Add intel_bufmgr_gem for new graphics execution manager.

10 files changed, 869 insertions, 40 deletions

diff --git a/src/mesa/drivers/dri/common/dri_bufmgr.c b/src/mesa/drivers/dri/common/dri_bufmgr.c
index 4df006fb9f..70ae091499 100644
--- a/src/mesa/drivers/dri/common/dri_bufmgr.c
+++ b/src/mesa/drivers/dri/common/dri_bufmgr.c
@@ -148,9 +148,9 @@ int dri_emit_reloc(dri_bo *reloc_buf, uint64_t flags, GLuint delta,
    return reloc_buf->bufmgr->emit_reloc(reloc_buf, flags, delta, offset, target_buf);
 }
 
-void *dri_process_relocs(dri_bo *batch_buf, GLuint *count)
+void *dri_process_relocs(dri_bo *batch_buf)
 {
-   return batch_buf->bufmgr->process_relocs(batch_buf, count);
+   return batch_buf->bufmgr->process_relocs(batch_buf);
 }
 
 void dri_post_submit(dri_bo *batch_buf, dri_fence **last_fence)
diff --git a/src/mesa/drivers/dri/common/dri_bufmgr.h b/src/mesa/drivers/dri/common/dri_bufmgr.h
index 4593eaf9f7..cbfeb9136d 100644
--- a/src/mesa/drivers/dri/common/dri_bufmgr.h
+++ b/src/mesa/drivers/dri/common/dri_bufmgr.h
@@ -41,7 +41,12 @@ typedef struct _dri_bo dri_bo;
 typedef struct _dri_fence dri_fence;
 
 struct _dri_bo {
-   /** Size in bytes of the buffer object. */
+   /**
+    * Size in bytes of the buffer object.
+    *
+    * The size may be larger than the size originally requested for the
+    * allocation, such as being aligned to page size.
+    */
    unsigned long size;
    /**
     * Card virtual address (offset from the beginning of the aperture) for the
@@ -169,10 +174,10 @@ struct _dri_bufmgr {
     * into them the appopriate order.
     *
     * \param batch_buf buffer at the root of the tree of relocations
-    * \param count returns the number of buffers validated.
-    * \return relocation record for use in command submission.
-    * */
-   void *(*process_relocs)(dri_bo *batch_buf, GLuint *count);
+    * \return argument to be completed and passed to the execbuffers ioctl
+    *   (if any).
+    */
+   void *(*process_relocs)(dri_bo *batch_buf);
 
    void (*post_submit)(dri_bo *batch_buf, dri_fence **fence);
 
@@ -214,7 +219,7 @@ void dri_bufmgr_destroy(dri_bufmgr *bufmgr);
 
 int dri_emit_reloc(dri_bo *reloc_buf, uint64_t flags, GLuint delta,
 		   GLuint offset, dri_bo *target_buf);
-void *dri_process_relocs(dri_bo *batch_buf, uint32_t *count);
+void *dri_process_relocs(dri_bo *batch_buf);
 void dri_post_process_relocs(dri_bo *batch_buf);
 void dri_post_submit(dri_bo *batch_buf, dri_fence **last_fence);
 int dri_bufmgr_check_aperture_space(dri_bo *bo);
diff --git a/src/mesa/drivers/dri/common/dri_bufmgr_fake.c b/src/mesa/drivers/dri/common/dri_bufmgr_fake.c
index 9bf3f3437c..9dd06b07eb 100644
--- a/src/mesa/drivers/dri/common/dri_bufmgr_fake.c
+++ b/src/mesa/drivers/dri/common/dri_bufmgr_fake.c
@@ -1098,7 +1098,7 @@ dri_fake_reloc_and_validate_buffer(dri_bo *bo)
 }
 
 static void *
-dri_fake_process_relocs(dri_bo *batch_buf, GLuint *count_p)
+dri_fake_process_relocs(dri_bo *batch_buf)
 {
    dri_bufmgr_fake *bufmgr_fake = (dri_bufmgr_fake *)batch_buf->bufmgr;
    dri_bo_fake *batch_fake = (dri_bo_fake *)batch_buf;
@@ -1126,8 +1126,6 @@ dri_fake_process_relocs(dri_bo *batch_buf, GLuint *count_p)
 
    assert(ret == 0);
 
-   *count_p = 0; /* junk */
-
    bufmgr_fake->current_total_size = 0;
    return NULL;
 }
diff --git a/src/mesa/drivers/dri/i965/Makefile b/src/mesa/drivers/dri/i965/Makefile
index d46b3428f5..ca9b7da40f 100644
--- a/src/mesa/drivers/dri/i965/Makefile
+++ b/src/mesa/drivers/dri/i965/Makefile
@@ -9,7 +9,7 @@ DRIVER_SOURCES = \
 	intel_blit.c \
 	intel_buffer_objects.c \
 	intel_buffers.c \
-	intel_bufmgr_ttm.c \
+	intel_bufmgr_gem.c \
 	intel_context.c \
 	intel_decode.c \
 	intel_depthstencil.c \
diff --git a/src/mesa/drivers/dri/i965/intel_bufmgr_gem.c b/src/mesa/drivers/dri/i965/intel_bufmgr_gem.c
new file mode 120000
index 0000000000..dee0daf9c0
--- /dev/null
+++ b/src/mesa/drivers/dri/i965/intel_bufmgr_gem.c
@@ -0,0 +1 @@
+../intel/intel_bufmgr_gem.c
\ No newline at end of file
diff --git a/src/mesa/drivers/dri/intel/intel_batchbuffer.c b/src/mesa/drivers/dri/intel/intel_batchbuffer.c
index a594fb6cc4..c5b0f531d4 100644
--- a/src/mesa/drivers/dri/intel/intel_batchbuffer.c
+++ b/src/mesa/drivers/dri/intel/intel_batchbuffer.c
@@ -131,11 +131,8 @@ do_flush_locked(struct intel_batchbuffer *batch,
 		GLuint used, GLboolean allow_unlock)
 {
    struct intel_context *intel = batch->intel;
-   void *start;
-   GLuint count;
 
    dri_bo_unmap(batch->buf);
-   start = dri_process_relocs(batch->buf, &count);
 
    batch->map = NULL;
    batch->ptr = NULL;
@@ -148,12 +145,16 @@ do_flush_locked(struct intel_batchbuffer *batch,
    if (!(intel->numClipRects == 0 &&
 	 batch->cliprect_mode == LOOP_CLIPRECTS)) {
       if (intel->ttm == GL_TRUE) {
+	 struct drm_i915_gem_execbuffer *execbuf;
+
+	 execbuf = dri_process_relocs(batch->buf);
 	 intel_exec_ioctl(batch->intel,
 			  used,
 			  batch->cliprect_mode != LOOP_CLIPRECTS,
 			  allow_unlock,
-			  start, count, &batch->last_fence);
+			  execbuf, &batch->last_fence);
       } else {
+	 dri_process_relocs(batch->buf);
 	 intel_batch_ioctl(batch->intel,
 			   batch->buf->offset,
 			   used,
diff --git a/src/mesa/drivers/dri/intel/intel_bufmgr_gem.c b/src/mesa/drivers/dri/intel/intel_bufmgr_gem.c
new file mode 100644
index 0000000000..2d8dced214
--- /dev/null
+++ b/src/mesa/drivers/dri/intel/intel_bufmgr_gem.c
@@ -0,0 +1,818 @@
+/**************************************************************************
+ *
+ * Copyright © 2007 Red Hat Inc.
+ * Copyright © 2007 Intel Corporation
+ * 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 Hellström <thomas-at-tungstengraphics-dot-com>
+ *          Keith Whitwell <keithw-at-tungstengraphics-dot-com>
+ *	    Eric Anholt <eric@anholt.net>
+ *	    Dave Airlie <airlied@linux.ie>
+ */
+
+#include <xf86drm.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+#include <assert.h>
+#include <sys/ioctl.h>
+#include <sys/mman.h>
+
+#include "errno.h"
+#include "mtypes.h"
+#include "dri_bufmgr.h"
+#include "string.h"
+#include "imports.h"
+
+#include "i915_drm.h"
+
+#include "intel_bufmgr_gem.h"
+
+#define DBG(...) do {					\
+   if (bufmgr_gem->bufmgr.debug)			\
+      fprintf(stderr, __VA_ARGS__);			\
+} while (0)
+
+struct intel_validate_entry {
+    dri_bo *bo;
+    struct drm_i915_op_arg bo_arg;
+};
+
+struct dri_gem_bo_bucket_entry {
+   uint32_t gem_handle;
+   struct dri_gem_bo_bucket_entry *next;
+};
+
+struct dri_gem_bo_bucket {
+   struct dri_gem_bo_bucket_entry *head;
+   struct dri_gem_bo_bucket_entry **tail;
+   /**
+    * Limit on the number of entries in this bucket.
+    *
+    * 0 means that this caching at this bucket size is disabled.
+    * -1 means that there is no limit to caching at this size.
+    */
+   int max_entries;
+   int num_entries;
+};
+
+/* Arbitrarily chosen, 16 means that the maximum size we'll cache for reuse
+ * is 1 << 16 pages, or 256MB.
+ */
+#define INTEL_GEM_BO_BUCKETS	16
+typedef struct _dri_bufmgr_gem {
+    dri_bufmgr bufmgr;
+
+    int fd;
+
+    uint32_t max_relocs;
+
+    struct drm_i915_gem_validate_entry *validate_array;
+    dri_bo **validate_bo;
+    int validate_array_size;
+    int validate_count;
+
+    /** Array of lists of cached gem objects of power-of-two sizes */
+    struct dri_gem_bo_bucket cache_bucket[INTEL_GEM_BO_BUCKETS];
+
+    struct drm_i915_gem_execbuffer exec_arg;
+} dri_bufmgr_gem;
+
+typedef struct _dri_bo_gem {
+    dri_bo bo;
+
+    int refcount;
+    unsigned int map_count;
+    uint32_t gem_handle;
+    const char *name;
+
+    /**
+     * Index of the buffer within the validation list while preparing a
+     * batchbuffer execution.
+     */
+    int validate_index;
+
+    /** Array passed to the DRM containing relocation information. */
+    struct drm_i915_gem_relocation_entry *relocs;
+    /** Array of bos corresponding to relocs[i].target_handle */
+    dri_bo **reloc_target_bo;
+    /** Number of entries in relocs */
+    int reloc_count;
+    /** Mapped address for the buffer */
+    void *virtual;
+} dri_bo_gem;
+
+typedef struct _dri_fence_gem
+{
+    dri_fence fence;
+
+    int refcount;
+    const char *name;
+    drmFence drm_fence;
+} dri_fence_gem;
+
+static int
+logbase2(int n)
+{
+   GLint i = 1;
+   GLint log2 = 0;
+
+   while (n > i) {
+      i *= 2;
+      log2++;
+   }
+
+   return log2;
+}
+
+static struct dri_gem_bo_bucket *
+dri_gem_bo_bucket_for_size(dri_bufmgr_gem *bufmgr_gem, unsigned long size)
+{
+    int i;
+
+    /* We only do buckets in power of two increments */
+    if ((size & (size - 1)) != 0)
+	return NULL;
+
+    /* We should only see sizes rounded to pages. */
+    assert((size % 4096) == 0);
+
+    /* We always allocate in units of pages */
+    i = ffs(size / 4096) - 1;
+    if (i >= INTEL_GEM_BO_BUCKETS)
+	return NULL;
+
+    return &bufmgr_gem->cache_bucket[i];
+}
+
+
+static void dri_gem_dump_validation_list(dri_bufmgr_gem *bufmgr_gem)
+{
+    int i, j;
+
+    for (i = 0; i < bufmgr_gem->validate_count; i++) {
+	dri_bo *bo = bufmgr_gem->validate_bo[i];
+	dri_bo_gem *bo_gem = (dri_bo_gem *)bo;
+
+	if (bo_gem->relocs == NULL) {
+	    DBG("%2d: %s\n", i, bo_gem->name);
+	    continue;
+	}
+
+	for (j = 0; j < bo_gem->reloc_count; j++) {
+	    dri_bo *target_bo = bo_gem->reloc_target_bo[j];
+	    dri_bo_gem *target_gem = (dri_bo_gem *)target_bo;
+
+	    DBG("%2d: %s@0x%08llx -> %s@0x%08lx + 0x%08x\n",
+		i,
+		bo_gem->name, bo_gem->relocs[j].offset,
+		target_gem->name, target_bo->offset,
+		bo_gem->relocs[j].delta);
+	}
+    }
+}
+
+/**
+ * Adds the given buffer to the list of buffers to be validated (moved into the
+ * appropriate memory type) with the next batch submission.
+ *
+ * If a buffer is validated multiple times in a batch submission, it ends up
+ * with the intersection of the memory type flags and the union of the
+ * access flags.
+ */
+static void
+intel_add_validate_buffer(dri_bo *bo)
+{
+    dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *)bo->bufmgr;
+    dri_bo_gem *bo_gem = (dri_bo_gem *)bo;
+    int index;
+
+    if (bo_gem->validate_index != -1)
+	return;
+
+    /* Extend the array of validation entries as necessary. */
+    if (bufmgr_gem->validate_count == bufmgr_gem->validate_array_size) {
+	int new_size = bufmgr_gem->validate_array_size * 2;
+
+	if (new_size == 0)
+	    new_size = 5;
+
+	bufmgr_gem->validate_array =
+	    realloc(bufmgr_gem->validate_array,
+		    sizeof(*bufmgr_gem->validate_array) * new_size);
+	bufmgr_gem->validate_bo =
+	    realloc(bufmgr_gem->validate_bo,
+		    sizeof(*bufmgr_gem->validate_bo) * new_size);
+	bufmgr_gem->validate_array_size = new_size;
+    }
+
+    index = bufmgr_gem->validate_count;
+    /* Fill in array entry */
+    bufmgr_gem->validate_array[index].buffer_handle = bo_gem->gem_handle;
+    bufmgr_gem->validate_array[index].relocation_count = bo_gem->reloc_count;
+    bufmgr_gem->validate_array[index].relocs_ptr = (uintptr_t)bo_gem->relocs;
+    bufmgr_gem->validate_bo[index] = bo;
+    dri_bo_reference(bo);
+    bufmgr_gem->validate_count++;
+}
+
+
+#define RELOC_BUF_SIZE(x) ((I915_RELOC_HEADER + x * I915_RELOC0_STRIDE) * \
+	sizeof(uint32_t))
+
+static int
+intel_setup_reloc_list(dri_bo *bo)
+{
+    dri_bo_gem *bo_gem = (dri_bo_gem *)bo;
+    dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *)bo->bufmgr;
+
+    bo_gem->relocs = calloc(bufmgr_gem->max_relocs,
+			    sizeof(struct drm_i915_gem_relocation_entry));
+    bo_gem->reloc_target_bo = calloc(1, sizeof(dri_bo *));
+
+    return 0;
+}
+
+static dri_bo *
+dri_gem_alloc(dri_bufmgr *bufmgr, const char *name,
+	      unsigned long size, unsigned int alignment,
+	      uint64_t location_mask)
+{
+    dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *)bufmgr;
+    dri_bo_gem *bo_gem;
+    unsigned int page_size = getpagesize();
+    int ret;
+    struct dri_gem_bo_bucket *bucket;
+    GLboolean alloc_from_cache = GL_FALSE;
+
+    bo_gem = calloc(1, sizeof(*bo_gem));
+    if (!bo_gem)
+	return NULL;
+
+    /* Round the allocated size up to a power of two number of pages. */
+    bo_gem->bo.size = 1 << logbase2(size);
+    if (bo_gem->bo.size < page_size)
+	bo_gem->bo.size = page_size;
+    bucket = dri_gem_bo_bucket_for_size(bufmgr_gem, bo_gem->bo.size);
+
+    /* If we don't have caching at this size, don't actually round the
+     * allocation up.
+     */
+    if (bucket == NULL || bucket->max_entries == 0) {
+	bo_gem->bo.size = size;
+	if (bo_gem->bo.size < page_size)
+	    bo_gem->bo.size = page_size;
+    }
+
+    /* Get a buffer out of the cache if available */
+    if (bucket != NULL && bucket->num_entries > 0) {
+	struct dri_gem_bo_bucket_entry *entry = bucket->head;
+#if 0
+	int busy;
+
+	/* XXX */
+	/* Check if the buffer is still in flight.  If not, reuse it. */
+	ret = drmBOBusy(bufmgr_gem->fd, &entry->drm_bo, &busy);
+	alloc_from_cache = (ret == 0 && busy == 0);
+#else
+	alloc_from_cache = 0;
+#endif
+
+	if (alloc_from_cache) {
+	    bucket->head = entry->next;
+	    if (entry->next == NULL)
+		bucket->tail = &bucket->head;
+	    bucket->num_entries--;
+
+	    bo_gem->gem_handle = entry->gem_handle;
+	    free(entry);
+	}
+    }
+
+    if (!alloc_from_cache) {
+	struct drm_gem_alloc alloc;
+
+	memset(&alloc, 0, sizeof(alloc));
+	alloc.size = bo_gem->bo.size;
+
+	ret = ioctl(bufmgr_gem->fd, DRM_IOCTL_GEM_ALLOC, &alloc);
+	bo_gem->gem_handle = alloc.handle;
+	if (ret != 0) {
+	    free(bo_gem);
+	    return NULL;
+	}
+    }
+
+    bo_gem->bo.offset = 0;
+    bo_gem->bo.virtual = NULL;
+    bo_gem->bo.bufmgr = bufmgr;
+    bo_gem->name = name;
+    bo_gem->refcount = 1;
+    bo_gem->validate_index = -1;
+
+    DBG("bo_create: %p (%s) %ldb\n", &bo_gem->bo, bo_gem->name, size);
+
+    return &bo_gem->bo;
+}
+
+/* Our GEM backend doesn't allow creation of static buffers, as that requires
+ * privelege for the non-fake case, and the lock in the fake case where we were
+ * working around the X Server not creating buffers and passing handles to us.
+ */
+static dri_bo *
+dri_gem_alloc_static(dri_bufmgr *bufmgr, const char *name,
+		     unsigned long offset, unsigned long size, void *virtual,
+		     uint64_t location_mask)
+{
+    return NULL;
+}
+
+/**
+ * Returns a dri_bo wrapping the given buffer object handle.
+ *
+ * This can be used when one application needs to pass a buffer object
+ * to another.
+ */
+dri_bo *
+intel_gem_bo_create_from_handle(dri_bufmgr *bufmgr, const char *name,
+			      unsigned int handle)
+{
+    dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *)bufmgr;
+    dri_bo_gem *bo_gem;
+    int ret;
+    struct drm_gem_open open_arg;
+
+    bo_gem = calloc(1, sizeof(*bo_gem));
+    if (!bo_gem)
+	return NULL;
+
+    memset(&open_arg, 0, sizeof(open_arg));
+    open_arg.name = handle;
+    ret = ioctl(bufmgr_gem->fd, DRM_IOCTL_GEM_OPEN, &open_arg);
+    if (ret != 0) {
+       fprintf(stderr, "Couldn't reference %s handle 0x%08x: %s\n",
+	       name, handle, strerror(-ret));
+	free(bo_gem);
+	return NULL;
+    }
+    bo_gem->bo.size = open_arg.size;
+    bo_gem->bo.offset = 0;
+    bo_gem->bo.virtual = NULL;
+    bo_gem->bo.bufmgr = bufmgr;
+    bo_gem->name = name;
+    bo_gem->refcount = 1;
+    bo_gem->validate_index = -1;
+
+    DBG("bo_create_from_handle: %p %08x (%s)\n",
+	&bo_gem->bo, handle, bo_gem->name);
+
+    return &bo_gem->bo;
+}
+
+static void
+dri_gem_bo_reference(dri_bo *bo)
+{
+    dri_bo_gem *bo_gem = (dri_bo_gem *)bo;
+
+    bo_gem->refcount++;
+}
+
+static void
+dri_gem_bo_unreference(dri_bo *bo)
+{
+    dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *)bo->bufmgr;
+    dri_bo_gem *bo_gem = (dri_bo_gem *)bo;
+
+    if (!bo)
+	return;
+
+    if (--bo_gem->refcount == 0) {
+	struct dri_gem_bo_bucket *bucket;
+	int ret;
+
+	assert(bo_gem->map_count == 0);
+
+	if (bo_gem->relocs != NULL) {
+	    int i;
+
+	    /* Unreference all the target buffers */
+	    for (i = 0; i < bo_gem->reloc_count; i++)
+		 dri_bo_unreference(bo_gem->reloc_target_bo[i]);
+	    free(bo_gem->reloc_target_bo);
+	    free(bo_gem->relocs);
+	}
+
+	bucket = dri_gem_bo_bucket_for_size(bufmgr_gem, bo->size);
+	/* Put the buffer into our internal cache for reuse if we can. */
+	if (bucket != NULL &&
+	    (bucket->max_entries == -1 ||
+	     (bucket->max_entries > 0 &&
+	      bucket->num_entries < bucket->max_entries)))
+	{
+	    struct dri_gem_bo_bucket_entry *entry;
+
+	    entry = calloc(1, sizeof(*entry));
+	    entry->gem_handle = bo_gem->gem_handle;
+
+	    entry->next = NULL;
+	    *bucket->tail = entry;
+	    bucket->tail = &entry->next;
+	    bucket->num_entries++;
+	} else {
+	    struct drm_gem_unreference unref;
+
+	    /* Decrement the kernel refcount for the buffer. */
+	    unref.handle = bo_gem->gem_handle;
+	    ret = ioctl(bufmgr_gem->fd, DRM_IOCTL_GEM_UNREFERENCE, &unref);
+	    if (ret != 0) {
+	       fprintf(stderr, "DRM_IOCTL_GEM_UNREFERENCE failed (%s): %s\n",
+		       bo_gem->name, strerror(-ret));
+	    }
+	}
+
+	DBG("bo_unreference final: %p (%s)\n", &bo_gem->bo, bo_gem->name);
+
+	free(bo);
+	return;
+    }
+}
+
+static int
+dri_gem_bo_map(dri_bo *bo, GLboolean write_enable)
+{
+    dri_bufmgr_gem *bufmgr_gem;
+    dri_bo_gem *bo_gem = (dri_bo_gem *)bo;
+    int ret;
+
+    bufmgr_gem = (dri_bufmgr_gem *)bo->bufmgr;
+
+    /* Allow recursive mapping. Mesa may recursively map buffers with
+     * nested display loops.
+     */
+    if (bo_gem->map_count++ != 0)
+	return 0;
+
+    assert(bo->virtual == NULL);
+
+    DBG("bo_map: %p (%s)\n", &bo_gem->bo, bo_gem->name);
+
+    if (bo_gem->virtual == NULL) {
+	struct drm_gem_mmap mmap_arg;
+
+	mmap_arg.handle = bo_gem->gem_handle;
+	ret = ioctl(bufmgr_gem->fd, DRM_IOCTL_GEM_MMAP, &mmap_arg);
+	if (ret != 0) {
+	    fprintf(stderr, "%s:%d: Error mapping buffer %s: %s .\n",
+		    __FILE__, __LINE__, bo_gem->name, strerror(-ret));
+	}
+	bo_gem->virtual = (void *)(uintptr_t)mmap_arg.addr_ptr;
+    }
+
+    /* XXX Synchronization with hardware */
+
+    bo->virtual = bo_gem->virtual;
+
+    return 0;
+}
+
+static int
+dri_gem_bo_unmap(dri_bo *bo)
+{
+    dri_bufmgr_gem *bufmgr_gem;
+    dri_bo_gem *bo_gem = (dri_bo_gem *)bo;
+
+    if (bo == NULL)
+	return 0;
+
+    assert(bo_gem->map_count != 0);
+    if (--bo_gem->map_count != 0)
+	return 0;
+
+    bufmgr_gem = (dri_bufmgr_gem *)bo->bufmgr;
+
+    assert(bo->virtual != NULL);
+
+    DBG("bo_unmap: %p (%s)\n", &bo_gem->bo, bo_gem->name);
+
+    munmap(bo_gem->virtual, bo->size);
+    bo_gem->virtual = NULL;
+    bo->virtual = NULL;
+
+    return 0;
+}
+
+static void
+dri_gem_fence_reference(dri_fence *fence)
+{
+    dri_fence_gem *fence_gem = (dri_fence_gem *)fence;
+    dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *)fence->bufmgr;
+
+    ++fence_gem->refcount;
+    DBG("fence_reference: %p (%s)\n", &fence_gem->fence, fence_gem->name);
+}
+
+static void
+dri_gem_fence_unreference(dri_fence *fence)
+{
+    dri_fence_gem *fence_gem = (dri_fence_gem *)fence;
+    dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *)fence->bufmgr;
+
+    if (!fence)
+	return;
+
+    DBG("fence_unreference: %p (%s)\n", &fence_gem->fence, fence_gem->name);
+
+    if (--fence_gem->refcount == 0) {
+	int ret;
+
+	ret = drmFenceUnreference(bufmgr_gem->fd, &fence_gem->drm_fence);
+	if (ret != 0) {
+	    fprintf(stderr, "drmFenceUnreference failed (%s): %s\n",
+		    fence_gem->name, strerror(-ret));
+	}
+
+	free(fence);
+	return;
+    }
+}
+
+static void
+dri_gem_fence_wait(dri_fence *fence)
+{
+    dri_fence_gem *fence_gem = (dri_fence_gem *)fence;
+    dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *)fence->bufmgr;
+    int ret;
+
+    ret = drmFenceWait(bufmgr_gem->fd, DRM_FENCE_FLAG_WAIT_LAZY, &fence_gem->drm_fence, 0);
+    if (ret != 0) {
+        fprintf(stderr, "%s:%d: Error waiting for fence %s: %s.\n",
+		__FILE__, __LINE__, fence_gem->name, strerror(-ret));
+	abort();
+    }
+
+    DBG("fence_wait: %p (%s)\n", &fence_gem->fence, fence_gem->name);
+}
+
+static void
+dri_bufmgr_gem_destroy(dri_bufmgr *bufmgr)
+{
+    dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *)bufmgr;
+    int i;
+
+    free(bufmgr_gem->validate_array);
+    free(bufmgr_gem->validate_bo);
+
+    /* Free any cached buffer objects we were going to reuse */
+    for (i = 0; i < INTEL_GEM_BO_BUCKETS; i++) {
+	struct dri_gem_bo_bucket *bucket = &bufmgr_gem->cache_bucket[i];
+	struct dri_gem_bo_bucket_entry *entry;
+
+	while ((entry = bucket->head) != NULL) {
+	    struct drm_gem_unreference unref;
+	    int ret;
+
+	    bucket->head = entry->next;
+	    if (entry->next == NULL)
+		bucket->tail = &bucket->head;
+	    bucket->num_entries--;
+
+	    /* Decrement the kernel refcount for the buffer. */
+	    unref.handle = entry->gem_handle;
+	    ret = ioctl(bufmgr_gem->fd, DRM_IOCTL_GEM_UNREFERENCE, &unref);
+	    if (ret != 0) {
+	       fprintf(stderr, "DRM_IOCTL_GEM_UNREFERENCE failed: %s\n",
+		       strerror(-ret));
+	    }
+
+	    free(entry);
+	}
+    }
+
+    free(bufmgr);
+}
+
+/**
+ * Adds the target buffer to the validation list and adds the relocation
+ * to the reloc_buffer's relocation list.
+ *
+ * The relocation entry at the given offset must already contain the
+ * precomputed relocation value, because the kernel will optimize out
+ * the relocation entry write when the buffer hasn't moved from the
+ * last known offset in target_bo.
+ */
+static int
+dri_gem_emit_reloc(dri_bo *bo, uint64_t flags, GLuint delta,
+		   GLuint offset, dri_bo *target_bo)
+{
+    dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *)bo->bufmgr;
+    dri_bo_gem *bo_gem = (dri_bo_gem *)bo;
+    dri_bo_gem *target_bo_gem = (dri_bo_gem *)target_bo;
+
+    /* Create a new relocation list if needed */
+    if (bo_gem->relocs == NULL)
+	intel_setup_reloc_list(bo);
+
+    /* Check overflow */
+    assert(bo_gem->reloc_count < bufmgr_gem->max_relocs);
+
+    bo_gem->relocs[bo_gem->reloc_count].offset = offset;
+    bo_gem->relocs[bo_gem->reloc_count].delta = delta;
+    bo_gem->relocs[bo_gem->reloc_count].target_handle =
+	target_bo_gem->gem_handle;
+
+    bo_gem->reloc_target_bo[bo_gem->reloc_count] = target_bo;
+    dri_bo_reference(target_bo);
+
+    bo_gem->reloc_count++;
+    return 0;
+}
+
+/**
+ * Walk the tree of relocations rooted at BO and accumulate the list of
+ * validations to be performed and update the relocation buffers with
+ * index values into the validation list.
+ */
+static void
+dri_gem_bo_process_reloc(dri_bo *bo)
+{
+    dri_bo_gem *bo_gem = (dri_bo_gem *)bo;
+    int i;
+
+    if (bo_gem->relocs == NULL)
+	return;
+
+    for (i = 0; i < bo_gem->reloc_count; i++) {
+	dri_bo *target_bo = bo_gem->reloc_target_bo[i];
+
+	/* Continue walking the tree depth-first. */
+	dri_gem_bo_process_reloc(target_bo);
+
+	/* Add the target to the validate list */
+	intel_add_validate_buffer(target_bo);
+    }
+}
+
+static void *
+dri_gem_process_reloc(dri_bo *batch_buf)
+{
+    dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *)batch_buf->bufmgr;
+
+    /* Update indices and set up the validate list. */
+    dri_gem_bo_process_reloc(batch_buf);
+
+    /* Add the batch buffer to the validation list.  There are no relocations
+     * pointing to it.
+     */
+    intel_add_validate_buffer(batch_buf);
+
+    bufmgr_gem->exec_arg.buffers_ptr = (uintptr_t)bufmgr_gem->validate_array;
+    bufmgr_gem->exec_arg.buffer_count = bufmgr_gem->validate_count;
+    bufmgr_gem->exec_arg.batch_start_offset = bufmgr_gem->validate_count;
+
+    return &bufmgr_gem->exec_arg;
+}
+
+static void
+intel_update_buffer_offsets (dri_bufmgr_gem *bufmgr_gem)
+{
+    int i;
+
+    for (i = 0; i < bufmgr_gem->validate_count; i++) {
+	dri_bo *bo = bufmgr_gem->validate_bo[i];
+	dri_bo_gem *bo_gem = (dri_bo_gem *)bo;
+
+	/* Update the buffer offset */
+	if (bufmgr_gem->validate_array[i].buffer_offset != bo->offset) {
+	    DBG("BO %s migrated: 0x%08lx -> 0x%08llx\n",
+		bo_gem->name, bo->offset,
+		bufmgr_gem->validate_array[i].buffer_offset);
+	    bo->offset = bufmgr_gem->validate_array[i].buffer_offset;
+	}
+    }
+}
+
+static void
+dri_gem_post_submit(dri_bo *batch_buf, dri_fence **last_fence)
+{
+    dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *)batch_buf->bufmgr;
+    int i;
+
+    intel_update_buffer_offsets (bufmgr_gem);
+
+    if (bufmgr_gem->bufmgr.debug)
+	dri_gem_dump_validation_list(bufmgr_gem);
+
+    for (i = 0; i < bufmgr_gem->validate_count; i++) {
+	dri_bo *bo = bufmgr_gem->validate_bo[i];
+	dri_bo_gem *bo_gem = (dri_bo_gem *)bo;
+
+	/* Disconnect the buffer from the validate list */
+	bo_gem->validate_index = -1;
+	dri_bo_unreference(bo);
+	bufmgr_gem->validate_bo[i] = NULL;
+    }
+    bufmgr_gem->validate_count = 0;
+}
+
+/**
+ * Enables unlimited caching of buffer objects for reuse.
+ *
+ * This is potentially very memory expensive, as the cache at each bucket
+ * size is only bounded by how many buffers of that size we've managed to have
+ * in flight at once.
+ */
+void
+intel_gem_enable_bo_reuse(dri_bufmgr *bufmgr)
+{
+    /*
+    dri_bufmgr_gem *bufmgr_gem = (dri_bufmgr_gem *)bufmgr;
+    int i;
+
+    for (i = 0; i < INTEL_GEM_BO_BUCKETS; i++) {
+	bufmgr_gem->cache_bucket[i].max_entries = -1;
+    }
+    */
+}
+
+/*
+ *
+ */
+static int
+dri_gem_check_aperture_space(dri_bo *bo)
+{
+    return 0;
+}
+
+/**
+ * Initializes the GEM buffer manager, which uses the kernel to allocate, map,
+ * and manage map buffer objections.
+ *
+ * \param fd File descriptor of the opened DRM device.
+ * \param fence_type Driver-specific fence type used for fences with no flush.
+ * \param fence_type_flush Driver-specific fence type used for fences with a
+ *	  flush.
+ */
+dri_bufmgr *
+intel_bufmgr_gem_init(int fd, int batch_size)
+{
+    dri_bufmgr_gem *bufmgr_gem;
+    int i;
+
+    bufmgr_gem = calloc(1, sizeof(*bufmgr_gem));
+    bufmgr_gem->fd = fd;
+
+    /* Let's go with one relocation per every 2 dwords (but round down a bit
+     * since a power of two will mean an extra page allocation for the reloc
+     * buffer).
+     *
+     * Every 4 was too few for the blender benchmark.
+     */
+    bufmgr_gem->max_relocs = batch_size / sizeof(uint32_t) / 2 - 2;
+
+    bufmgr_gem->bufmgr.bo_alloc = dri_gem_alloc;
+    bufmgr_gem->bufmgr.bo_alloc_static = dri_gem_alloc_static;
+    bufmgr_gem->bufmgr.bo_reference = dri_gem_bo_reference;
+    bufmgr_gem->bufmgr.bo_unreference = dri_gem_bo_unreference;
+    bufmgr_gem->bufmgr.bo_map = dri_gem_bo_map;
+    bufmgr_gem->bufmgr.bo_unmap = dri_gem_bo_unmap;
+    bufmgr_gem->bufmgr.fence_reference = dri_gem_fence_reference;
+    bufmgr_gem->bufmgr.fence_unreference = dri_gem_fence_unreference;
+    bufmgr_gem->bufmgr.fence_wait = dri_gem_fence_wait;
+    bufmgr_gem->bufmgr.destroy = dri_bufmgr_gem_destroy;
+    bufmgr_gem->bufmgr.emit_reloc = dri_gem_emit_reloc;
+    bufmgr_gem->bufmgr.process_relocs = dri_gem_process_reloc;
+    bufmgr_gem->bufmgr.post_submit = dri_gem_post_submit;
+    bufmgr_gem->bufmgr.debug = GL_FALSE;
+    bufmgr_gem->bufmgr.check_aperture_space = dri_gem_check_aperture_space;
+    /* Initialize the linked lists for BO reuse cache. */
+    for (i = 0; i < INTEL_GEM_BO_BUCKETS; i++)
+	bufmgr_gem->cache_bucket[i].tail = &bufmgr_gem->cache_bucket[i].head;
+
+    return &bufmgr_gem->bufmgr;
+}
+
diff --git a/src/mesa/drivers/dri/intel/intel_bufmgr_gem.h b/src/mesa/drivers/dri/intel/intel_bufmgr_gem.h
new file mode 100644
index 0000000000..a28f5ae814
--- /dev/null
+++ b/src/mesa/drivers/dri/intel/intel_bufmgr_gem.h
@@ -0,0 +1,19 @@
+
+#ifndef INTEL_BUFMGR_GEM_H
+#define INTEL_BUFMGR_GEM_H
+
+#include "dri_bufmgr.h"
+
+extern dri_bo *intel_gem_bo_create_from_handle(dri_bufmgr *bufmgr, const char *name,
+					       unsigned int handle);
+
+dri_fence *intel_gem_fence_create_from_arg(dri_bufmgr *bufmgr, const char *name,
+					   drm_fence_arg_t *arg);
+
+
+dri_bufmgr *intel_bufmgr_gem_init(int fd, int batch_size);
+
+void
+intel_gem_enable_bo_reuse(dri_bufmgr *bufmgr);
+
+#endif /* INTEL_BUFMGR_GEM_H */
diff --git a/src/mesa/drivers/dri/intel/intel_ioctl.c b/src/mesa/drivers/dri/intel/intel_ioctl.c
index 66e36102b9..f9624a6abe 100644
--- a/src/mesa/drivers/dri/intel/intel_ioctl.c
+++ b/src/mesa/drivers/dri/intel/intel_ioctl.c
@@ -151,9 +151,9 @@ void
 intel_exec_ioctl(struct intel_context *intel,
 		 GLuint used,
 		 GLboolean ignore_cliprects, GLboolean allow_unlock,
-		 void *start, GLuint count, dri_fence **fence)
+		 struct drm_i915_gem_execbuffer *execbuf,
+		 dri_fence **fence)
 {
-   struct drm_i915_execbuffer execbuf;
    dri_fence *fo;
    int ret;
 
@@ -169,16 +169,13 @@ intel_exec_ioctl(struct intel_context *intel,
 
    memset(&execbuf, 0, sizeof(execbuf));
 
-   execbuf.num_buffers = count;
-   execbuf.batch.used = used;
-   execbuf.batch.cliprects = intel->pClipRects;
-   execbuf.batch.num_cliprects = ignore_cliprects ? 0 : intel->numClipRects;
-   execbuf.batch.DR1 = 0;
-   execbuf.batch.DR4 = ((((GLuint) intel->drawX) & 0xffff) |
-			(((GLuint) intel->drawY) << 16));
-
-   execbuf.ops_list = (unsigned long)start; // TODO
-   execbuf.fence_arg.flags = DRM_FENCE_FLAG_SHAREABLE | DRM_I915_FENCE_FLAG_FLUSHED;
+   execbuf->batch_start_offset = 0;
+   execbuf->batch_len = used;
+   execbuf->cliprects_ptr = (uintptr_t)intel->pClipRects;
+   execbuf->num_cliprects = ignore_cliprects ? 0 : intel->numClipRects;
+   execbuf->DR1 = 0;
+   execbuf->DR4 = ((((GLuint) intel->drawX) & 0xffff) |
+		   (((GLuint) intel->drawY) << 16));
 
    do {
       ret = drmCommandWriteRead(intel->driFd, DRM_I915_EXECBUFFER, &execbuf,
@@ -191,17 +188,6 @@ intel_exec_ioctl(struct intel_context *intel,
       exit(1);
    }
 
-   if (execbuf.fence_arg.error != 0) {
-
-      /*
-       * Fence creation has failed, but the GPU has been
-       * idled by the kernel. Safe to continue.
-       */ 
-
-      *fence = NULL;
-      return;
-   }
-
    fo = intel_ttm_fence_create_from_arg(intel->bufmgr, "fence buffers",
 					&execbuf.fence_arg);
    if (!fo) {
diff --git a/src/mesa/drivers/dri/intel/intel_ioctl.h b/src/mesa/drivers/dri/intel/intel_ioctl.h
index 8674aef723..7691a27f92 100644
--- a/src/mesa/drivers/dri/intel/intel_ioctl.h
+++ b/src/mesa/drivers/dri/intel/intel_ioctl.h
@@ -41,6 +41,7 @@ void intel_batch_ioctl( struct intel_context *intel,
 void intel_exec_ioctl(struct intel_context *intel,
 		      GLuint used,
 		      GLboolean ignore_cliprects, GLboolean allow_unlock,
-		      void *start, GLuint count, dri_fence **fence);
+		      struct drm_i915_gem_execbuffer *execbuf,
+		      dri_fence **fence);
 
 #endif