#define MMIO_REGS_PS_DEPTH_COUNT                8
+
+typedef struct {
+        unsigned int read_write:1;
+        unsigned int reg:31;
+        void __user *data;
+} drmI830MMIO;
+
 #endif /* _I830_DRM_H_ */
diff --git a/src/gallium/winsys/dri/intel/ws_dri_bufmgr.c b/src/gallium/winsys/dri/intel/ws_dri_bufmgr.c
new file mode 100644
index 00000000000..eb5216cb372
--- /dev/null
+++ b/src/gallium/winsys/dri/intel/ws_dri_bufmgr.c
@@ -0,0 +1,945 @@
+/**************************************************************************
+ * 
+ * Copyright 2006 Tungsten Graphics, Inc., Bismarck, ND., USA
+ * 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 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
+ * THE COPYRIGHT HOLDERS, AUTHORS 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.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ * 
+ * 
+ **************************************************************************/
+/*
+ * Authors: Thomas Hellstr�m <thomas-at-tungstengraphics-dot-com>
+ *          Keith Whitwell <keithw-at-tungstengraphics-dot-com>
+ */
+
+#include <xf86drm.h>
+#include <stdlib.h>
+#include "glthread.h"
+#include "errno.h"
+#include "ws_dri_bufmgr.h"
+#include "string.h"
+#include "imports.h"
+#include "ws_dri_bufpool.h"
+#include "ws_dri_fencemgr.h"
+
+/*
+ * This lock is here to protect drmBO structs changing underneath us during a
+ * validate list call, since validatelist cannot take individiual locks for
+ * each drmBO. Validatelist takes this lock in write mode. Any access to an
+ * individual drmBO should take this lock in read mode, since in that case, the
+ * driBufferObject mutex will protect the access. Locking order is 
+ * driBufferObject mutex - > this rw lock.
+ */
+
+_glthread_DECLARE_STATIC_MUTEX(bmMutex);
+_glthread_DECLARE_STATIC_COND(bmCond);
+
+static int kernelReaders = 0;
+
+static drmBO *drmBOListBuf(void *iterator)
+{
+    drmBONode *node;
+    drmMMListHead *l = (drmMMListHead *) iterator;
+    node = DRMLISTENTRY(drmBONode, l, head);
+    return node->buf;
+}
+
+static void *drmBOListIterator(drmBOList *list)
+{
+    void *ret = list->list.next;
+
+    if (ret == &list->list)
+	return NULL;
+    return ret;
+}
+
+static void *drmBOListNext(drmBOList *list, void *iterator)
+{
+    void *ret;
+
+    drmMMListHead *l = (drmMMListHead *) iterator;
+    ret = l->next;
+    if (ret == &list->list)
+	return NULL;
+    return ret;
+}
+
+static drmBONode *drmAddListItem(drmBOList *list, drmBO *item, 
+				 uint64_t arg0,
+				 uint64_t arg1)
+{
+    drmBONode *node;
+    drmMMListHead *l;
+
+    l = list->free.next;
+    if (l == &list->free) {
+	node = (drmBONode *) malloc(sizeof(*node));
+	if (!node) {
+	    return NULL;
+	}
+	list->numCurrent++;
+    }
+    else {
+	DRMLISTDEL(l);
+	node = DRMLISTENTRY(drmBONode, l, head);
+    }
+    node->buf = item;
+    node->arg0 = arg0;
+    node->arg1 = arg1;
+    DRMLISTADD(&node->head, &list->list);
+    list->numOnList++;
+    return node;
+}
+  
+static int drmAddValidateItem(drmBOList *list, drmBO *buf, uint64_t flags, 
+			      uint64_t mask, int *newItem)
+{
+    drmBONode *node, *cur;
+    drmMMListHead *l;
+
+    *newItem = 0;
+    cur = NULL;
+
+    for (l = list->list.next; l != &list->list; l = l->next) {
+	node = DRMLISTENTRY(drmBONode, l, head);
+	if (node->buf == buf) {
+	    cur = node;
+	    break;
+	}
+    }
+    if (!cur) {
+	cur = drmAddListItem(list, buf, flags, mask);
+	if (!cur) {
+	    return -ENOMEM;
+	}
+	*newItem = 1;
+	cur->arg0 = flags;
+	cur->arg1 = mask;
+    }
+    else {
+        uint64_t memFlags = cur->arg0 & flags & DRM_BO_MASK_MEM;
+	uint64_t accFlags = (cur->arg0 | flags) & ~DRM_BO_MASK_MEM;
+
+	if (mask & cur->arg1 & ~DRM_BO_MASK_MEM  & (cur->arg0 ^ flags)) {
+	    return -EINVAL;
+	}
+
+	cur->arg1 |= mask;
+	cur->arg0 = (cur->arg0 & ~mask) | ((memFlags | accFlags) & mask);
+
+	if (((cur->arg1 & DRM_BO_MASK_MEM) != 0) &&
+	    (cur->arg0 & DRM_BO_MASK_MEM) == 0) {
+	    return -EINVAL;
+	}
+    }
+    return 0;
+}
+
+static void drmBOFreeList(drmBOList *list)
+{
+    drmBONode *node;
+    drmMMListHead *l;
+
+    l = list->list.next;
+    while(l != &list->list) {
+	DRMLISTDEL(l);
+	node = DRMLISTENTRY(drmBONode, l, head);
+	free(node);
+	l = list->list.next;
+	list->numCurrent--;
+	list->numOnList--;
+    }
+
+    l = list->free.next;
+    while(l != &list->free) {
+	DRMLISTDEL(l);
+	node = DRMLISTENTRY(drmBONode, l, head);
+	free(node);
+	l = list->free.next;
+	list->numCurrent--;
+    }
+}
+
+static int drmAdjustListNodes(drmBOList *list)
+{
+    drmBONode *node;
+    drmMMListHead *l;
+    int ret = 0;
+
+    while(list->numCurrent < list->numTarget) {
+	node = (drmBONode *) malloc(sizeof(*node));
+	if (!node) {
+	    ret = -ENOMEM;
+	    break;
+	}
+	list->numCurrent++;
+	DRMLISTADD(&node->head, &list->free);
+    }
+
+    while(list->numCurrent > list->numTarget) {
+	l = list->free.next;
+	if (l == &list->free)
+	    break;
+	DRMLISTDEL(l);
+	node = DRMLISTENTRY(drmBONode, l, head);
+	free(node);
+	list->numCurrent--;
+    }
+    return ret;
+}
+
+static int drmBOCreateList(int numTarget, drmBOList *list)
+{
+    DRMINITLISTHEAD(&list->list);
+    DRMINITLISTHEAD(&list->free);
+    list->numTarget = numTarget;
+    list->numCurrent = 0;
+    list->numOnList = 0;
+    return drmAdjustListNodes(list);
+}
+
+static int drmBOResetList(drmBOList *list)
+{
+    drmMMListHead *l;
+    int ret;
+
+    ret = drmAdjustListNodes(list);
+    if (ret)
+	return ret;
+
+    l = list->list.next;
+    while (l != &list->list) {
+	DRMLISTDEL(l);
+	DRMLISTADD(l, &list->free);
+	list->numOnList--;
+	l = list->list.next;
+    }
+    return drmAdjustListNodes(list);
+}
+
+void driWriteLockKernelBO(void)
+{
+    _glthread_LOCK_MUTEX(bmMutex);
+    while(kernelReaders != 0)
+	_glthread_COND_WAIT(bmCond, bmMutex);
+}
+    
+void driWriteUnlockKernelBO(void)
+{
+    _glthread_UNLOCK_MUTEX(bmMutex);
+}    
+
+void driReadLockKernelBO(void)
+{
+    _glthread_LOCK_MUTEX(bmMutex);
+    kernelReaders++;
+    _glthread_UNLOCK_MUTEX(bmMutex);
+}
+
+void driReadUnlockKernelBO(void)
+{
+    _glthread_LOCK_MUTEX(bmMutex);
+    if (--kernelReaders == 0)
+        _glthread_COND_BROADCAST(bmCond);
+    _glthread_UNLOCK_MUTEX(bmMutex);
+}
+
+
+
+
+/*
+ * TODO: Introduce fence pools in the same way as 
+ * buffer object pools.
+ */
+
+typedef struct _DriBufferObject
+{
+   DriBufferPool *pool;
+   _glthread_Mutex mutex;
+   int refCount;
+   const char *name;
+   uint64_t flags;
+   unsigned hint;
+   unsigned alignment;
+   unsigned createdByReference;
+   void *private;
+   /* user-space buffer: */
+   unsigned userBuffer;
+   void *userData;
+   unsigned userSize;
+} DriBufferObject;
+
+typedef struct _DriBufferList {
+    drmBOList drmBuffers;  /* List of kernel buffers needing validation */
+    drmBOList driBuffers;  /* List of user-space buffers needing validation */
+} DriBufferList;
+
+
+void
+bmError(int val, const char *file, const char *function, int line)
+{
+   _mesa_printf("Fatal video memory manager error \"%s\".\n"
+                "Check kernel logs or set the LIBGL_DEBUG\n"
+                "environment variable to \"verbose\" for more info.\n"
+                "Detected in file %s, line %d, function %s.\n",
+                strerror(-val), file, line, function);
+#ifndef NDEBUG
+   abort();
+#else
+   abort();
+#endif
+}
+
+extern drmBO *
+driBOKernel(struct _DriBufferObject *buf)
+{
+   drmBO *ret;
+
+   driReadLockKernelBO();
+   _glthread_LOCK_MUTEX(buf->mutex);
+   assert(buf->private != NULL);
+   ret = buf->pool->kernel(buf->pool, buf->private);
+   if (!ret)
+      BM_CKFATAL(-EINVAL);
+   _glthread_UNLOCK_MUTEX(buf->mutex);
+   driReadUnlockKernelBO();
+
+   return ret;
+}
+
+void
+driBOWaitIdle(struct _DriBufferObject *buf, int lazy)
+{
+
+  /*
+   * This function may block. Is it sane to keep the mutex held during
+   * that time??
+   */
+
+   _glthread_LOCK_MUTEX(buf->mutex);
+   BM_CKFATAL(buf->pool->waitIdle(buf->pool, buf->private, &buf->mutex, lazy));
+   _glthread_UNLOCK_MUTEX(buf->mutex);
+}
+
+void *
+driBOMap(struct _DriBufferObject *buf, unsigned flags, unsigned hint)
+{
+   void *virtual;
+   int retval;
+
+   if (buf->userBuffer) {
+      return buf->userData;
+   }
+
+   _glthread_LOCK_MUTEX(buf->mutex);
+   assert(buf->private != NULL);
+   retval = buf->pool->map(buf->pool, buf->private, flags, hint, 
+			   &buf->mutex, &virtual);
+   _glthread_UNLOCK_MUTEX(buf->mutex);
+
+   return retval == 0 ? virtual : NULL;
+}
+
+void
+driBOUnmap(struct _DriBufferObject *buf)
+{
+   if (buf->userBuffer)
+      return;
+
+   assert(buf->private != NULL);
+   _glthread_LOCK_MUTEX(buf->mutex);
+   BM_CKFATAL(buf->pool->unmap(buf->pool, buf->private));
+   _glthread_UNLOCK_MUTEX(buf->mutex);
+}
+
+unsigned long
+driBOOffset(struct _DriBufferObject *buf)
+{
+   unsigned long ret;
+
+   assert(buf->private != NULL);
+
+   _glthread_LOCK_MUTEX(buf->mutex);
+   ret = buf->pool->offset(buf->pool, buf->private);
+   _glthread_UNLOCK_MUTEX(buf->mutex);
+   return ret;
+}
+
+unsigned long
+driBOPoolOffset(struct _DriBufferObject *buf)
+{
+   unsigned long ret;
+
+   assert(buf->private != NULL);
+
+   _glthread_LOCK_MUTEX(buf->mutex);
+   ret = buf->pool->poolOffset(buf->pool, buf->private);
+   _glthread_UNLOCK_MUTEX(buf->mutex);
+   return ret;
+}
+
+uint64_t 
+driBOFlags(struct _DriBufferObject *buf)
+{
+   uint64_t ret;
+
+   assert(buf->private != NULL);
+
+   driReadLockKernelBO();
+   _glthread_LOCK_MUTEX(buf->mutex);
+   ret = buf->pool->flags(buf->pool, buf->private);
+   _glthread_UNLOCK_MUTEX(buf->mutex);
+   driReadUnlockKernelBO();
+   return ret;
+}
+
+struct _DriBufferObject *
+driBOReference(struct _DriBufferObject *buf)
+{
+   _glthread_LOCK_MUTEX(buf->mutex);
+   if (++buf->refCount == 1) {
+      _glthread_UNLOCK_MUTEX(buf->mutex);
+      BM_CKFATAL(-EINVAL);
+   }
+   _glthread_UNLOCK_MUTEX(buf->mutex);
+   return buf;
+}
+
+void
+driBOUnReference(struct _DriBufferObject *buf)
+{
+   int tmp;
+
+   if (!buf)
+      return;
+
+
+   _glthread_LOCK_MUTEX(buf->mutex);
+   tmp = --buf->refCount;
+   if (!tmp) {
+      _glthread_UNLOCK_MUTEX(buf->mutex);
+      if (buf->private) {
+	 if (buf->createdByReference)
+	    buf->pool->unreference(buf->pool, buf->private);
+	 else
+	    buf->pool->destroy(buf->pool, buf->private);
+      }
+      free(buf);
+   } else 
+     _glthread_UNLOCK_MUTEX(buf->mutex);
+
+}
+
+
+int
+driBOData(struct _DriBufferObject *buf,
+          unsigned size, const void *data, 
+	  DriBufferPool *newPool, 
+	  uint64_t flags)
+{
+   void *virtual = NULL;
+   int newBuffer;
+   int retval = 0;
+   struct _DriBufferPool *pool;
+
+   assert(!buf->userBuffer); /* XXX just do a memcpy? */
+
+   _glthread_LOCK_MUTEX(buf->mutex);
+   pool = buf->pool;
+
+   if (pool == NULL && newPool != NULL) {
+       buf->pool = newPool;
+       pool = newPool;
+   }
+   if (newPool == NULL)
+       newPool = pool;
+
+   if (!pool->create) {
+      _mesa_error(NULL, GL_INVALID_OPERATION,
+                  "driBOData called on invalid buffer\n");
+      BM_CKFATAL(-EINVAL);
+   }
+
+   newBuffer = (!buf->private || pool != newPool ||
+ 		pool->size(pool, buf->private) < size);
+
+   if (!flags)
+       flags = buf->flags;
+
+   if (newBuffer) {
+
+       if (buf->createdByReference) {
+	  _mesa_error(NULL, GL_INVALID_OPERATION,
+		     "driBOData requiring resizing called on "
+		     "shared buffer.\n");
+	  BM_CKFATAL(-EINVAL);
+       }
+
+       if (buf->private)
+	   buf->pool->destroy(buf->pool, buf->private);
+
+       pool = newPool;
+       buf->pool = newPool;
+       buf->private = pool->create(pool, size, flags, DRM_BO_HINT_DONT_FENCE,
+				  buf->alignment);
+      if (!buf->private)
+	  retval = -ENOMEM;
+
+      if (retval == 0)
+	  retval = pool->map(pool, buf->private,
+			     DRM_BO_FLAG_WRITE,
+			     DRM_BO_HINT_DONT_BLOCK, &buf->mutex, &virtual);
+   } else if (pool->map(pool, buf->private, DRM_BO_FLAG_WRITE,
+			DRM_BO_HINT_DONT_BLOCK, &buf->mutex, &virtual)) {
+       /*
+	* Buffer is busy. need to create a new one.
+	*/
+
+       void *newBuf;
+
+       newBuf = pool->create(pool, size, flags, DRM_BO_HINT_DONT_FENCE,
+			     buf->alignment);
+       if (newBuf) {
+	   buf->pool->destroy(buf->pool, buf->private);
+	   buf->private = newBuf;
+       }
+
+       retval = pool->map(pool, buf->private,
+			  DRM_BO_FLAG_WRITE, 0, &buf->mutex, &virtual);
+   } else {
+       uint64_t flag_diff = flags ^ buf->flags;
+       
+       /*
+	* We might need to change buffer flags.
+	*/
+
+       if (flag_diff){
+	   assert(pool->setStatus != NULL);
+	   BM_CKFATAL(pool->unmap(pool, buf->private));
+	   BM_CKFATAL(pool->setStatus(pool, buf->private, flag_diff,
+				      buf->flags));
+	   if (!data)
+	     goto out;
+
+	   retval = pool->map(pool, buf->private,
+			      DRM_BO_FLAG_WRITE, 0, &buf->mutex, &virtual);
+       }
+   }
+
+   if (retval == 0) {
+      if (data)
+	 memcpy(virtual, data, size);
+
+      BM_CKFATAL(pool->unmap(pool, buf->private));
+   }
+
+ out:
+   _glthread_UNLOCK_MUTEX(buf->mutex);
+   
+   return retval;
+}
+
+void
+driBOSubData(struct _DriBufferObject *buf,
+             unsigned long offset, unsigned long size, const void *data)
+{
+   void *virtual;
+
+   assert(!buf->userBuffer); /* XXX just do a memcpy? */
+
+   _glthread_LOCK_MUTEX(buf->mutex);
+   if (size && data) {
+      BM_CKFATAL(buf->pool->map(buf->pool, buf->private,
+                                DRM_BO_FLAG_WRITE, 0, &buf->mutex,
+				&virtual));
+      memcpy((unsigned char *) virtual + offset, data, size);
+      BM_CKFATAL(buf->pool->unmap(buf->pool, buf->private));
+   }
+   _glthread_UNLOCK_MUTEX(buf->mutex);
+}
+
+void
+driBOGetSubData(struct _DriBufferObject *buf,
+                unsigned long offset, unsigned long size, void *data)
+{
+   void *virtual;
+
+   assert(!buf->userBuffer); /* XXX just do a memcpy? */
+
+   _glthread_LOCK_MUTEX(buf->mutex);
+   if (size && data) {
+      BM_CKFATAL(buf->pool->map(buf->pool, buf->private,
+                                DRM_BO_FLAG_READ, 0, &buf->mutex, &virtual));
+      memcpy(data, (unsigned char *) virtual + offset, size);
+      BM_CKFATAL(buf->pool->unmap(buf->pool, buf->private));
+   }
+   _glthread_UNLOCK_MUTEX(buf->mutex);
+}
+
+void
+driBOSetReferenced(struct _DriBufferObject *buf,
+		   unsigned long handle)
+{
+   _glthread_LOCK_MUTEX(buf->mutex);
+   if (buf->private != NULL) {
+      _mesa_error(NULL, GL_INVALID_OPERATION,
+                  "Invalid buffer for setReferenced\n");
+      BM_CKFATAL(-EINVAL);
+   
+   }
+   if (buf->pool->reference == NULL) {
+      _mesa_error(NULL, GL_INVALID_OPERATION,
+                  "Invalid buffer pool for setReferenced\n");
+      BM_CKFATAL(-EINVAL);
+   }
+   buf->private = buf->pool->reference(buf->pool, handle);
+   if (!buf->private) {
+      _mesa_error(NULL, GL_OUT_OF_MEMORY,
+                  "Invalid buffer pool for setStatic\n");
+      BM_CKFATAL(-ENOMEM);
+   }
+   buf->createdByReference = GL_TRUE;
+   buf->flags = buf->pool->kernel(buf->pool, buf->private)->flags;
+   _glthread_UNLOCK_MUTEX(buf->mutex);
+}
+
+int
+driGenBuffers(struct _DriBufferPool *pool,
+              const char *name,
+              unsigned n,
+              struct _DriBufferObject *buffers[],
+              unsigned alignment, uint64_t flags, unsigned hint)
+{
+   struct _DriBufferObject *buf;
+   int i;
+
+   flags = (flags) ? flags : DRM_BO_FLAG_MEM_TT | DRM_BO_FLAG_MEM_VRAM |
+      DRM_BO_FLAG_MEM_LOCAL | DRM_BO_FLAG_READ | DRM_BO_FLAG_WRITE;
+
+
+   assert(pool);
+
+   for (i = 0; i < n; ++i) {
+      buf = (struct _DriBufferObject *) calloc(1, sizeof(*buf));
+      if (!buf)
+	 return -ENOMEM;
+
+      _glthread_INIT_MUTEX(buf->mutex);
+      _glthread_LOCK_MUTEX(buf->mutex);
+      buf->refCount = 1;
+      buf->flags = flags;
+      buf->hint = hint;
+      buf->name = name;
+      buf->alignment = alignment;
+      buf->pool = pool;
+      buf->createdByReference = 0;
+      _glthread_UNLOCK_MUTEX(buf->mutex);
+      buffers[i] = buf;
+   }
+   return 0;
+}
+
+void
+driGenUserBuffer(struct _DriBufferPool *pool,
+                 const char *name,
+                 struct _DriBufferObject **buffers,
+                 void *ptr, unsigned bytes)
+{
+   const unsigned alignment = 1, flags = 0, hint = 0;
+
+   driGenBuffers(pool, name, 1, buffers, alignment, flags, hint);
+
+   (*buffers)->userBuffer = 1;
+   (*buffers)->userData = ptr;
+   (*buffers)->userSize = bytes;
+}
+
+void
+driDeleteBuffers(unsigned n, struct _DriBufferObject *buffers[])
+{
+   int i;
+
+   for (i = 0; i < n; ++i) {
+      driBOUnReference(buffers[i]);
+   }
+}
+
+
+void
+driInitBufMgr(int fd)
+{
+   ;
+}
+
+/*
+ * Note that lists are per-context and don't need mutex protection.
+ */
+
+struct _DriBufferList *
+driBOCreateList(int target)
+{
+    struct _DriBufferList *list = calloc(sizeof(*list), 1);
+
+    BM_CKFATAL(drmBOCreateList(target, &list->drmBuffers));
+    BM_CKFATAL(drmBOCreateList(target, &list->driBuffers));
+    return list;
+}
+
+int
+driBOResetList(struct _DriBufferList * list)
+{
+    int ret;
+    ret = drmBOResetList(&list->drmBuffers);
+    if (ret)
+	return ret;
+    ret = drmBOResetList(&list->driBuffers);
+    return ret;
+}
+
+void
+driBOFreeList(struct _DriBufferList * list)
+{
+   drmBOFreeList(&list->drmBuffers);
+   drmBOFreeList(&list->driBuffers);
+   free(list);
+}
+
+
+/*
+ * Copied from libdrm, because it is needed by driAddValidateItem.
+ */
+
+static drmBONode *
+driAddListItem(drmBOList * list, drmBO * item,
+	       uint64_t arg0, uint64_t arg1)
+{
+    drmBONode *node;
+    drmMMListHead *l;
+
+    l = list->free.next;
+    if (l == &list->free) {
+	node = (drmBONode *) malloc(sizeof(*node));
+	if (!node) {
+	    return NULL;
+	}
+	list->numCurrent++;
+    } else {
+	DRMLISTDEL(l);
+	node = DRMLISTENTRY(drmBONode, l, head);
+    }
+    memset(&node->bo_arg, 0, sizeof(node->bo_arg));
+    node->buf = item;
+    node->arg0 = arg0;
+    node->arg1 = arg1;
+    DRMLISTADDTAIL(&node->head, &list->list);
+    list->numOnList++;
+    return node;
+}
+
+/*
+ * Slightly modified version compared to the libdrm version.
+ * This one returns the list index of the buffer put on the list.
+ */
+
+static int
+driAddValidateItem(drmBOList * list, drmBO * buf, uint64_t flags,
+		   uint64_t mask, int *itemLoc, 
+		   struct _drmBONode **pnode)
+{
+    drmBONode *node, *cur;
+    drmMMListHead *l;
+    int count = 0;
+
+    cur = NULL;
+
+    for (l = list->list.next; l != &list->list; l = l->next) {
+	node = DRMLISTENTRY(drmBONode, l, head);
+	if (node->buf == buf) {
+	    cur = node;
+	    break;
+	}
+	count++;
+    }
+    if (!cur) {
+	cur = driAddListItem(list, buf, flags, mask);
+	if (!cur)
+	    return -ENOMEM;
+
+	cur->arg0 = flags;
+	cur->arg1 = mask;
+    } else {
+        uint64_t memFlags = cur->arg0 & flags & DRM_BO_MASK_MEM;
+	uint64_t accFlags = (cur->arg0 | flags) & ~DRM_BO_MASK_MEM;
+
+	if (mask & cur->arg1 & ~DRM_BO_MASK_MEM  & (cur->arg0 ^ flags)) {
+	    return -EINVAL;
+	}
+
+	cur->arg1 |= mask;
+	cur->arg0 = (cur->arg0 & ~mask) | ((memFlags | accFlags) & mask);
+
+	if (((cur->arg1 & DRM_BO_MASK_MEM) != 0) &&
+	    (cur->arg0 & DRM_BO_MASK_MEM) == 0) {
+	    return -EINVAL;
+	}
+    }
+    *itemLoc = count;
+    *pnode = cur;
+    return 0;
+}
+
+
+void
+driBOAddListItem(struct _DriBufferList * list, struct _DriBufferObject *buf,
+                 uint64_t flags, uint64_t mask, int *itemLoc, 
+		 struct _drmBONode **node)
+{
+   int newItem;
+   
+   _glthread_LOCK_MUTEX(buf->mutex);
+   BM_CKFATAL(driAddValidateItem(&list->drmBuffers,
+				 buf->pool->kernel(buf->pool, buf->private),
+                                 flags, mask, itemLoc, node));
+   BM_CKFATAL(drmAddValidateItem(&list->driBuffers, (drmBO *) buf,
+				 flags, mask, &newItem));
+   if (newItem) 
+     buf->refCount++;
+
+   _glthread_UNLOCK_MUTEX(buf->mutex);
+}
+
+drmBOList *driGetdrmBOList(struct _DriBufferList *list)
+{
+	driWriteLockKernelBO();
+	return &list->drmBuffers;
+}
+
+void driPutdrmBOList(struct _DriBufferList *list)
+{
+	driWriteUnlockKernelBO();
+}
+
+
+void
+driBOFence(struct _DriBufferObject *buf, struct _DriFenceObject *fence)
+{
+   _glthread_LOCK_MUTEX(buf->mutex);
+   if (buf->pool->fence)
+       BM_CKFATAL(buf->pool->fence(buf->pool, buf->private, fence));
+   _glthread_UNLOCK_MUTEX(buf->mutex);
+
+}
+
+void
+driBOUnrefUserList(struct _DriBufferList *list)
+{
+    struct _DriBufferObject *buf;
+    void *curBuf;
+
+    curBuf = drmBOListIterator(&list->driBuffers);
+    while (curBuf) {
+	buf = (struct _DriBufferObject *)drmBOListBuf(curBuf);
+	driBOUnReference(buf);
+	curBuf = drmBOListNext(&list->driBuffers, curBuf);
+    }
+}
+
+struct _DriFenceObject *
+driBOFenceUserList(struct _DriFenceMgr *mgr,
+		   struct _DriBufferList *list, const char *name,
+		   drmFence *kFence)
+{
+    struct _DriFenceObject *fence;
+    struct _DriBufferObject *buf;
+    void *curBuf;
+
+    fence = driFenceCreate(mgr, kFence->fence_class, kFence->type,
+			   kFence, sizeof(*kFence));
+    curBuf = drmBOListIterator(&list->driBuffers);
+
+   /*
+    * User-space fencing callbacks.
+    */
+
+   while (curBuf) {
+        buf = (struct _DriBufferObject *) drmBOListBuf(curBuf);
+	driBOFence(buf, fence);
+	driBOUnReference(buf);
+	curBuf = drmBOListNext(&list->driBuffers, curBuf);
+   }
+
+   driBOResetList(list);
+   return fence;
+}
+   
+void
+driBOValidateUserList(struct _DriBufferList * list)
+{
+    void *curBuf;
+    struct _DriBufferObject *buf;
+
+    curBuf = drmBOListIterator(&list->driBuffers);
+
+    /*
+     * User-space validation callbacks.
+     */
+
+    while (curBuf) {
+	buf = (struct _DriBufferObject *) drmBOListBuf(curBuf);
+	_glthread_LOCK_MUTEX(buf->mutex);
+	if (buf->pool->validate)
+	    BM_CKFATAL(buf->pool->validate(buf->pool, buf->private, &buf->mutex));
+	_glthread_UNLOCK_MUTEX(buf->mutex);
+	curBuf = drmBOListNext(&list->driBuffers, curBuf);
+    }
+}
+
+
+void
+driPoolTakeDown(struct _DriBufferPool *pool)
+{
+   pool->takeDown(pool);
+
+}
+
+unsigned long 
+driBOSize(struct _DriBufferObject *buf)
+{
+  unsigned long size;
+
+   _glthread_LOCK_MUTEX(buf->mutex);
+   size = buf->pool->size(buf->pool, buf->private);
+   _glthread_UNLOCK_MUTEX(buf->mutex);
+
+  return size;
+
+}
+
+drmBOList *driBOGetDRMBuffers(struct _DriBufferList *list)
+{
+    return &list->drmBuffers;
+}
+
+drmBOList *driBOGetDRIBuffers(struct _DriBufferList *list)
+{
+    return &list->driBuffers;
+}
+    
diff --git a/src/gallium/winsys/dri/intel/ws_dri_bufmgr.h b/src/gallium/winsys/dri/intel/ws_dri_bufmgr.h
new file mode 100644
index 00000000000..fdaf5ee93ac
--- /dev/null
+++ b/src/gallium/winsys/dri/intel/ws_dri_bufmgr.h
@@ -0,0 +1,138 @@
+/**************************************************************************
+ * 
+ * Copyright 2006 Tungsten Graphics, Inc., Bismarck, ND., USA
+ * 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 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
+ * THE COPYRIGHT HOLDERS, AUTHORS 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.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ * 
+ * 
+ **************************************************************************/
+/*
+ * Authors: Thomas Hellstr�m <thomas-at-tungstengraphics-dot-com>
+ *          Keith Whitwell <keithw-at-tungstengraphics-dot-com>
+ */
+
+#ifndef _PSB_BUFMGR_H_
+#define _PSB_BUFMGR_H_
+#include <xf86mm.h>
+#include "i915_drm.h"
+#include "ws_dri_fencemgr.h"
+
+typedef struct _drmBONode
+{
+    drmMMListHead head;
+    drmBO *buf;
+    struct drm_i915_op_arg bo_arg;
+    uint64_t arg0;
+    uint64_t arg1;
+} drmBONode;
+
+typedef struct _drmBOList {
+    unsigned numTarget;
+    unsigned numCurrent;
+    unsigned numOnList;
+    drmMMListHead list;
+    drmMMListHead free;
+} drmBOList;
+
+
+struct _DriFenceObject;
+struct _DriBufferObject;
+struct _DriBufferPool;
+struct _DriBufferList;
+
+/*
+ * Return a pointer to the libdrm buffer object this DriBufferObject
+ * uses.
+ */
+
+extern drmBO *driBOKernel(struct _DriBufferObject *buf);
+extern void *driBOMap(struct _DriBufferObject *buf, unsigned flags,
+                      unsigned hint);
+extern void driBOUnmap(struct _DriBufferObject *buf);
+extern unsigned long driBOOffset(struct _DriBufferObject *buf);
+extern unsigned long driBOPoolOffset(struct _DriBufferObject *buf);
+
+extern uint64_t driBOFlags(struct _DriBufferObject *buf);
+extern struct _DriBufferObject *driBOReference(struct _DriBufferObject *buf);
+extern void driBOUnReference(struct _DriBufferObject *buf);
+
+extern int driBOData(struct _DriBufferObject *r_buf,
+		     unsigned size, const void *data, 
+		     struct _DriBufferPool *pool, uint64_t flags);
+
+extern void driBOSubData(struct _DriBufferObject *buf,
+                         unsigned long offset, unsigned long size,
+                         const void *data);
+extern void driBOGetSubData(struct _DriBufferObject *buf,
+                            unsigned long offset, unsigned long size,
+                            void *data);
+extern int driGenBuffers(struct _DriBufferPool *pool,
+			 const char *name,
+			 unsigned n,
+			 struct _DriBufferObject *buffers[],
+			 unsigned alignment, uint64_t flags, unsigned hint);
+extern void driGenUserBuffer(struct _DriBufferPool *pool,
+                             const char *name,
+                             struct _DriBufferObject *buffers[],
+                             void *ptr, unsigned bytes);
+extern void driDeleteBuffers(unsigned n, struct _DriBufferObject *buffers[]);
+extern void driInitBufMgr(int fd);
+extern struct _DriBufferList *driBOCreateList(int target);
+extern int driBOResetList(struct _DriBufferList * list);
+extern void driBOAddListItem(struct _DriBufferList * list, 
+			     struct _DriBufferObject *buf,
+                             uint64_t flags, uint64_t mask, int *itemLoc,
+			     struct _drmBONode **node);
+
+extern void driBOValidateList(int fd, struct _DriBufferList * list);
+extern void driBOFreeList(struct _DriBufferList * list);
+extern struct _DriFenceObject *driBOFenceUserList(struct _DriFenceMgr *mgr,
+						  struct _DriBufferList *list,
+						  const char *name,
+						  drmFence *kFence);
+extern void driBOUnrefUserList(struct _DriBufferList *list);
+extern void driBOValidateUserList(struct _DriBufferList * list);
+extern drmBOList *driGetdrmBOList(struct _DriBufferList *list);
+extern void driPutdrmBOList(struct _DriBufferList *list);
+
+extern void driBOFence(struct _DriBufferObject *buf,
+                       struct _DriFenceObject *fence);
+
+extern void driPoolTakeDown(struct _DriBufferPool *pool);
+extern void driBOSetReferenced(struct _DriBufferObject *buf,
+			       unsigned long handle);
+unsigned long driBOSize(struct _DriBufferObject *buf);
+extern void driBOWaitIdle(struct _DriBufferObject *buf, int lazy);
+extern void driPoolTakeDown(struct _DriBufferPool *pool);
+
+extern void driReadLockKernelBO(void);
+extern void driReadUnlockKernelBO(void);
+extern void driWriteLockKernelBO(void);
+extern void driWriteUnlockKernelBO(void);
+
+/*
+ * For debugging purposes.
+ */
+
+extern drmBOList *driBOGetDRMBuffers(struct _DriBufferList *list);
+extern drmBOList *driBOGetDRIBuffers(struct _DriBufferList *list);
+#endif
diff --git a/src/gallium/winsys/dri/intel/ws_dri_bufpool.h b/src/gallium/winsys/dri/intel/ws_dri_bufpool.h
new file mode 100644
index 00000000000..3a302e13d3b
--- /dev/null
+++ b/src/gallium/winsys/dri/intel/ws_dri_bufpool.h
@@ -0,0 +1,102 @@
+/**************************************************************************
+ * 
+ * Copyright 2006 Tungsten Graphics, Inc., Bismarck, ND., USA
+ * 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 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
+ * THE COPYRIGHT HOLDERS, AUTHORS 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.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ * 
+ * 
+ **************************************************************************/
+/*
+ * Authors: Thomas Hellstr�m <thomas-at-tungstengraphics-dot-com>
+ */
+
+#ifndef _PSB_BUFPOOL_H_
+#define _PSB_BUFPOOL_H_
+
+#include <xf86drm.h>
+#include <glthread.h>
+struct _DriFenceObject;
+
+typedef struct _DriBufferPool
+{
+   int fd;
+   int (*map) (struct _DriBufferPool * pool, void *private,
+               unsigned flags, int hint, _glthread_Mutex *mutex, 
+	       void **virtual);
+   int (*unmap) (struct _DriBufferPool * pool, void *private);
+   int (*destroy) (struct _DriBufferPool * pool, void *private);
+   unsigned long (*offset) (struct _DriBufferPool * pool, void *private);
+   unsigned long (*poolOffset) (struct _DriBufferPool * pool, void *private);
+   uint64_t (*flags) (struct _DriBufferPool * pool, void *private);
+   unsigned long (*size) (struct _DriBufferPool * pool, void *private);
+   void *(*create) (struct _DriBufferPool * pool, unsigned long size,
+                    uint64_t flags, unsigned hint, unsigned alignment);
+   void *(*reference) (struct _DriBufferPool * pool, unsigned handle);
+   int (*unreference) (struct _DriBufferPool * pool, void *private);
+   int (*fence) (struct _DriBufferPool * pool, void *private,
+                 struct _DriFenceObject * fence);
+   drmBO *(*kernel) (struct _DriBufferPool * pool, void *private);
+   int (*validate) (struct _DriBufferPool * pool, void *private, _glthread_Mutex *mutex);
+   int (*waitIdle) (struct _DriBufferPool *pool, void *private, _glthread_Mutex *mutex,
+		    int lazy);
+   int (*setStatus)  (struct _DriBufferPool *pool, void *private,
+		      uint64_t flag_diff, uint64_t old_flags);
+   void (*takeDown) (struct _DriBufferPool * pool);
+   void *data;
+} DriBufferPool;
+
+extern void bmError(int val, const char *file, const char *function,
+                    int line);
+#define BM_CKFATAL(val)					       \
+  do{							       \
+    int tstVal = (val);					       \
+    if (tstVal) 					       \
+      bmError(tstVal, __FILE__, __FUNCTION__, __LINE__);       \
+  } while(0);
+
+
+/*
+ * Builtin pools.
+ */
+
+/*
+ * Kernel buffer objects. Size in multiples of page size. Page size aligned.
+ */
+
+extern struct _DriBufferPool *driDRMPoolInit(int fd);
+extern struct _DriBufferPool *driMallocPoolInit(void);
+
+struct _DriFreeSlabManager;
+extern struct _DriBufferPool * driSlabPoolInit(int fd, uint64_t flags,
+					       uint64_t validMask,
+					       uint32_t smallestSize,
+					       uint32_t numSizes,
+					       uint32_t desiredNumBuffers,
+					       uint32_t maxSlabSize,
+					       uint32_t pageAlignment,
+					       struct _DriFreeSlabManager *fMan);
+extern void driFinishFreeSlabManager(struct _DriFreeSlabManager *fMan);
+extern struct _DriFreeSlabManager *
+driInitFreeSlabManager(uint32_t checkIntervalMsec, uint32_t slabTimeoutMsec);
+
+
+#endif
diff --git a/src/gallium/winsys/dri/intel/ws_dri_drmpool.c b/src/gallium/winsys/dri/intel/ws_dri_drmpool.c
new file mode 100644
index 00000000000..7c55dbc6740
--- /dev/null
+++ b/src/gallium/winsys/dri/intel/ws_dri_drmpool.c
@@ -0,0 +1,268 @@
+/**************************************************************************
+ * 
+ * Copyright 2006 Tungsten Graphics, Inc., Bismarck, ND., USA
+ * 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 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
+ * THE COPYRIGHT HOLDERS, AUTHORS 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.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ * 
+ * 
+ **************************************************************************/
+/*
+ * Authors: Thomas Hellstr�m <thomas-at-tungstengraphics-dot-com>
+ */
+
+#include <xf86drm.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include "ws_dri_bufpool.h"
+#include "ws_dri_bufmgr.h"
+#include "assert.h"
+
+/*
+ * Buffer pool implementation using DRM buffer objects as DRI buffer objects.
+ */
+
+static void *
+pool_create(struct _DriBufferPool *pool,
+            unsigned long size, uint64_t flags, unsigned hint,
+            unsigned alignment)
+{
+   drmBO *buf = (drmBO *) malloc(sizeof(*buf));
+   int ret;
+   unsigned pageSize = getpagesize();
+
+   if (!buf)
+      return NULL;
+
+   if ((alignment > pageSize) && (alignment % pageSize)) {
+      free(buf);
+      return NULL;
+   }
+
+   ret = drmBOCreate(pool->fd, size, alignment / pageSize,
+		     NULL,
+                     flags, hint, buf);
+   if (ret) {
+      free(buf);
+      return NULL;
+   }
+
+   return (void *) buf;
+}
+
+static void *
+pool_reference(struct _DriBufferPool *pool, unsigned handle)
+{
+   drmBO *buf = (drmBO *) malloc(sizeof(*buf));
+   int ret;
+
+   if (!buf)
+      return NULL;
+
+   ret = drmBOReference(pool->fd, handle, buf);
+
+   if (ret) {
+      free(buf);
+      return NULL;
+   }
+
+   return (void *) buf;
+}
+
+static int
+pool_destroy(struct _DriBufferPool *pool, void *private)
+{
+   int ret;
+   drmBO *buf = (drmBO *) private;
+   driReadLockKernelBO();
+   ret = drmBOUnreference(pool->fd, buf);
+   free(buf);
+   driReadUnlockKernelBO();
+   return ret;
+}
+
+static int
+pool_unreference(struct _DriBufferPool *pool, void *private)
+{
+   int ret;
+   drmBO *buf = (drmBO *) private;
+   driReadLockKernelBO();
+   ret = drmBOUnreference(pool->fd, buf);
+   free(buf);
+   driReadUnlockKernelBO();
+   return ret;
+}
+
+static int
+pool_map(struct _DriBufferPool *pool, void *private, unsigned flags,
+         int hint, _glthread_Mutex *mutex, void **virtual)
+{
+   drmBO *buf = (drmBO *) private;
+   int ret;
+   
+   driReadLockKernelBO();
+   ret = drmBOMap(pool->fd, buf, flags, hint, virtual);
+   driReadUnlockKernelBO();
+   return ret;
+}
+
+static int
+pool_unmap(struct _DriBufferPool *pool, void *private)
+{
+   drmBO *buf = (drmBO *) private;
+   int ret;
+
+   driReadLockKernelBO();
+   ret = drmBOUnmap(pool->fd, buf);
+   driReadUnlockKernelBO();
+   
+   return ret;
+}
+
+static unsigned long
+pool_offset(struct _DriBufferPool *pool, void *private)
+{
+   drmBO *buf = (drmBO *) private;
+   unsigned long offset;
+
+   driReadLockKernelBO();
+   assert(buf->flags & DRM_BO_FLAG_NO_MOVE);   
+   offset = buf->offset;
+   driReadUnlockKernelBO();
+
+   return buf->offset;
+}
+
+static unsigned long
+pool_poolOffset(struct _DriBufferPool *pool, void *private)
+{
+   return 0;
+}
+
+static uint64_t
+pool_flags(struct _DriBufferPool *pool, void *private)
+{
+   drmBO *buf = (drmBO *) private;
+   uint64_t flags;
+
+   driReadLockKernelBO();
+   flags = buf->flags;
+   driReadUnlockKernelBO();
+
+   return flags;
+}
+
+
+static unsigned long
+pool_size(struct _DriBufferPool *pool, void *private)
+{
+   drmBO *buf = (drmBO *) private;
+   unsigned long size;
+
+   driReadLockKernelBO();
+   size = buf->size;
+   driReadUnlockKernelBO();
+
+   return buf->size;
+}
+
+static int
+pool_fence(struct _DriBufferPool *pool, void *private,
+           struct _DriFenceObject *fence)
+{
+   /*
+    * Noop. The kernel handles all fencing.
+    */
+
+   return 0;
+}
+
+static drmBO *
+pool_kernel(struct _DriBufferPool *pool, void *private)
+{
+   return (drmBO *) private;
+}
+
+static int
+pool_waitIdle(struct _DriBufferPool *pool, void *private, _glthread_Mutex *mutex, 
+	      int lazy)
+{
+   drmBO *buf = (drmBO *) private;
+   int ret;
+
+   driReadLockKernelBO();
+   ret = drmBOWaitIdle(pool->fd, buf, (lazy) ? DRM_BO_HINT_WAIT_LAZY:0);
+   driReadUnlockKernelBO();
+
+   return ret;
+}
+
+   
+static void
+pool_takedown(struct _DriBufferPool *pool)
+{
+   free(pool);
+}
+
+/*static int
+pool_setStatus(struct _DriBufferPool *pool, void *private, 
+	       uint64_t flag_diff, uint64_t old_flags)
+{
+   drmBO *buf = (drmBO *) private;
+   uint64_t new_flags = old_flags ^ flag_diff;
+   int ret;
+
+   driReadLockKernelBO();
+   ret = drmBOSetStatus(pool->fd, buf, new_flags, flag_diff,
+			0, 0, 0);
+   driReadUnlockKernelBO();
+   return ret;
+}*/
+
+struct _DriBufferPool *
+driDRMPoolInit(int fd)
+{
+   struct _DriBufferPool *pool;
+
+   pool = (struct _DriBufferPool *) malloc(sizeof(*pool));
+
+   if (!pool)
+      return NULL;
+
+   pool->fd = fd;
+   pool->map = &pool_map;
+   pool->unmap = &pool_unmap;
+   pool->destroy = &pool_destroy;
+   pool->offset = &pool_offset;
+   pool->poolOffset = &pool_poolOffset;
+   pool->flags = &pool_flags;
+   pool->size = &pool_size;
+   pool->create = &pool_create;
+   pool->fence = &pool_fence;
+   pool->kernel = &pool_kernel;
+   pool->validate = NULL;
+   pool->waitIdle = &pool_waitIdle;
+   pool->takeDown = &pool_takedown;
+   pool->reference = &pool_reference;
+   pool->unreference = &pool_unreference;
+   pool->data = NULL;
+   return pool;
+}
diff --git a/src/gallium/winsys/dri/intel/ws_dri_fencemgr.c b/src/gallium/winsys/dri/intel/ws_dri_fencemgr.c
new file mode 100644
index 00000000000..8aaef1c6200
--- /dev/null
+++ b/src/gallium/winsys/dri/intel/ws_dri_fencemgr.c
@@ -0,0 +1,372 @@
+#include "ws_dri_fencemgr.h"
+#include "glthread.h"
+#include <xf86mm.h>
+#include <string.h>
+#include <unistd.h>
+
+/*
+ * Note: Locking order is
+ * _DriFenceObject::mutex
+ * _DriFenceMgr::mutex
+ */
+
+struct _DriFenceMgr {
+    /*
+     * Constant members. Need no mutex protection.
+     */
+    struct _DriFenceMgrCreateInfo info;
+    void *private;
+
+    /*
+     * These members are protected by this->mutex
+     */
+    _glthread_Mutex mutex;
+    int refCount;
+    drmMMListHead *heads;
+};
+
+struct _DriFenceObject {
+
+    /*
+     * These members are constant and need no mutex protection.
+     */
+    struct _DriFenceMgr *mgr;
+    uint32_t fence_class;
+    uint32_t fence_type;
+
+    /*
+     * These members are protected by mgr->mutex.
+     */
+    drmMMListHead head;
+    int refCount;
+
+    /*
+     * These members are protected by this->mutex.
+     */
+    _glthread_Mutex mutex;
+    uint32_t signaled_type;
+    void *private;
+};
+
+uint32_t
+driFenceType(struct _DriFenceObject *fence)
+{
+  return fence->fence_type;
+}
+
+struct _DriFenceMgr *
+driFenceMgrCreate(const struct _DriFenceMgrCreateInfo *info)
+{
+  struct _DriFenceMgr *tmp;
+  uint32_t i;
+
+  tmp = calloc(1, sizeof(*tmp));
+  if (!tmp)
+      return NULL;
+
+  _glthread_INIT_MUTEX(tmp->mutex);
+  _glthread_LOCK_MUTEX(tmp->mutex);
+  tmp->refCount = 1;
+  tmp->info = *info;
+  tmp->heads = calloc(tmp->info.num_classes, sizeof(*tmp->heads));
+  if (!tmp->heads)
+      goto out_err;
+
+  for (i=0; i<tmp->info.num_classes; ++i) {
+      DRMINITLISTHEAD(&tmp->heads[i]);
+  }
+  _glthread_UNLOCK_MUTEX(tmp->mutex);
+  return tmp;
+
+  out_err:
+  if (tmp)
+      free(tmp);
+  return NULL;
+}
+
+static void
+driFenceMgrUnrefUnlock(struct _DriFenceMgr **pMgr)
+{
+    struct _DriFenceMgr *mgr = *pMgr;
+
+    *pMgr = NULL;
+    if (--mgr->refCount == 0)
+	free(mgr);
+    else
+	_glthread_UNLOCK_MUTEX(mgr->mutex);
+}
+
+void
+driFenceMgrUnReference(struct _DriFenceMgr **pMgr)
+{
+    _glthread_LOCK_MUTEX((*pMgr)->mutex);
+    driFenceMgrUnrefUnlock(pMgr);
+}
+
+static void
+driFenceUnReferenceLocked(struct _DriFenceObject **pFence)
+{
+    struct _DriFenceObject *fence = *pFence;
+    struct _DriFenceMgr *mgr = fence->mgr;
+
+    *pFence = NULL;
+    if (--fence->refCount == 0) {
+	DRMLISTDELINIT(&fence->head);
+	if (fence->private)
+	    mgr->info.unreference(mgr, &fence->private);
+	fence->mgr = NULL;
+	--mgr->refCount;
+	free(fence);
+    }
+}
+
+
+static void
+driSignalPreviousFencesLocked(struct _DriFenceMgr *mgr,
+			      drmMMListHead *list,
+			      uint32_t fence_class,
+			      uint32_t fence_type)
+{
+    struct _DriFenceObject *entry;
+    drmMMListHead *prev;
+
+    while(list != &mgr->heads[fence_class]) {
+	entry = DRMLISTENTRY(struct _DriFenceObject, list, head);
+
+	/*
+	 * Up refcount so that entry doesn't disappear from under us
+	 * when we unlock-relock mgr to get the correct locking order.
+	 */
+
+	++entry->refCount;
+	_glthread_UNLOCK_MUTEX(mgr->mutex);
+	_glthread_LOCK_MUTEX(entry->mutex);
+	_glthread_LOCK_MUTEX(mgr->mutex);
+
+	prev = list->prev;
+
+
+
+	if (list->prev == list) {
+
+		/*
+		 * Somebody else removed the entry from the list.
+		 */
+
+		_glthread_UNLOCK_MUTEX(entry->mutex);
+		driFenceUnReferenceLocked(&entry);
+		return;
+	}
+
+	entry->signaled_type |= (fence_type & entry->fence_type);
+	if (entry->signaled_type == entry->fence_type) {
+	    DRMLISTDELINIT(list);
+	    mgr->info.unreference(mgr, &entry->private);
+	}
+	_glthread_UNLOCK_MUTEX(entry->mutex);
+	driFenceUnReferenceLocked(&entry);
+	list = prev;
+    }
+}
+
+
+int
+driFenceFinish(struct _DriFenceObject *fence, uint32_t fence_type,
+	       int lazy_hint)
+{
+    struct _DriFenceMgr *mgr = fence->mgr;
+    int ret = 0;
+
+    _glthread_LOCK_MUTEX(fence->mutex);
+
+    if ((fence->signaled_type & fence_type) == fence_type)
+	goto out0;
+
+    ret = mgr->info.finish(mgr, fence->private, fence_type, lazy_hint);
+    if (ret)
+	goto out0;
+
+    _glthread_LOCK_MUTEX(mgr->mutex);
+    _glthread_UNLOCK_MUTEX(fence->mutex);
+
+    driSignalPreviousFencesLocked(mgr, &fence->head, fence->fence_class,
+				  fence_type);
+    _glthread_UNLOCK_MUTEX(mgr->mutex);
+    return 0;
+
+  out0:
+    _glthread_UNLOCK_MUTEX(fence->mutex);
+    return ret;
+}
+
+uint32_t driFenceSignaledTypeCached(struct _DriFenceObject *fence)
+{
+    uint32_t ret;
+
+    _glthread_LOCK_MUTEX(fence->mutex);
+    ret = fence->signaled_type;
+    _glthread_UNLOCK_MUTEX(fence->mutex);
+
+    return ret;
+}
+
+int
+driFenceSignaledType(struct _DriFenceObject *fence, uint32_t flush_type,
+		 uint32_t *signaled)
+{
+    int ret = 0;
+    struct _DriFenceMgr *mgr;
+
+    _glthread_LOCK_MUTEX(fence->mutex);
+    mgr = fence->mgr;
+    *signaled = fence->signaled_type;
+    if ((fence->signaled_type & flush_type) == flush_type)
+	goto out0;
+
+    ret = mgr->info.signaled(mgr, fence->private, flush_type, signaled);
+    if (ret) {
+	*signaled = fence->signaled_type;
+	goto out0;
+    }
+
+    if ((fence->signaled_type | *signaled) == fence->signaled_type)
+	goto out0;
+
+    _glthread_LOCK_MUTEX(mgr->mutex);
+    _glthread_UNLOCK_MUTEX(fence->mutex);
+
+    driSignalPreviousFencesLocked(mgr, &fence->head, fence->fence_class,
+				  *signaled);
+
+    _glthread_UNLOCK_MUTEX(mgr->mutex);
+    return 0;
+  out0:
+    _glthread_UNLOCK_MUTEX(fence->mutex);
+    return ret;
+}
+
+struct _DriFenceObject *
+driFenceReference(struct _DriFenceObject *fence)
+{
+    _glthread_LOCK_MUTEX(fence->mgr->mutex);
+    ++fence->refCount;
+    _glthread_UNLOCK_MUTEX(fence->mgr->mutex);
+    return fence;
+}
+
+void
+driFenceUnReference(struct _DriFenceObject **pFence)
+{
+    struct _DriFenceMgr *mgr; 
+
+    if (*pFence == NULL)
+	return;
+
+    mgr = (*pFence)->mgr;
+    _glthread_LOCK_MUTEX(mgr->mutex);
+    ++mgr->refCount;
+    driFenceUnReferenceLocked(pFence);
+    driFenceMgrUnrefUnlock(&mgr);
+}
+
+struct _DriFenceObject
+*driFenceCreate(struct _DriFenceMgr *mgr, uint32_t fence_class,
+		uint32_t fence_type, void *private, size_t private_size)
+{
+    struct _DriFenceObject *fence;
+    size_t fence_size = sizeof(*fence);
+
+    if (private_size)
+	fence_size = ((fence_size + 15) & ~15);
+
+    fence = calloc(1, fence_size + private_size);
+
+    if (!fence) {
+	int ret = mgr->info.finish(mgr, private, fence_type, 0);
+
+	if (ret)
+	    usleep(10000000);
+
+	return NULL;
+    }
+
+    _glthread_INIT_MUTEX(fence->mutex);
+    _glthread_LOCK_MUTEX(fence->mutex);
+    _glthread_LOCK_MUTEX(mgr->mutex);
+    fence->refCount = 1;
+    DRMLISTADDTAIL(&fence->head, &mgr->heads[fence_class]);
+    fence->mgr = mgr;
+    ++mgr->refCount;
+    _glthread_UNLOCK_MUTEX(mgr->mutex);
+    fence->fence_class = fence_class;
+    fence->fence_type = fence_type;
+    fence->signaled_type = 0;
+    fence->private = private;
+    if (private_size) {
+        fence->private = (void *)(((uint8_t *) fence) + fence_size);
+	memcpy(fence->private, private, private_size);
+    }
+
+    _glthread_UNLOCK_MUTEX(fence->mutex);
+    return fence;
+}
+
+
+static int
+tSignaled(struct _DriFenceMgr *mgr, void *private, uint32_t flush_type,
+	  uint32_t *signaled_type)
+{
+  long fd = (long) mgr->private;
+  int dummy;
+  drmFence *fence = (drmFence *) private;
+  int ret;
+
+  *signaled_type = 0;
+  ret = drmFenceSignaled((int) fd, fence, flush_type, &dummy);
+  if (ret)
+    return ret;
+
+  *signaled_type = fence->signaled;
+
+  return 0;
+}
+
+static int
+tFinish(struct _DriFenceMgr *mgr, void *private, uint32_t fence_type,
+	int lazy_hint)
+{
+  long fd = (long) mgr->private;
+  unsigned flags = lazy_hint ? DRM_FENCE_FLAG_WAIT_LAZY : 0;
+
+  return drmFenceWait((int)fd, flags, (drmFence *) private, fence_type);
+}
+
+static int
+tUnref(struct _DriFenceMgr *mgr, void **private)
+{
+  long fd = (long) mgr->private;
+  drmFence *fence = (drmFence *) *private;
+  *private = NULL;
+
+  return drmFenceUnreference(fd, fence);
+}
+
+struct _DriFenceMgr *driFenceMgrTTMInit(int fd)
+{
+  struct _DriFenceMgrCreateInfo info;
+  struct _DriFenceMgr *mgr;
+
+  info.flags = DRI_FENCE_CLASS_ORDERED;
+  info.num_classes = 4;
+  info.signaled = tSignaled;
+  info.finish = tFinish;
+  info.unreference = tUnref;
+
+  mgr = driFenceMgrCreate(&info);
+  if (mgr == NULL)
+    return NULL;
+
+  mgr->private = (void *) (long) fd;
+  return mgr;
+}
+
diff --git a/src/gallium/winsys/dri/intel/ws_dri_fencemgr.h b/src/gallium/winsys/dri/intel/ws_dri_fencemgr.h
new file mode 100644
index 00000000000..4ea58dfe183
--- /dev/null
+++ b/src/gallium/winsys/dri/intel/ws_dri_fencemgr.h
@@ -0,0 +1,115 @@
+#ifndef DRI_FENCEMGR_H
+#define DRI_FENCEMGR_H
+
+#include <stdint.h>
+#include <stdlib.h>
+
+struct _DriFenceObject;
+struct _DriFenceMgr;
+
+/*
+ * Do a quick check to see if the fence manager has registered the fence
+ * object as signaled. Note that this function may return a false negative
+ * answer.
+ */
+extern uint32_t driFenceSignaledTypeCached(struct _DriFenceObject *fence);
+
+/*
+ * Check if the fence object is signaled. This function can be substantially
+ * more expensive to call than the above function, but will not return a false
+ * negative answer. The argument "flush_type" sets the types that the
+ * underlying mechanism must make sure will eventually signal.
+ */
+extern int driFenceSignaledType(struct _DriFenceObject *fence,
+				uint32_t flush_type, uint32_t *signaled);
+
+/*
+ * Convenience functions.
+ */
+
+static inline int driFenceSignaled(struct _DriFenceObject *fence,
+				   uint32_t flush_type)
+{
+    uint32_t signaled_types;
+    int ret = driFenceSignaledType(fence, flush_type, &signaled_types);
+    if (ret)
+	return 0;
+    return ((signaled_types & flush_type) == flush_type);
+}
+
+static inline int driFenceSignaledCached(struct _DriFenceObject *fence,
+					 uint32_t flush_type)
+{
+    uint32_t signaled_types =
+	driFenceSignaledTypeCached(fence);
+
+    return ((signaled_types & flush_type) == flush_type);
+}
+
+/*
+ * Reference a fence object.
+ */
+extern struct _DriFenceObject *driFenceReference(struct _DriFenceObject *fence);
+
+/*
+ * Unreference a fence object. The fence object pointer will be reset to NULL.
+ */
+
+extern void driFenceUnReference(struct _DriFenceObject **pFence);
+
+
+/*
+ * Wait for a fence to signal the indicated fence_type.
+ * If "lazy_hint" is true, it indicates that the wait may sleep to avoid
+ * busy-wait polling.
+ */
+extern int driFenceFinish(struct _DriFenceObject *fence, uint32_t fence_type,
+			  int lazy_hint);
+
+/*
+ * Create a DriFenceObject for manager "mgr".
+ *
+ * "private" is a pointer that should be used for the callbacks in
+ * struct _DriFenceMgrCreateInfo.
+ *
+ * if private_size is nonzero, then the info stored at *private, with size
+ * private size will be copied and the fence manager will instead use a
+ * pointer to the copied data for the callbacks in
+ * struct _DriFenceMgrCreateInfo. In that case, the object pointed to by
+ * "private" may be destroyed after the call to driFenceCreate.
+ */
+extern struct _DriFenceObject *driFenceCreate(struct _DriFenceMgr *mgr,
+					      uint32_t fence_class,
+					      uint32_t fence_type,
+					      void *private,
+					      size_t private_size);
+
+extern uint32_t driFenceType(struct _DriFenceObject *fence);
+
+/*
+ * Fence creations are ordered. If a fence signals a fence_type,
+ * it is safe to assume that all fences of the same class that was
+ * created before that fence has signaled the same type.
+ */
+
+#define DRI_FENCE_CLASS_ORDERED (1 << 0)
+
+struct _DriFenceMgrCreateInfo {
+    uint32_t flags;
+    uint32_t num_classes;
+    int (*signaled) (struct _DriFenceMgr *mgr, void *private, uint32_t flush_type,
+		     uint32_t *signaled_type);
+    int (*finish) (struct _DriFenceMgr *mgr, void *private, uint32_t fence_type, int lazy_hint);
+    int (*unreference) (struct _DriFenceMgr *mgr, void **private);
+};
+
+extern struct _DriFenceMgr *
+driFenceMgrCreate(const struct _DriFenceMgrCreateInfo *info);
+
+void
+driFenceMgrUnReference(struct _DriFenceMgr **pMgr);
+
+extern struct _DriFenceMgr *
+driFenceMgrTTMInit(int fd);
+
+#endif
diff --git a/src/gallium/winsys/dri/intel/ws_dri_mallocpool.c b/src/gallium/winsys/dri/intel/ws_dri_mallocpool.c
new file mode 100644
index 00000000000..bf97d7e440b
--- /dev/null
+++ b/src/gallium/winsys/dri/intel/ws_dri_mallocpool.c
@@ -0,0 +1,162 @@
+/**************************************************************************
+ * 
+ * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, TX., USA
+ * 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 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
+ * THE COPYRIGHT HOLDERS, AUTHORS 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.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ * 
+ * 
+ **************************************************************************/
+/*
+ * Authors: Thomas Hellstr�m <thomas-at-tungstengraphics-dot-com>
+ */
+
+#include <xf86drm.h>
+#include <stdlib.h>
+#include <errno.h>
+#include "imports.h"
+#include "glthread.h"
+#include "ws_dri_bufpool.h"
+#include "ws_dri_bufmgr.h"
+#include "intel_screen.h"
+
+static void *
+pool_create(struct _DriBufferPool *pool,
+            unsigned long size, uint64_t flags, unsigned hint,
+            unsigned alignment)
+{
+    unsigned long *private = malloc(size + 2*sizeof(unsigned long));
+    if ((flags & DRM_BO_MASK_MEM) != DRM_BO_FLAG_MEM_LOCAL)
+      abort();
+
+    *private = size;
+    return (void *)private;
+}
+
+
+static int
+pool_destroy(struct _DriBufferPool *pool, void *private)
+{
+    free(private);
+    return 0;
+}
+
+static int
+pool_waitIdle(struct _DriBufferPool *pool, void *private, 
+	      _glthread_Mutex *mutex, int lazy)
+{
+    return 0;
+}
+
+static int
+pool_map(struct _DriBufferPool *pool, void *private, unsigned flags,
+         int hint, _glthread_Mutex *mutex, void **virtual)
+{
+    *virtual = (void *)((unsigned long *)private + 2);
+    return 0;
+}
+
+static int
+pool_unmap(struct _DriBufferPool *pool, void *private)
+{
+    return 0;
+}
+
+static unsigned long
+pool_offset(struct _DriBufferPool *pool, void *private)
+{
+    /*
+     * BUG
+     */
+    abort();
+    return 0UL;
+}
+
+static unsigned long
+pool_poolOffset(struct _DriBufferPool *pool, void *private)
+{
+    /*
+     * BUG
+     */
+    abort();
+}
+
+static uint64_t
+pool_flags(struct _DriBufferPool *pool, void *private)
+{
+    return DRM_BO_FLAG_MEM_LOCAL | DRM_BO_FLAG_CACHED;
+}
+
+static unsigned long
+pool_size(struct _DriBufferPool *pool, void *private)
+{
+    return *(unsigned long *) private;
+}
+
+
+static int
+pool_fence(struct _DriBufferPool *pool, void *private,
+           struct _DriFenceObject *fence)
+{
+    abort();
+    return 0UL;
+}
+
+static drmBO *
+pool_kernel(struct _DriBufferPool *pool, void *private)
+{
+    abort();
+    return NULL;
+}
+
+static void
+pool_takedown(struct _DriBufferPool *pool)
+{
+    free(pool);
+}
+
+
+struct _DriBufferPool *
+driMallocPoolInit(void)
+{
+   struct _DriBufferPool *pool;
+
+   pool = (struct _DriBufferPool *) malloc(sizeof(*pool));
+   if (!pool)
+       return NULL;
+
+   pool->data = NULL;
+   pool->fd = -1;
+   pool->map = &pool_map;
+   pool->unmap = &pool_unmap;
+   pool->destroy = &pool_destroy;
+   pool->offset = &pool_offset;
+   pool->poolOffset = &pool_poolOffset;
+   pool->flags = &pool_flags;
+   pool->size = &pool_size;
+   pool->create = &pool_create;
+   pool->fence = &pool_fence;
+   pool->kernel = &pool_kernel;
+   pool->validate = NULL;
+   pool->waitIdle = &pool_waitIdle;
+   pool->takeDown = &pool_takedown;
+   return pool;
+}
diff --git a/src/gallium/winsys/dri/intel/ws_dri_slabpool.c b/src/gallium/winsys/dri/intel/ws_dri_slabpool.c
new file mode 100644
index 00000000000..235f0ac9726
--- /dev/null
+++ b/src/gallium/winsys/dri/intel/ws_dri_slabpool.c
@@ -0,0 +1,945 @@
+/**************************************************************************
+ *
+ * Copyright 2006-2008 Tungsten Graphics, Inc., Cedar Park, TX., USA
+ * 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 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
+ * THE COPYRIGHT HOLDERS, AUTHORS 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.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ *
+ **************************************************************************/
+/*
+ * Authors: Thomas Hellstrom <thomas-at-tungstengraphics-dot-com>
+ */
+
+#include <stdint.h>
+#include <sys/time.h>
+#include <errno.h>
+#include <unistd.h>
+#include <assert.h>
+#include "ws_dri_bufpool.h"
+#include "ws_dri_fencemgr.h"
+#include "ws_dri_bufmgr.h"
+#include "glthread.h"
+
+#define DRI_SLABPOOL_ALLOC_RETRIES 100
+
+struct _DriSlab;
+
+struct _DriSlabBuffer {
+    int isSlabBuffer;
+    drmBO *bo;
+    struct _DriFenceObject *fence;
+    struct _DriSlab *parent;
+    drmMMListHead head;
+    uint32_t mapCount;
+    uint32_t start;
+    uint32_t fenceType;
+    int unFenced;
+    _glthread_Cond event;
+};
+
+struct _DriKernelBO {
+    int fd;
+    drmBO bo;
+    drmMMListHead timeoutHead;
+    drmMMListHead head;
+    struct timeval timeFreed;
+    uint32_t pageAlignment;
+    void *virtual;
+};
+
+struct _DriSlab{
+    drmMMListHead head;
+    drmMMListHead freeBuffers;
+    uint32_t numBuffers;
+    uint32_t numFree;
+    struct _DriSlabBuffer *buffers;
+    struct _DriSlabSizeHeader *header;
+    struct _DriKernelBO *kbo;
+};
+
+
+struct _DriSlabSizeHeader {
+    drmMMListHead slabs;
+    drmMMListHead freeSlabs;
+    drmMMListHead delayedBuffers;
+    uint32_t numDelayed;
+    struct _DriSlabPool *slabPool;
+    uint32_t bufSize;
+    _glthread_Mutex mutex;
+};
+
+struct _DriFreeSlabManager {
+    struct timeval slabTimeout;
+    struct timeval checkInterval;
+    struct timeval nextCheck;
+    drmMMListHead timeoutList;
+    drmMMListHead unCached;
+    drmMMListHead cached;
+    _glthread_Mutex mutex;
+};
+
+
+struct _DriSlabPool {
+
+    /*
+     * The data of this structure remains constant after
+     * initialization and thus needs no mutex protection.
+     */
+
+    struct _DriFreeSlabManager *fMan;
+    uint64_t proposedFlags;
+    uint64_t validMask;
+    uint32_t *bucketSizes;
+    uint32_t numBuckets;
+    uint32_t pageSize;
+    int fd;
+    int pageAlignment;
+    int maxSlabSize;
+    int desiredNumBuffers;
+    struct _DriSlabSizeHeader *headers;
+};
+
+/*
+ * FIXME: Perhaps arrange timeout slabs in size buckets for fast
+ * retreival??
+ */
+
+
+static inline int
+driTimeAfterEq(struct timeval *arg1, struct timeval *arg2)
+{
+    return ((arg1->tv_sec > arg2->tv_sec) ||
+	    ((arg1->tv_sec == arg2->tv_sec) &&
+	     (arg1->tv_usec > arg2->tv_usec)));
+}
+
+static inline void
+driTimeAdd(struct timeval *arg, struct timeval *add)
+{
+    unsigned int sec;
+
+    arg->tv_sec += add->tv_sec;
+    arg->tv_usec += add->tv_usec;
+    sec = arg->tv_usec / 1000000;
+    arg->tv_sec += sec;
+    arg->tv_usec -= sec*1000000;
+}
+
+static void
+driFreeKernelBO(struct _DriKernelBO *kbo)
+{
+    if (!kbo)
+	return;
+
+    (void) drmBOUnreference(kbo->fd, &kbo->bo);
+    free(kbo);
+}
+
+
+static void
+driFreeTimeoutKBOsLocked(struct _DriFreeSlabManager *fMan,
+			 struct timeval *time)
+{
+    drmMMListHead *list, *next;
+    struct _DriKernelBO *kbo;
+
+    if (!driTimeAfterEq(time, &fMan->nextCheck))
+	return;
+
+    for (list = fMan->timeoutList.next, next = list->next;
+	 list != &fMan->timeoutList;
+	 list = next, next = list->next) {
+
+	kbo = DRMLISTENTRY(struct _DriKernelBO, list, timeoutHead);
+
+	if (!driTimeAfterEq(time, &kbo->timeFreed))
+	    break;
+
+	DRMLISTDELINIT(&kbo->timeoutHead);
+	DRMLISTDELINIT(&kbo->head);
+	driFreeKernelBO(kbo);
+    }
+
+    fMan->nextCheck = *time;
+    driTimeAdd(&fMan->nextCheck, &fMan->checkInterval);
+}
+
+
+/*
+ * Add a _DriKernelBO to the free slab manager.
+ * This means that it is available for reuse, but if it's not
+ * reused in a while, it will be freed.
+ */
+
+static void
+driSetKernelBOFree(struct _DriFreeSlabManager *fMan,
+		   struct _DriKernelBO *kbo)
+{
+    struct timeval time;
+
+    _glthread_LOCK_MUTEX(fMan->mutex);
+    gettimeofday(&time, NULL);
+    driTimeAdd(&time, &fMan->slabTimeout);
+
+    kbo->timeFreed = time;
+
+    if (kbo->bo.flags & DRM_BO_FLAG_CACHED)
+	DRMLISTADD(&kbo->head, &fMan->cached);
+    else
+	DRMLISTADD(&kbo->head, &fMan->unCached);
+
+    DRMLISTADDTAIL(&kbo->timeoutHead, &fMan->timeoutList);
+    driFreeTimeoutKBOsLocked(fMan, &time);
+
+    _glthread_UNLOCK_MUTEX(fMan->mutex);
+}
+
+/*
+ * Get a _DriKernelBO for us to use as storage for a slab.
+ *
+ */
+
+static struct _DriKernelBO *
+driAllocKernelBO(struct _DriSlabSizeHeader *header)
+
+{
+    struct _DriSlabPool *slabPool = header->slabPool;
+    struct _DriFreeSlabManager *fMan = slabPool->fMan;
+    drmMMListHead *list, *next, *head;
+    uint32_t size = header->bufSize * slabPool->desiredNumBuffers;
+    struct _DriKernelBO *kbo;
+    int ret;
+
+    /*
+     * FIXME: We should perhaps allow some variation in slabsize in order
+     * to efficiently reuse slabs.
+     */
+
+    size = (size <= slabPool->maxSlabSize) ? size : slabPool->maxSlabSize;
+    size = (size + slabPool->pageSize - 1) & ~(slabPool->pageSize - 1);
+    _glthread_LOCK_MUTEX(fMan->mutex);
+
+    kbo = NULL;
+
+  retry:
+    head = (slabPool->proposedFlags & DRM_BO_FLAG_CACHED) ?
+	&fMan->cached : &fMan->unCached;
+
+    for (list = head->next, next = list->next;
+	 list != head;
+	 list = next, next = list->next) {
+
+	kbo = DRMLISTENTRY(struct _DriKernelBO, list, head);
+
+	if ((kbo->bo.size == size) &&
+	    (slabPool->pageAlignment == 0 ||
+	     (kbo->pageAlignment % slabPool->pageAlignment) == 0)) {
+
+	    DRMLISTDELINIT(&kbo->head);
+	    DRMLISTDELINIT(&kbo->timeoutHead);
+	    break;
+	}
+
+	kbo = NULL;
+    }
+
+    _glthread_UNLOCK_MUTEX(fMan->mutex);
+
+    if (kbo) {
+	ret = 0;
+	ret = drmBOSetStatus(kbo->fd, &kbo->bo, slabPool->proposedFlags,
+			     (slabPool->proposedFlags ^ kbo->bo.flags),
+			     DRM_BO_HINT_DONT_FENCE, 0, 0);
+	if (ret == 0)
+	    return kbo;
+
+	driFreeKernelBO(kbo);
+	kbo = NULL;
+	goto retry;
+    }
+
+    kbo = calloc(1, sizeof(struct _DriKernelBO));
+    if (!kbo)
+	return NULL;
+
+    kbo->fd = slabPool->fd;
+    DRMINITLISTHEAD(&kbo->head);
+    DRMINITLISTHEAD(&kbo->timeoutHead);
+
+    ret = drmBOCreate(kbo->fd, size, slabPool->pageAlignment, NULL,
+		      slabPool->proposedFlags,
+		      DRM_BO_HINT_DONT_FENCE, &kbo->bo);
+    if (ret)
+	goto out_err0;
+
+    ret = drmBOMap(kbo->fd, &kbo->bo,
+		   DRM_BO_FLAG_READ | DRM_BO_FLAG_WRITE,
+		   0, &kbo->virtual);
+
+    if (ret)
+	goto out_err1;
+
+    ret = drmBOUnmap(kbo->fd, &kbo->bo);
+    if (ret)
+	goto out_err1;
+
+    return kbo;
+
+  out_err1:
+    drmBOUnreference(kbo->fd, &kbo->bo);
+  out_err0:
+    free(kbo);
+    return NULL;
+}
+
+
+static int
+driAllocSlab(struct _DriSlabSizeHeader *header)
+{
+    struct _DriSlab *slab;
+    struct _DriSlabBuffer *buf;
+    uint32_t numBuffers;
+    int ret;
+    int i;
+
+    slab = calloc(1, sizeof(*slab));
+    if (!slab)
+	return -ENOMEM;
+
+    slab->kbo = driAllocKernelBO(header);
+    if (!slab->kbo) {
+	ret = -ENOMEM;
+	goto out_err0;
+    }
+
+    numBuffers = slab->kbo->bo.size / header->bufSize;
+
+    slab->buffers = calloc(numBuffers, sizeof(*slab->buffers));
+    if (!slab->buffers) {
+	ret = -ENOMEM;
+	goto out_err1;
+    }
+
+    DRMINITLISTHEAD(&slab->head);
+    DRMINITLISTHEAD(&slab->freeBuffers);
+    slab->numBuffers = numBuffers;
+    slab->numFree = 0;
+    slab->header = header;
+
+    buf = slab->buffers;
+    for (i=0; i < numBuffers; ++i) {
+	buf->parent = slab;
+	buf->start = i* header->bufSize;
+	buf->mapCount = 0;
+	buf->isSlabBuffer = 1;
+	_glthread_INIT_COND(buf->event);
+	DRMLISTADDTAIL(&buf->head, &slab->freeBuffers);
+	slab->numFree++;
+	buf++;
+    }
+
+    DRMLISTADDTAIL(&slab->head, &header->slabs);
+
+    return 0;
+
+  out_err1:
+    driSetKernelBOFree(header->slabPool->fMan, slab->kbo);
+    free(slab->buffers);
+  out_err0:
+    free(slab);
+    return ret;
+}
+
+/*
+ * Delete a buffer from the slab header delayed list and put
+ * it on the slab free list.
+ */
+
+static void
+driSlabFreeBufferLocked(struct _DriSlabBuffer *buf)
+{
+    struct _DriSlab *slab = buf->parent;
+    struct _DriSlabSizeHeader *header = slab->header;
+    drmMMListHead *list = &buf->head;
+
+    DRMLISTDEL(list);
+    DRMLISTADDTAIL(list, &slab->freeBuffers);
+    slab->numFree++;
+
+    if (slab->head.next == &slab->head)
+	DRMLISTADDTAIL(&slab->head, &header->slabs);
+
+    if (slab->numFree == slab->numBuffers) {
+	list = &slab->head;
+	DRMLISTDEL(list);
+	DRMLISTADDTAIL(list, &header->freeSlabs);
+    }
+
+    if (header->slabs.next == &header->slabs ||
+	slab->numFree != slab->numBuffers) {
+
+	drmMMListHead *next;
+	struct _DriFreeSlabManager *fMan = header->slabPool->fMan;
+
+	for (list = header->freeSlabs.next, next = list->next;
+	     list != &header->freeSlabs;
+	     list = next, next = list->next) {
+
+	    slab = DRMLISTENTRY(struct _DriSlab, list, head);
+
+	    DRMLISTDELINIT(list);
+	    driSetKernelBOFree(fMan, slab->kbo);
+	    free(slab->buffers);
+	    free(slab);
+	}
+    }
+}
+
+static void
+driSlabCheckFreeLocked(struct _DriSlabSizeHeader *header, int wait)
+{
+   drmMMListHead *list, *prev;
+   struct _DriSlabBuffer *buf;
+   struct _DriSlab *slab;
+
+   int signaled = 0;
+   int i;
+   int ret;
+
+   list = header->delayedBuffers.next;
+
+   /* Only examine the oldest 1/3 of delayed buffers:
+    */
+   if (header->numDelayed > 3) {
+      for (i = 0; i < header->numDelayed; i += 3) {
+         list = list->next;
+      }
+   }
+
+   prev = list->prev;
+   for (; list != &header->delayedBuffers; list = prev, prev = list->prev) {
+       buf = DRMLISTENTRY(struct _DriSlabBuffer, list, head);
+       slab = buf->parent;
+
+       if (!signaled) {
+	   if (wait) {
+	       ret = driFenceFinish(buf->fence, buf->fenceType, 0);
+	       if (ret)
+		   break;
+	       signaled = 1;
+	   } else {
+	       signaled = driFenceSignaled(buf->fence, buf->fenceType);
+	   }
+	   if (signaled) {
+	       driFenceUnReference(&buf->fence);
+	       header->numDelayed--;
+	       driSlabFreeBufferLocked(buf);
+	   }
+       } else if (driFenceSignaledCached(buf->fence, buf->fenceType)) {
+	   driFenceUnReference(&buf->fence);
+	   header->numDelayed--;
+	   driSlabFreeBufferLocked(buf);
+       }
+   }
+}
+
+
+static struct _DriSlabBuffer *
+driSlabAllocBuffer(struct _DriSlabSizeHeader *header)
+{
+    static struct _DriSlabBuffer *buf;
+    struct _DriSlab *slab;
+    drmMMListHead *list;
+    int count = DRI_SLABPOOL_ALLOC_RETRIES;
+
+    _glthread_LOCK_MUTEX(header->mutex);
+    while(header->slabs.next == &header->slabs && count > 0) {
+        driSlabCheckFreeLocked(header, 0);
+	if (header->slabs.next != &header->slabs)
+	  break;
+
+	_glthread_UNLOCK_MUTEX(header->mutex);
+	if (count != DRI_SLABPOOL_ALLOC_RETRIES)
+	    usleep(1);
+	(void) driAllocSlab(header);
+	_glthread_LOCK_MUTEX(header->mutex);
+	count--;
+    }
+
+    list = header->slabs.next;
+    if (list == &header->slabs) {
+	_glthread_UNLOCK_MUTEX(header->mutex);
+	//assert(0, "no buffers in slab");
+	return NULL;
+    }
+    slab = DRMLISTENTRY(struct _DriSlab, list, head);
+    if (--slab->numFree == 0)
+	DRMLISTDELINIT(list);
+
+    list = slab->freeBuffers.next;
+    DRMLISTDELINIT(list);
+
+    _glthread_UNLOCK_MUTEX(header->mutex);
+    buf = DRMLISTENTRY(struct _DriSlabBuffer, list, head);
+    return buf;
+}
+
+static void *
+pool_create(struct _DriBufferPool *driPool, unsigned long size,
+	    uint64_t flags, unsigned hint, unsigned alignment)
+{
+    struct _DriSlabPool *pool = (struct _DriSlabPool *) driPool->data;
+    struct _DriSlabSizeHeader *header;
+    struct _DriSlabBuffer *buf;
+    void *dummy;
+    int i;
+    int ret;
+
+
+    /*
+     * FIXME: Check for compatibility.
+     */
+
+    header = pool->headers;
+    for (i=0; i<pool->numBuckets; ++i) {
+      if (header->bufSize >= size)
+	break;
+      header++;
+    }
+
+    if (i < pool->numBuckets)
+	return driSlabAllocBuffer(header);
+
+
+    /*
+     * Fall back to allocate a buffer object directly from DRM.
+     * and wrap it in a driBO structure.
+     */
+
+
+    buf = calloc(1, sizeof(*buf));
+
+    if (!buf)
+	return NULL;
+
+    buf->bo = calloc(1, sizeof(*buf->bo));
+    if (!buf->bo)
+	goto out_err0;
+
+    if (alignment) {
+	if ((alignment < pool->pageSize) && (pool->pageSize % alignment))
+	    goto out_err1;
+	if ((alignment > pool->pageSize) && (alignment % pool->pageSize))
+	    goto out_err1;
+    }
+
+    ret = drmBOCreate(pool->fd, size, alignment / pool->pageSize, NULL,
+			flags, hint, buf->bo);
+    if (ret)
+	goto out_err1;
+
+    ret  = drmBOMap(pool->fd, buf->bo, DRM_BO_FLAG_READ | DRM_BO_FLAG_WRITE,
+		    0, &dummy);
+    if (ret)
+	goto out_err2;
+
+    ret = drmBOUnmap(pool->fd, buf->bo);
+    if (ret)
+	goto out_err2;
+
+    return buf;
+  out_err2:
+    drmBOUnreference(pool->fd, buf->bo);
+  out_err1:
+    free(buf->bo);
+  out_err0:
+    free(buf);
+    return NULL;
+}
+
+static int
+pool_destroy(struct _DriBufferPool *driPool, void *private)
+{
+    struct _DriSlabBuffer *buf =
+	(struct _DriSlabBuffer *) private;
+    struct _DriSlab *slab;
+    struct _DriSlabSizeHeader *header;
+
+    if (!buf->isSlabBuffer) {
+	struct _DriSlabPool *pool = (struct _DriSlabPool *) driPool->data;
+	int ret;
+
+	ret = drmBOUnreference(pool->fd, buf->bo);
+	free(buf->bo);
+	free(buf);
+	return ret;
+    }
+
+    slab = buf->parent;
+    header = slab->header;
+
+    _glthread_LOCK_MUTEX(header->mutex);
+    buf->unFenced = 0;
+    buf->mapCount = 0;
+
+    if (buf->fence && !driFenceSignaledCached(buf->fence, buf->fenceType)) {
+	DRMLISTADDTAIL(&buf->head, &header->delayedBuffers);
+	header->numDelayed++;
+    } else {
+	driSlabFreeBufferLocked(buf);
+    }
+
+    _glthread_UNLOCK_MUTEX(header->mutex);
+    return 0;
+}
+
+static int
+pool_waitIdle(struct _DriBufferPool *driPool, void *private, 
+	      _glthread_Mutex *mutex, int lazy)
+{
+   struct _DriSlabBuffer *buf = (struct _DriSlabBuffer *) private;
+
+   while(buf->unFenced)
+       _glthread_COND_WAIT(buf->event, *mutex);
+
+   if (!buf->fence)
+     return 0;
+
+   driFenceFinish(buf->fence, buf->fenceType, lazy);
+   driFenceUnReference(&buf->fence);
+
+   return 0;
+}
+
+static int
+pool_map(struct _DriBufferPool *pool, void *private, unsigned flags,
+         int hint, _glthread_Mutex *mutex, void **virtual)
+{
+   struct _DriSlabBuffer *buf = (struct _DriSlabBuffer *) private;
+   int busy;
+
+   if (buf->isSlabBuffer)
+       busy = buf->unFenced || (buf->fence && !driFenceSignaledCached(buf->fence, buf->fenceType));
+   else
+       busy = buf->fence && !driFenceSignaled(buf->fence, buf->fenceType);
+
+
+   if (busy) {
+       if (hint & DRM_BO_HINT_DONT_BLOCK)
+	   return -EBUSY;
+       else {
+	   (void) pool_waitIdle(pool, private, mutex, 0);
+       }
+   }
+
+   ++buf->mapCount;
+   *virtual = (buf->isSlabBuffer) ?
+       (void *) ((uint8_t *) buf->parent->kbo->virtual + buf->start) :
+       (void *) buf->bo->virtual;
+
+   return 0;
+}
+
+static int
+pool_unmap(struct _DriBufferPool *pool, void *private)
+{
+   struct _DriSlabBuffer *buf = (struct _DriSlabBuffer *) private;
+
+   --buf->mapCount;
+   if (buf->mapCount == 0 && buf->isSlabBuffer) 
+       _glthread_COND_BROADCAST(buf->event);
+
+   return 0;
+}
+
+static unsigned long
+pool_offset(struct _DriBufferPool *pool, void *private)
+{
+   struct _DriSlabBuffer *buf = (struct _DriSlabBuffer *) private;
+   struct _DriSlab *slab;
+   struct _DriSlabSizeHeader *header;
+
+   if (!buf->isSlabBuffer) {
+       assert(buf->bo->proposedFlags & DRM_BO_FLAG_NO_MOVE);
+       return buf->bo->offset;
+   }
+
+   slab = buf->parent;
+   header = slab->header;
+
+   (void) header;
+   assert(header->slabPool->proposedFlags & DRM_BO_FLAG_NO_MOVE);
+   return slab->kbo->bo.offset + buf->start;
+}
+
+static unsigned long
+pool_poolOffset(struct _DriBufferPool *pool, void *private)
+{
+   struct _DriSlabBuffer *buf = (struct _DriSlabBuffer *) private;
+
+   return buf->start;
+}
+
+static uint64_t
+pool_flags(struct _DriBufferPool *pool, void *private)
+{
+   struct _DriSlabBuffer *buf = (struct _DriSlabBuffer *) private;
+
+   if (!buf->isSlabBuffer)
+       return buf->bo->flags;
+
+   return buf->parent->kbo->bo.flags;
+}
+
+static unsigned long
+pool_size(struct _DriBufferPool *pool, void *private)
+{
+   struct _DriSlabBuffer *buf = (struct _DriSlabBuffer *) private;
+   if (!buf->isSlabBuffer)
+       return buf->bo->size;
+
+   return buf->parent->header->bufSize;
+}
+
+static int
+pool_fence(struct _DriBufferPool *pool, void *private,
+           struct _DriFenceObject *fence)
+{
+   struct _DriSlabBuffer *buf = (struct _DriSlabBuffer *) private;
+   drmBO *bo;
+
+   if (buf->fence)
+      driFenceUnReference(&buf->fence);
+
+   buf->fence = driFenceReference(fence);
+   bo = (buf->isSlabBuffer) ?
+     &buf->parent->kbo->bo:
+     buf->bo;
+   buf->fenceType = bo->fenceFlags;
+
+   buf->unFenced = 0;
+   _glthread_COND_BROADCAST(buf->event);
+
+   return 0;
+}
+
+static drmBO *
+pool_kernel(struct _DriBufferPool *pool, void *private)
+{
+   struct _DriSlabBuffer *buf = (struct _DriSlabBuffer *) private;
+
+   return (buf->isSlabBuffer) ? &buf->parent->kbo->bo : buf->bo;
+}
+
+static int
+pool_validate(struct _DriBufferPool *pool, void *private, 
+	      _glthread_Mutex *mutex)
+{
+   struct _DriSlabBuffer *buf = (struct _DriSlabBuffer *) private;
+
+   if (!buf->isSlabBuffer) 
+       return 0;
+   
+   while(buf->mapCount != 0)
+       _glthread_COND_WAIT(buf->event, *mutex);
+
+   buf->unFenced = 1;
+   return 0;
+}
+
+
+struct _DriFreeSlabManager *
+driInitFreeSlabManager(uint32_t checkIntervalMsec, uint32_t slabTimeoutMsec)
+{
+    struct _DriFreeSlabManager *tmp;
+
+    tmp = calloc(1, sizeof(*tmp));
+    if (!tmp)
+	return NULL;
+
+    _glthread_INIT_MUTEX(tmp->mutex);
+    _glthread_LOCK_MUTEX(tmp->mutex);
+    tmp->slabTimeout.tv_usec = slabTimeoutMsec*1000;
+    tmp->slabTimeout.tv_sec = tmp->slabTimeout.tv_usec / 1000000;
+    tmp->slabTimeout.tv_usec -=  tmp->slabTimeout.tv_sec*1000000;
+
+    tmp->checkInterval.tv_usec = checkIntervalMsec*1000;
+    tmp->checkInterval.tv_sec = tmp->checkInterval.tv_usec / 1000000;
+    tmp->checkInterval.tv_usec -=  tmp->checkInterval.tv_sec*1000000;
+
+    gettimeofday(&tmp->nextCheck, NULL);
+    driTimeAdd(&tmp->nextCheck, &tmp->checkInterval);
+    DRMINITLISTHEAD(&tmp->timeoutList);
+    DRMINITLISTHEAD(&tmp->unCached);
+    DRMINITLISTHEAD(&tmp->cached);
+    _glthread_UNLOCK_MUTEX(tmp->mutex);
+
+    return tmp;
+}
+
+void
+driFinishFreeSlabManager(struct _DriFreeSlabManager *fMan)
+{
+    struct timeval time;
+
+    time = fMan->nextCheck;
+    driTimeAdd(&time, &fMan->checkInterval);
+
+    _glthread_LOCK_MUTEX(fMan->mutex);
+    driFreeTimeoutKBOsLocked(fMan, &time);
+    _glthread_UNLOCK_MUTEX(fMan->mutex);
+
+    assert(fMan->timeoutList.next == &fMan->timeoutList);
+    assert(fMan->unCached.next == &fMan->unCached);
+    assert(fMan->cached.next == &fMan->cached);
+
+    free(fMan);
+}
+
+static void
+driInitSizeHeader(struct _DriSlabPool *pool, uint32_t size,
+		  struct _DriSlabSizeHeader *header)
+{
+    _glthread_INIT_MUTEX(header->mutex);
+    _glthread_LOCK_MUTEX(header->mutex);
+
+    DRMINITLISTHEAD(&header->slabs);
+    DRMINITLISTHEAD(&header->freeSlabs);
+    DRMINITLISTHEAD(&header->delayedBuffers);
+
+    header->numDelayed = 0;
+    header->slabPool = pool;
+    header->bufSize = size;
+
+    _glthread_UNLOCK_MUTEX(header->mutex);
+}
+
+static void
+driFinishSizeHeader(struct _DriSlabSizeHeader *header)
+{
+    drmMMListHead *list, *next;
+    struct _DriSlabBuffer *buf;
+
+    _glthread_LOCK_MUTEX(header->mutex);
+    for (list = header->delayedBuffers.next, next = list->next;
+	 list != &header->delayedBuffers;
+	 list = next, next = list->next) {
+
+	buf = DRMLISTENTRY(struct _DriSlabBuffer, list , head);
+	if (buf->fence) {
+	    (void) driFenceFinish(buf->fence, buf->fenceType, 0);
+	    driFenceUnReference(&buf->fence);
+	}
+	header->numDelayed--;
+	driSlabFreeBufferLocked(buf);
+    }
+    _glthread_UNLOCK_MUTEX(header->mutex);
+}
+
+static void
+pool_takedown(struct _DriBufferPool *driPool)
+{
+   struct _DriSlabPool *pool = driPool->data;
+   int i;
+
+   for (i=0; i<pool->numBuckets; ++i) {
+     driFinishSizeHeader(&pool->headers[i]);
+   }
+
+   free(pool->headers);
+   free(pool->bucketSizes);
+   free(pool);
+   free(driPool);
+}
+
+struct _DriBufferPool *
+driSlabPoolInit(int fd, uint64_t flags,
+		uint64_t validMask,
+		uint32_t smallestSize,
+		uint32_t numSizes,
+		uint32_t desiredNumBuffers,
+		uint32_t maxSlabSize,
+		uint32_t pageAlignment,
+		struct _DriFreeSlabManager *fMan)
+{
+    struct _DriBufferPool *driPool;
+    struct _DriSlabPool *pool;
+    uint32_t i;
+
+    driPool = calloc(1, sizeof(*driPool));
+    if (!driPool)
+	return NULL;
+
+    pool = calloc(1, sizeof(*pool));
+    if (!pool)
+	goto out_err0;
+
+    pool->bucketSizes = calloc(numSizes, sizeof(*pool->bucketSizes));
+    if (!pool->bucketSizes)
+	goto out_err1;
+
+    pool->headers = calloc(numSizes, sizeof(*pool->headers));
+    if (!pool->headers)
+	goto out_err2;
+
+    pool->fMan = fMan;
+    pool->proposedFlags = flags;
+    pool->validMask = validMask;
+    pool->numBuckets = numSizes;
+    pool->pageSize = getpagesize();
+    pool->fd = fd;
+    pool->pageAlignment = pageAlignment;
+    pool->maxSlabSize = maxSlabSize;
+    pool->desiredNumBuffers = desiredNumBuffers;
+
+    for (i=0; i<pool->numBuckets; ++i) {
+	pool->bucketSizes[i] = (smallestSize << i);
+	driInitSizeHeader(pool, pool->bucketSizes[i],
+			  &pool->headers[i]);
+    }
+
+    driPool->data = (void *) pool;
+    driPool->map = &pool_map;
+    driPool->unmap = &pool_unmap;
+    driPool->destroy = &pool_destroy;
+    driPool->offset = &pool_offset;
+    driPool->poolOffset = &pool_poolOffset;
+    driPool->flags = &pool_flags;
+    driPool->size = &pool_size;
+    driPool->create = &pool_create;
+    driPool->fence = &pool_fence;
+    driPool->kernel = &pool_kernel;
+    driPool->validate = &pool_validate;
+    driPool->waitIdle = &pool_waitIdle;
+    driPool->takeDown = &pool_takedown;
+
+    return driPool;
+
+  out_err2:
+    free(pool->bucketSizes);
+  out_err1:
+    free(pool);
+  out_err0:
+    free(driPool);
+
+    return NULL;
+}