drm/i915: Generalise GPU activity tracking

We currently track GPU memory usage inside VMA, such that we never
release memory used by the GPU until after it has finished accessing it.
However, we may want to track other resources aside from VMA, or we may
want to split a VMA into multiple independent regions and track each
separately. For this purpose, generalise our request tracking (akin to
struct reservation_object) so that we can embed it into other objects.

v2: Tweak error handling during selftest setup.

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/20190205130005.2807-2-chris@chris-wilson.co.uk

1 files changed, 28 insertions, 145 deletions

diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c
index d83b8ad5f859..d4772061e642 100644
--- a/drivers/gpu/drm/i915/i915_vma.c
+++ b/drivers/gpu/drm/i915/i915_vma.c
@@ -63,22 +63,23 @@ static void vma_print_allocator(struct i915_vma *vma, const char *reason)
 
 #endif
 
-struct i915_vma_active {
-	struct i915_gem_active base;
-	struct i915_vma *vma;
-	struct rb_node node;
-	u64 timeline;
-};
+static void obj_bump_mru(struct drm_i915_gem_object *obj)
+{
+	struct drm_i915_private *i915 = to_i915(obj->base.dev);
 
-static void
-__i915_vma_retire(struct i915_vma *vma, struct i915_request *rq)
+	spin_lock(&i915->mm.obj_lock);
+	if (obj->bind_count)
+		list_move_tail(&obj->mm.link, &i915->mm.bound_list);
+	spin_unlock(&i915->mm.obj_lock);
+
+	obj->mm.dirty = true; /* be paranoid  */
+}
+
+static void __i915_vma_retire(struct i915_active *ref)
 {
+	struct i915_vma *vma = container_of(ref, typeof(*vma), active);
 	struct drm_i915_gem_object *obj = vma->obj;
 
-	GEM_BUG_ON(!i915_vma_is_active(vma));
-	if (--vma->active_count)
-		return;
-
 	GEM_BUG_ON(!i915_gem_object_is_active(obj));
 	if (--obj->active_count)
 		return;
@@ -90,16 +91,12 @@ __i915_vma_retire(struct i915_vma *vma, struct i915_request *rq)
 		reservation_object_unlock(obj->resv);
 	}
 
-	/* Bump our place on the bound list to keep it roughly in LRU order
+	/*
+	 * Bump our place on the bound list to keep it roughly in LRU order
 	 * so that we don't steal from recently used but inactive objects
 	 * (unless we are forced to ofc!)
 	 */
-	spin_lock(&rq->i915->mm.obj_lock);
-	if (obj->bind_count)
-		list_move_tail(&obj->mm.link, &rq->i915->mm.bound_list);
-	spin_unlock(&rq->i915->mm.obj_lock);
-
-	obj->mm.dirty = true; /* be paranoid  */
+	obj_bump_mru(obj);
 
 	if (i915_gem_object_has_active_reference(obj)) {
 		i915_gem_object_clear_active_reference(obj);
@@ -107,21 +104,6 @@ __i915_vma_retire(struct i915_vma *vma, struct i915_request *rq)
 	}
 }
 
-static void
-i915_vma_retire(struct i915_gem_active *base, struct i915_request *rq)
-{
-	struct i915_vma_active *active =
-		container_of(base, typeof(*active), base);
-
-	__i915_vma_retire(active->vma, rq);
-}
-
-static void
-i915_vma_last_retire(struct i915_gem_active *base, struct i915_request *rq)
-{
-	__i915_vma_retire(container_of(base, struct i915_vma, last_active), rq);
-}
-
 static struct i915_vma *
 vma_create(struct drm_i915_gem_object *obj,
 	   struct i915_address_space *vm,
@@ -137,10 +119,9 @@ vma_create(struct drm_i915_gem_object *obj,
 	if (vma == NULL)
 		return ERR_PTR(-ENOMEM);
 
-	vma->active = RB_ROOT;
-
-	init_request_active(&vma->last_active, i915_vma_last_retire);
+	i915_active_init(vm->i915, &vma->active, __i915_vma_retire);
 	init_request_active(&vma->last_fence, NULL);
+
 	vma->vm = vm;
 	vma->ops = &vm->vma_ops;
 	vma->obj = obj;
@@ -823,7 +804,6 @@ void i915_vma_reopen(struct i915_vma *vma)
 static void __i915_vma_destroy(struct i915_vma *vma)
 {
 	struct drm_i915_private *i915 = vma->vm->i915;
-	struct i915_vma_active *iter, *n;
 
 	GEM_BUG_ON(vma->node.allocated);
 	GEM_BUG_ON(vma->fence);
@@ -843,10 +823,7 @@ static void __i915_vma_destroy(struct i915_vma *vma)
 		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));
-		kfree(iter);
-	}
+	i915_active_fini(&vma->active);
 
 	kmem_cache_free(i915->vmas, vma);
 }
@@ -931,104 +908,15 @@ static void export_fence(struct i915_vma *vma,
 	reservation_object_unlock(resv);
 }
 
-static struct i915_gem_active *active_instance(struct i915_vma *vma, u64 idx)
-{
-	struct i915_vma_active *active;
-	struct rb_node **p, *parent;
-	struct i915_request *old;
-
-	/*
-	 * We track the most recently used timeline to skip a rbtree search
-	 * for the common case, under typical loads we never need the rbtree
-	 * at all. We can reuse the last_active slot if it is empty, that is
-	 * after the previous activity has been retired, or if the active
-	 * matches the current timeline.
-	 *
-	 * Note that we allow the timeline to be active simultaneously in
-	 * the rbtree and the last_active cache. We do this to avoid having
-	 * to search and replace the rbtree element for a new timeline, with
-	 * the cost being that we must be aware that the vma may be retired
-	 * twice for the same timeline (as the older rbtree element will be
-	 * retired before the new request added to last_active).
-	 */
-	old = i915_gem_active_raw(&vma->last_active,
-				  &vma->vm->i915->drm.struct_mutex);
-	if (!old || old->fence.context == idx)
-		goto out;
-
-	/* Move the currently active fence into the rbtree */
-	idx = old->fence.context;
-
-	parent = NULL;
-	p = &vma->active.rb_node;
-	while (*p) {
-		parent = *p;
-
-		active = rb_entry(parent, struct i915_vma_active, node);
-		if (active->timeline == idx)
-			goto replace;
-
-		if (active->timeline < idx)
-			p = &parent->rb_right;
-		else
-			p = &parent->rb_left;
-	}
-
-	active = kmalloc(sizeof(*active), GFP_KERNEL);
-
-	/* kmalloc may retire the vma->last_active request (thanks shrinker)! */
-	if (unlikely(!i915_gem_active_raw(&vma->last_active,
-					  &vma->vm->i915->drm.struct_mutex))) {
-		kfree(active);
-		goto out;
-	}
-
-	if (unlikely(!active))
-		return ERR_PTR(-ENOMEM);
-
-	init_request_active(&active->base, i915_vma_retire);
-	active->vma = vma;
-	active->timeline = idx;
-
-	rb_link_node(&active->node, parent, p);
-	rb_insert_color(&active->node, &vma->active);
-
-replace:
-	/*
-	 * Overwrite the previous active slot in the rbtree with last_active,
-	 * leaving last_active zeroed. If the previous slot is still active,
-	 * we must be careful as we now only expect to receive one retire
-	 * callback not two, and so much undo the active counting for the
-	 * overwritten slot.
-	 */
-	if (i915_gem_active_isset(&active->base)) {
-		/* Retire ourselves from the old rq->active_list */
-		__list_del_entry(&active->base.link);
-		vma->active_count--;
-		GEM_BUG_ON(!vma->active_count);
-	}
-	GEM_BUG_ON(list_empty(&vma->last_active.link));
-	list_replace_init(&vma->last_active.link, &active->base.link);
-	active->base.request = fetch_and_zero(&vma->last_active.request);
-
-out:
-	return &vma->last_active;
-}
-
 int i915_vma_move_to_active(struct i915_vma *vma,
 			    struct i915_request *rq,
 			    unsigned int flags)
 {
 	struct drm_i915_gem_object *obj = vma->obj;
-	struct i915_gem_active *active;
 
 	lockdep_assert_held(&rq->i915->drm.struct_mutex);
 	GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
 
-	active = active_instance(vma, rq->fence.context);
-	if (IS_ERR(active))
-		return PTR_ERR(active);
-
 	/*
 	 * Add a reference if we're newly entering the active list.
 	 * The order in which we add operations to the retirement queue is
@@ -1037,9 +925,15 @@ int i915_vma_move_to_active(struct i915_vma *vma,
 	 * add the active reference first and queue for it to be dropped
 	 * *last*.
 	 */
-	if (!i915_gem_active_isset(active) && !vma->active_count++)
+	if (!vma->active.count)
 		obj->active_count++;
-	i915_gem_active_set(active, rq);
+
+	if (unlikely(i915_active_ref(&vma->active, rq->fence.context, rq))) {
+		if (!vma->active.count)
+			obj->active_count--;
+		return -ENOMEM;
+	}
+
 	GEM_BUG_ON(!i915_vma_is_active(vma));
 	GEM_BUG_ON(!obj->active_count);
 
@@ -1073,8 +967,6 @@ int i915_vma_unbind(struct i915_vma *vma)
 	 */
 	might_sleep();
 	if (i915_vma_is_active(vma)) {
-		struct i915_vma_active *active, *n;
-
 		/*
 		 * When a closed VMA is retired, it is unbound - eek.
 		 * In order to prevent it from being recursively closed,
@@ -1090,19 +982,10 @@ int i915_vma_unbind(struct i915_vma *vma)
 		 */
 		__i915_vma_pin(vma);
 
-		ret = i915_gem_active_retire(&vma->last_active,
-					     &vma->vm->i915->drm.struct_mutex);
+		ret = i915_active_wait(&vma->active);
 		if (ret)
 			goto unpin;
 
-		rbtree_postorder_for_each_entry_safe(active, n,
-						     &vma->active, node) {
-			ret = i915_gem_active_retire(&active->base,
-						     &vma->vm->i915->drm.struct_mutex);
-			if (ret)
-				goto unpin;
-		}
-
 		ret = i915_gem_active_retire(&vma->last_fence,
 					     &vma->vm->i915->drm.struct_mutex);
 unpin: