drm/i915: Move vma lookup to its own lock

Remove the struct_mutex requirement for looking up the vma for an
object.

v2: Highlight how the race for duplicate vma creation is resolved on
reacquiring the lock with a short comment.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190128102356.15037-3-chris@chris-wilson.co.uk

drivers/gpu/drm/i915/i915_debugfs.c

@@ -159,14 +159,14 @@ describe_obj(struct seq_file *m, struct drm_i915_gem_object *obj)
 		   obj->mm.madv == I915_MADV_DONTNEED ? " purgeable" : "");
 	if (obj->base.name)
 		seq_printf(m, " (name: %d)", obj->base.name);
-	list_for_each_entry(vma, &obj->vma_list, obj_link) {
+	list_for_each_entry(vma, &obj->vma.list, obj_link) {
 		if (i915_vma_is_pinned(vma))
 			pin_count++;
 	}
 	seq_printf(m, " (pinned x %d)", pin_count);
 	if (obj->pin_global)
 		seq_printf(m, " (global)");
-	list_for_each_entry(vma, &obj->vma_list, obj_link) {
+	list_for_each_entry(vma, &obj->vma.list, obj_link) {
 		if (!drm_mm_node_allocated(&vma->node))
 			continue;
 
@@ -322,7 +322,7 @@ static int per_file_stats(int id, void *ptr, void *data)
 	if (obj->base.name || obj->base.dma_buf)
 		stats->shared += obj->base.size;
 
-	list_for_each_entry(vma, &obj->vma_list, obj_link) {
+	list_for_each_entry(vma, &obj->vma.list, obj_link) {
 		if (!drm_mm_node_allocated(&vma->node))
 			continue;
 

drivers/gpu/drm/i915/i915_gem.c

@@ -437,15 +437,19 @@ int i915_gem_object_unbind(struct drm_i915_gem_object *obj)
 	if (ret)
 		return ret;
 
-	while ((vma = list_first_entry_or_null(&obj->vma_list,
+	spin_lock(&obj->vma.lock);
+	while (!ret && (vma = list_first_entry_or_null(&obj->vma.list,
 						       struct i915_vma,
 						       obj_link))) {
 		list_move_tail(&vma->obj_link, &still_in_list);
+		spin_unlock(&obj->vma.lock);
+
 		ret = i915_vma_unbind(vma);
-		if (ret)
-			break;
+
+		spin_lock(&obj->vma.lock);
 	}
-	list_splice(&still_in_list, &obj->vma_list);
+	list_splice(&still_in_list, &obj->vma.list);
+	spin_unlock(&obj->vma.lock);
 
 	return ret;
 }
@@ -3489,7 +3493,7 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
 	 * reading an invalid PTE on older architectures.
 	 */
 restart:
-	list_for_each_entry(vma, &obj->vma_list, obj_link) {
+	list_for_each_entry(vma, &obj->vma.list, obj_link) {
 		if (!drm_mm_node_allocated(&vma->node))
 			continue;
 
@@ -3567,7 +3571,7 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
 			 */
 		}
 
-		list_for_each_entry(vma, &obj->vma_list, obj_link) {
+		list_for_each_entry(vma, &obj->vma.list, obj_link) {
 			if (!drm_mm_node_allocated(&vma->node))
 				continue;
 
@@ -3577,7 +3581,7 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,
 		}
 	}
 
-	list_for_each_entry(vma, &obj->vma_list, obj_link)
+	list_for_each_entry(vma, &obj->vma.list, obj_link)
 		vma->node.color = cache_level;
 	i915_gem_object_set_cache_coherency(obj, cache_level);
 	obj->cache_dirty = true; /* Always invalidate stale cachelines */
@@ -4153,7 +4157,9 @@ void i915_gem_object_init(struct drm_i915_gem_object *obj,
 {
 	mutex_init(&obj->mm.lock);
 
-	INIT_LIST_HEAD(&obj->vma_list);
+	spin_lock_init(&obj->vma.lock);
+	INIT_LIST_HEAD(&obj->vma.list);
+
 	INIT_LIST_HEAD(&obj->lut_list);
 	INIT_LIST_HEAD(&obj->batch_pool_link);
 
@@ -4319,14 +4325,13 @@ static void __i915_gem_free_objects(struct drm_i915_private *i915,
 		mutex_lock(&i915->drm.struct_mutex);
 
 		GEM_BUG_ON(i915_gem_object_is_active(obj));
-		list_for_each_entry_safe(vma, vn,
-					 &obj->vma_list, obj_link) {
+		list_for_each_entry_safe(vma, vn, &obj->vma.list, obj_link) {
 			GEM_BUG_ON(i915_vma_is_active(vma));
 			vma->flags &= ~I915_VMA_PIN_MASK;
 			i915_vma_destroy(vma);
 		}
-		GEM_BUG_ON(!list_empty(&obj->vma_list));
-		GEM_BUG_ON(!RB_EMPTY_ROOT(&obj->vma_tree));
+		GEM_BUG_ON(!list_empty(&obj->vma.list));
+		GEM_BUG_ON(!RB_EMPTY_ROOT(&obj->vma.tree));
 
 		/* This serializes freeing with the shrinker. Since the free
 		 * is delayed, first by RCU then by the workqueue, we want the

drivers/gpu/drm/i915/i915_gem_object.h

@@ -87,24 +87,33 @@ struct drm_i915_gem_object {
 
 	const struct drm_i915_gem_object_ops *ops;
 
+	struct {
 		/**
-	 * @vma_list: List of VMAs backed by this object
+		 * @vma.lock: protect the list/tree of vmas
+		 */
+		spinlock_t lock;
+
+		/**
+		 * @vma.list: List of VMAs backed by this object
 		 *
-	 * The VMA on this list are ordered by type, all GGTT vma are placed
-	 * at the head and all ppGTT vma are placed at the tail. The different
-	 * types of GGTT vma are unordered between themselves, use the
-	 * @vma_tree (which has a defined order between all VMA) to find an
-	 * exact match.
+		 * The VMA on this list are ordered by type, all GGTT vma are
+		 * placed at the head and all ppGTT vma are placed at the tail.
+		 * The different types of GGTT vma are unordered between
+		 * themselves, use the @vma.tree (which has a defined order
+		 * between all VMA) to quickly find an exact match.
 		 */
-	struct list_head vma_list;
+		struct list_head list;
+
 		/**
-	 * @vma_tree: Ordered tree of VMAs backed by this object
+		 * @vma.tree: Ordered tree of VMAs backed by this object
 		 *
-	 * All VMA created for this object are placed in the @vma_tree for
-	 * fast retrieval via a binary search in i915_vma_instance().
-	 * They are also added to @vma_list for easy iteration.
+		 * All VMA created for this object are placed in the @vma.tree
+		 * for fast retrieval via a binary search in
+		 * i915_vma_instance(). They are also added to @vma.list for
+		 * easy iteration.
 		 */
-	struct rb_root vma_tree;
+		struct rb_root tree;
+	} vma;
 
 	/**
 	 * @lut_list: List of vma lookup entries in use for this object.

drivers/gpu/drm/i915/i915_vma.c

@@ -187,32 +187,52 @@ vma_create(struct drm_i915_gem_object *obj,
 								i915_gem_object_get_stride(obj));
 		GEM_BUG_ON(!is_power_of_2(vma->fence_alignment));
 
-		/*
-		 * We put the GGTT vma at the start of the vma-list, followed
-		 * by the ppGGTT vma. This allows us to break early when
-		 * iterating over only the GGTT vma for an object, see
-		 * for_each_ggtt_vma()
-		 */
 		vma->flags |= I915_VMA_GGTT;
-		list_add(&vma->obj_link, &obj->vma_list);
-	} else {
-		list_add_tail(&vma->obj_link, &obj->vma_list);
 	}
 
+	spin_lock(&obj->vma.lock);
+
 	rb = NULL;
-	p = &obj->vma_tree.rb_node;
+	p = &obj->vma.tree.rb_node;
 	while (*p) {
 		struct i915_vma *pos;
+		long cmp;
 
 		rb = *p;
 		pos = rb_entry(rb, struct i915_vma, obj_node);
-		if (i915_vma_compare(pos, vm, view) < 0)
+
+		/*
+		 * If the view already exists in the tree, another thread
+		 * already created a matching vma, so return the older instance
+		 * and dispose of ours.
+		 */
+		cmp = i915_vma_compare(pos, vm, view);
+		if (cmp == 0) {
+			spin_unlock(&obj->vma.lock);
+			kmem_cache_free(vm->i915->vmas, vma);
+			return pos;
+		}
+
+		if (cmp < 0)
 			p = &rb->rb_right;
 		else
 			p = &rb->rb_left;
 	}
 	rb_link_node(&vma->obj_node, rb, p);
-	rb_insert_color(&vma->obj_node, &obj->vma_tree);
+	rb_insert_color(&vma->obj_node, &obj->vma.tree);
+
+	if (i915_vma_is_ggtt(vma))
+		/*
+		 * We put the GGTT vma at the start of the vma-list, followed
+		 * by the ppGGTT vma. This allows us to break early when
+		 * iterating over only the GGTT vma for an object, see
+		 * for_each_ggtt_vma()
+		 */
+		list_add(&vma->obj_link, &obj->vma.list);
+	else
+		list_add_tail(&vma->obj_link, &obj->vma.list);
+
+	spin_unlock(&obj->vma.lock);
 
 	mutex_lock(&vm->mutex);
 	list_add(&vma->vm_link, &vm->unbound_list);
@@ -232,7 +252,7 @@ vma_lookup(struct drm_i915_gem_object *obj,
 {
 	struct rb_node *rb;
 
-	rb = obj->vma_tree.rb_node;
+	rb = obj->vma.tree.rb_node;
 	while (rb) {
 		struct i915_vma *vma = rb_entry(rb, struct i915_vma, obj_node);
 		long cmp;
@@ -272,16 +292,18 @@ i915_vma_instance(struct drm_i915_gem_object *obj,
 {
 	struct i915_vma *vma;
 
-	lockdep_assert_held(&obj->base.dev->struct_mutex);
 	GEM_BUG_ON(view && !i915_is_ggtt(vm));
 	GEM_BUG_ON(vm->closed);
 
+	spin_lock(&obj->vma.lock);
 	vma = vma_lookup(obj, vm, view);
-	if (!vma)
+	spin_unlock(&obj->vma.lock);
+
+	/* vma_create() will resolve the race if another creates the vma */
+	if (unlikely(!vma))
 		vma = vma_create(obj, vm, view);
 
 	GEM_BUG_ON(!IS_ERR(vma) && i915_vma_compare(vma, vm, view));
-	GEM_BUG_ON(!IS_ERR(vma) && vma_lookup(obj, vm, view) != vma);
 	return vma;
 }
 
@@ -808,14 +830,18 @@ static void __i915_vma_destroy(struct i915_vma *vma)
 
 	GEM_BUG_ON(i915_gem_active_isset(&vma->last_fence));
 
-	list_del(&vma->obj_link);
-
 	mutex_lock(&vma->vm->mutex);
 	list_del(&vma->vm_link);
 	mutex_unlock(&vma->vm->mutex);
 
-	if (vma->obj)
-		rb_erase(&vma->obj_node, &vma->obj->vma_tree);
+	if (vma->obj) {
+		struct drm_i915_gem_object *obj = vma->obj;
+
+		spin_lock(&obj->vma.lock);
+		list_del(&vma->obj_link);
+		rb_erase(&vma->obj_node, &vma->obj->vma.tree);
+		spin_unlock(&obj->vma.lock);
+	}
 
 	rbtree_postorder_for_each_entry_safe(iter, n, &vma->active, node) {
 		GEM_BUG_ON(i915_gem_active_isset(&iter->base));

drivers/gpu/drm/i915/i915_vma.h

@@ -425,7 +425,7 @@ void i915_vma_parked(struct drm_i915_private *i915);
  * or the list is empty ofc.
  */
 #define for_each_ggtt_vma(V, OBJ) \
-	list_for_each_entry(V, &(OBJ)->vma_list, obj_link)		\
+	list_for_each_entry(V, &(OBJ)->vma.list, obj_link)		\
 		for_each_until(!i915_vma_is_ggtt(V))
 
 #endif

drivers/gpu/drm/i915/selftests/i915_vma.c

@@ -672,7 +672,7 @@ static int igt_vma_partial(void *arg)
 		}
 
 		count = 0;
-		list_for_each_entry(vma, &obj->vma_list, obj_link)
+		list_for_each_entry(vma, &obj->vma.list, obj_link)
 			count++;
 		if (count != nvma) {
 			pr_err("(%s) All partial vma were not recorded on the obj->vma_list: found %u, expected %u\n",
@@ -701,7 +701,7 @@ static int igt_vma_partial(void *arg)
 		i915_vma_unpin(vma);
 
 		count = 0;
-		list_for_each_entry(vma, &obj->vma_list, obj_link)
+		list_for_each_entry(vma, &obj->vma.list, obj_link)
 			count++;
 		if (count != nvma) {
 			pr_err("(%s) allocated an extra full vma!\n", p->name);