]> git.karo-electronics.de Git - karo-tx-linux.git/commitdiff
drm/i915/execbuffer: Reorder binding of objects to favour restrictions
authorChris Wilson <chris@chris-wilson.co.uk>
Mon, 10 Jan 2011 17:35:37 +0000 (17:35 +0000)
committerChris Wilson <chris@chris-wilson.co.uk>
Tue, 11 Jan 2011 22:55:48 +0000 (22:55 +0000)
As the mappable portion of the aperture is always a small subset at the
start of the GTT, it is allocated preferentially by drm_mm. This is
useful in case we ever need to map an object later. However, if you have
a large object that can consume the entire mappable region of the
GTT this prevents the batchbuffer from fitting and so causing an error.
Instead allocate all those that require a mapping up front in order to
improve the likelihood of finding sufficient space to bind them.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/i915_gem_execbuffer.c

index 6130f77c26bf9db7651823b91b4ea0d02cebcb8b..385fc7ec39d3c987f3b4cc4c420922382da0ca69 100644 (file)
@@ -796,6 +796,7 @@ struct drm_i915_gem_object {
         */
        struct hlist_node exec_node;
        unsigned long exec_handle;
+       struct drm_i915_gem_exec_object2 *exec_entry;
 
        /**
         * Current offset of the object in GTT space.
index 0445770cc23c1cfedf5dfe8ceef0de36744d565e..e69834341ef068b4ebbbb500fc4484d64826a879 100644 (file)
@@ -268,7 +268,6 @@ eb_destroy(struct eb_objects *eb)
 static int
 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
                                   struct eb_objects *eb,
-                                  struct drm_i915_gem_exec_object2 *entry,
                                   struct drm_i915_gem_relocation_entry *reloc)
 {
        struct drm_device *dev = obj->base.dev;
@@ -411,10 +410,10 @@ i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
 
 static int
 i915_gem_execbuffer_relocate_object(struct drm_i915_gem_object *obj,
-                                   struct eb_objects *eb,
-                                   struct drm_i915_gem_exec_object2 *entry)
+                                   struct eb_objects *eb)
 {
        struct drm_i915_gem_relocation_entry __user *user_relocs;
+       struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
        int i, ret;
 
        user_relocs = (void __user *)(uintptr_t)entry->relocs_ptr;
@@ -426,7 +425,7 @@ i915_gem_execbuffer_relocate_object(struct drm_i915_gem_object *obj,
                                              sizeof(reloc)))
                        return -EFAULT;
 
-               ret = i915_gem_execbuffer_relocate_entry(obj, eb, entry, &reloc);
+               ret = i915_gem_execbuffer_relocate_entry(obj, eb, &reloc);
                if (ret)
                        return ret;
 
@@ -442,13 +441,13 @@ i915_gem_execbuffer_relocate_object(struct drm_i915_gem_object *obj,
 static int
 i915_gem_execbuffer_relocate_object_slow(struct drm_i915_gem_object *obj,
                                         struct eb_objects *eb,
-                                        struct drm_i915_gem_exec_object2 *entry,
                                         struct drm_i915_gem_relocation_entry *relocs)
 {
+       const struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
        int i, ret;
 
        for (i = 0; i < entry->relocation_count; i++) {
-               ret = i915_gem_execbuffer_relocate_entry(obj, eb, entry, &relocs[i]);
+               ret = i915_gem_execbuffer_relocate_entry(obj, eb, &relocs[i]);
                if (ret)
                        return ret;
        }
@@ -459,8 +458,7 @@ i915_gem_execbuffer_relocate_object_slow(struct drm_i915_gem_object *obj,
 static int
 i915_gem_execbuffer_relocate(struct drm_device *dev,
                             struct eb_objects *eb,
-                            struct list_head *objects,
-                            struct drm_i915_gem_exec_object2 *exec)
+                            struct list_head *objects)
 {
        struct drm_i915_gem_object *obj;
        int ret;
@@ -468,7 +466,7 @@ i915_gem_execbuffer_relocate(struct drm_device *dev,
        list_for_each_entry(obj, objects, exec_list) {
                obj->base.pending_read_domains = 0;
                obj->base.pending_write_domain = 0;
-               ret = i915_gem_execbuffer_relocate_object(obj, eb, exec++);
+               ret = i915_gem_execbuffer_relocate_object(obj, eb);
                if (ret)
                        return ret;
        }
@@ -479,13 +477,36 @@ i915_gem_execbuffer_relocate(struct drm_device *dev,
 static int
 i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
                            struct drm_file *file,
-                           struct list_head *objects,
-                           struct drm_i915_gem_exec_object2 *exec)
+                           struct list_head *objects)
 {
        struct drm_i915_gem_object *obj;
-       struct drm_i915_gem_exec_object2 *entry;
        int ret, retry;
        bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
+       struct list_head ordered_objects;
+
+       INIT_LIST_HEAD(&ordered_objects);
+       while (!list_empty(objects)) {
+               struct drm_i915_gem_exec_object2 *entry;
+               bool need_fence, need_mappable;
+
+               obj = list_first_entry(objects,
+                                      struct drm_i915_gem_object,
+                                      exec_list);
+               entry = obj->exec_entry;
+
+               need_fence =
+                       has_fenced_gpu_access &&
+                       entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
+                       obj->tiling_mode != I915_TILING_NONE;
+               need_mappable =
+                       entry->relocation_count ? true : need_fence;
+
+               if (need_mappable)
+                       list_move(&obj->exec_list, &ordered_objects);
+               else
+                       list_move_tail(&obj->exec_list, &ordered_objects);
+       }
+       list_splice(&ordered_objects, objects);
 
        /* Attempt to pin all of the buffers into the GTT.
         * This is done in 3 phases:
@@ -504,14 +525,11 @@ i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
                ret = 0;
 
                /* Unbind any ill-fitting objects or pin. */
-               entry = exec;
                list_for_each_entry(obj, objects, exec_list) {
+                       struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
                        bool need_fence, need_mappable;
-
-                       if (!obj->gtt_space) {
-                               entry++;
+                       if (!obj->gtt_space)
                                continue;
-                       }
 
                        need_fence =
                                has_fenced_gpu_access &&
@@ -534,8 +552,8 @@ i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
                }
 
                /* Bind fresh objects */
-               entry = exec;
                list_for_each_entry(obj, objects, exec_list) {
+                       struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
                        bool need_fence;
 
                        need_fence =
@@ -570,7 +588,6 @@ i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
                        }
 
                        entry->offset = obj->gtt_offset;
-                       entry++;
                }
 
                /* Decrement pin count for bound objects */
@@ -680,10 +697,11 @@ i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
 
                list_add_tail(&obj->exec_list, objects);
                obj->exec_handle = exec[i].handle;
+               obj->exec_entry = &exec[i];
                eb_add_object(eb, obj);
        }
 
-       ret = i915_gem_execbuffer_reserve(ring, file, objects, exec);
+       ret = i915_gem_execbuffer_reserve(ring, file, objects);
        if (ret)
                goto err;
 
@@ -692,7 +710,6 @@ i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
                obj->base.pending_read_domains = 0;
                obj->base.pending_write_domain = 0;
                ret = i915_gem_execbuffer_relocate_object_slow(obj, eb,
-                                                              exec,
                                                               reloc + total);
                if (ret)
                        goto err;
@@ -1110,16 +1127,22 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
 
                list_add_tail(&obj->exec_list, &objects);
                obj->exec_handle = exec[i].handle;
+               obj->exec_entry = &exec[i];
                eb_add_object(eb, obj);
        }
 
+       /* take note of the batch buffer before we might reorder the lists */
+       batch_obj = list_entry(objects.prev,
+                              struct drm_i915_gem_object,
+                              exec_list);
+
        /* Move the objects en-masse into the GTT, evicting if necessary. */
-       ret = i915_gem_execbuffer_reserve(ring, file, &objects, exec);
+       ret = i915_gem_execbuffer_reserve(ring, file, &objects);
        if (ret)
                goto err;
 
        /* The objects are in their final locations, apply the relocations. */
-       ret = i915_gem_execbuffer_relocate(dev, eb, &objects, exec);
+       ret = i915_gem_execbuffer_relocate(dev, eb, &objects);
        if (ret) {
                if (ret == -EFAULT) {
                        ret = i915_gem_execbuffer_relocate_slow(dev, file, ring,
@@ -1133,9 +1156,6 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
        }
 
        /* Set the pending read domains for the batch buffer to COMMAND */
-       batch_obj = list_entry(objects.prev,
-                              struct drm_i915_gem_object,
-                              exec_list);
        if (batch_obj->base.pending_write_domain) {
                DRM_ERROR("Attempting to use self-modifying batch buffer\n");
                ret = -EINVAL;