Commit 8b293eb5 authored by Chris Wilson's avatar Chris Wilson

drm/i915: Track the last-active inside the i915_vma

Using a VMA on more than one timeline concurrently is the exception
rather than the rule (using it concurrently on multiple engines). As we
expect to only use one active tracker, store the most recently used
tracker inside the i915_vma itself and only fallback to the rbtree if
we need a second or more concurrent active trackers.

v2: Comments on how we overwrite any existing last_active cache.
v3: __list_del_entry() before list_replace_init() is confusing and, much
more important, entirely redundant.
v4: Note that both last_active and the rbtree may be simultaneously
tracking this timeline, albeit with different requests, and so the vma
may be retired twice for the same timeline.
v5: No, that list_del is required!
Signed-off-by: default avatarChris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: default avatarTvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180706123157.9645-1-chris@chris-wilson.co.uk
parent 5c3f8c22
...@@ -119,6 +119,12 @@ i915_vma_retire(struct i915_gem_active *base, struct i915_request *rq) ...@@ -119,6 +119,12 @@ i915_vma_retire(struct i915_gem_active *base, struct i915_request *rq)
__i915_vma_retire(active->vma, rq); __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 * static struct i915_vma *
vma_create(struct drm_i915_gem_object *obj, vma_create(struct drm_i915_gem_object *obj,
struct i915_address_space *vm, struct i915_address_space *vm,
...@@ -136,6 +142,7 @@ vma_create(struct drm_i915_gem_object *obj, ...@@ -136,6 +142,7 @@ vma_create(struct drm_i915_gem_object *obj,
vma->active = RB_ROOT; vma->active = RB_ROOT;
init_request_active(&vma->last_active, i915_vma_last_retire);
init_request_active(&vma->last_fence, NULL); init_request_active(&vma->last_fence, NULL);
vma->vm = vm; vma->vm = vm;
vma->ops = &vm->vma_ops; vma->ops = &vm->vma_ops;
...@@ -895,6 +902,29 @@ static struct i915_gem_active *active_instance(struct i915_vma *vma, u64 idx) ...@@ -895,6 +902,29 @@ static struct i915_gem_active *active_instance(struct i915_vma *vma, u64 idx)
{ {
struct i915_vma_active *active; struct i915_vma_active *active;
struct rb_node **p, *parent; 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; parent = NULL;
p = &vma->active.rb_node; p = &vma->active.rb_node;
...@@ -903,7 +933,7 @@ static struct i915_gem_active *active_instance(struct i915_vma *vma, u64 idx) ...@@ -903,7 +933,7 @@ static struct i915_gem_active *active_instance(struct i915_vma *vma, u64 idx)
active = rb_entry(parent, struct i915_vma_active, node); active = rb_entry(parent, struct i915_vma_active, node);
if (active->timeline == idx) if (active->timeline == idx)
return &active->base; goto replace;
if (active->timeline < idx) if (active->timeline < idx)
p = &parent->rb_right; p = &parent->rb_right;
...@@ -922,7 +952,26 @@ static struct i915_gem_active *active_instance(struct i915_vma *vma, u64 idx) ...@@ -922,7 +952,26 @@ static struct i915_gem_active *active_instance(struct i915_vma *vma, u64 idx)
rb_link_node(&active->node, parent, p); rb_link_node(&active->node, parent, p);
rb_insert_color(&active->node, &vma->active); rb_insert_color(&active->node, &vma->active);
return &active->base; 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, int i915_vma_move_to_active(struct i915_vma *vma,
...@@ -1002,6 +1051,11 @@ int i915_vma_unbind(struct i915_vma *vma) ...@@ -1002,6 +1051,11 @@ int i915_vma_unbind(struct i915_vma *vma)
*/ */
__i915_vma_pin(vma); __i915_vma_pin(vma);
ret = i915_gem_active_retire(&vma->last_active,
&vma->vm->i915->drm.struct_mutex);
if (ret)
goto unpin;
rbtree_postorder_for_each_entry_safe(active, n, rbtree_postorder_for_each_entry_safe(active, n,
&vma->active, node) { &vma->active, node) {
ret = i915_gem_active_retire(&active->base, ret = i915_gem_active_retire(&active->base,
......
...@@ -97,6 +97,7 @@ struct i915_vma { ...@@ -97,6 +97,7 @@ struct i915_vma {
unsigned int active_count; unsigned int active_count;
struct rb_root active; struct rb_root active;
struct i915_gem_active last_active;
struct i915_gem_active last_fence; struct i915_gem_active last_fence;
/** /**
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment