drm/i915/scheduler: Record all dependencies upon request construction

The scheduler needs to know the dependencies of each request for the
lifetime of the request, as it may choose to reschedule the requests at
any time and must ensure the dependency tree is not broken. This is in
additional to using the fence to only allow execution after all
dependencies have been completed.

One option was to extend the fence to support the bidirectional
dependency tracking required by the scheduler. However the mismatch in
lifetimes between the submit fence and the request essentially meant
that we had to build a completely separate struct (and we could not
simply reuse the existing waitqueue in the fence for one half of the
dependency tracking). The extra dependency tracking simply did not mesh
well with the fence, and keeping it separate both keeps the fence
implementation simpler and allows us to extend the dependency tracking
into a priority tree (whilst maintaining support for reordering the
tree).

To avoid the additional allocations and list manipulations, the use of
the priotree is disabled when there are no schedulers to use it.

v2: Create a dedicated slab for i915_dependency.
    Rename the lists.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20161114204105.29171-7-chris@chris-wilson.co.uk

drivers/gpu/drm/i915/i915_drv.h

@@ -1778,6 +1778,7 @@ struct drm_i915_private {
 	struct kmem_cache *objects;
 	struct kmem_cache *vmas;
 	struct kmem_cache *requests;
+	struct kmem_cache *dependencies;
 
 	const struct intel_device_info info;
 

drivers/gpu/drm/i915/i915_gem.c

@@ -4434,12 +4434,18 @@ i915_gem_load_init(struct drm_device *dev)
 	if (!dev_priv->requests)
 		goto err_vmas;
 
+	dev_priv->dependencies = KMEM_CACHE(i915_dependency,
+					    SLAB_HWCACHE_ALIGN |
+					    SLAB_RECLAIM_ACCOUNT);
+	if (!dev_priv->dependencies)
+		goto err_requests;
+
 	mutex_lock(&dev_priv->drm.struct_mutex);
 	INIT_LIST_HEAD(&dev_priv->gt.timelines);
 	err = i915_gem_timeline_init__global(dev_priv);
 	mutex_unlock(&dev_priv->drm.struct_mutex);
 	if (err)
-		goto err_requests;
+		goto err_dependencies;
 
 	INIT_LIST_HEAD(&dev_priv->context_list);
 	INIT_WORK(&dev_priv->mm.free_work, __i915_gem_free_work);
@@ -4467,6 +4473,8 @@ i915_gem_load_init(struct drm_device *dev)
 
 	return 0;
 
+err_dependencies:
+	kmem_cache_destroy(dev_priv->dependencies);
 err_requests:
 	kmem_cache_destroy(dev_priv->requests);
 err_vmas:
@@ -4483,6 +4491,7 @@ void i915_gem_load_cleanup(struct drm_device *dev)
 
 	WARN_ON(!llist_empty(&dev_priv->mm.free_list));
 
+	kmem_cache_destroy(dev_priv->dependencies);
 	kmem_cache_destroy(dev_priv->requests);
 	kmem_cache_destroy(dev_priv->vmas);
 	kmem_cache_destroy(dev_priv->objects);

drivers/gpu/drm/i915/i915_gem_request.c

@@ -113,6 +113,77 @@ i915_gem_request_remove_from_client(struct drm_i915_gem_request *request)
 	spin_unlock(&file_priv->mm.lock);
 }
 
+static struct i915_dependency *
+i915_dependency_alloc(struct drm_i915_private *i915)
+{
+	return kmem_cache_alloc(i915->dependencies, GFP_KERNEL);
+}
+
+static void
+i915_dependency_free(struct drm_i915_private *i915,
+		     struct i915_dependency *dep)
+{
+	kmem_cache_free(i915->dependencies, dep);
+}
+
+static void
+__i915_priotree_add_dependency(struct i915_priotree *pt,
+			       struct i915_priotree *signal,
+			       struct i915_dependency *dep,
+			       unsigned long flags)
+{
+	list_add(&dep->wait_link, &signal->waiters_list);
+	list_add(&dep->signal_link, &pt->signalers_list);
+	dep->signaler = signal;
+	dep->flags = flags;
+}
+
+static int
+i915_priotree_add_dependency(struct drm_i915_private *i915,
+			     struct i915_priotree *pt,
+			     struct i915_priotree *signal)
+{
+	struct i915_dependency *dep;
+
+	dep = i915_dependency_alloc(i915);
+	if (!dep)
+		return -ENOMEM;
+
+	__i915_priotree_add_dependency(pt, signal, dep, I915_DEPENDENCY_ALLOC);
+	return 0;
+}
+
+static void
+i915_priotree_fini(struct drm_i915_private *i915, struct i915_priotree *pt)
+{
+	struct i915_dependency *dep, *next;
+
+	/* Everyone we depended upon (the fences we wait to be signaled)
+	 * should retire before us and remove themselves from our list.
+	 * However, retirement is run independently on each timeline and
+	 * so we may be called out-of-order.
+	 */
+	list_for_each_entry_safe(dep, next, &pt->signalers_list, signal_link) {
+		list_del(&dep->wait_link);
+		if (dep->flags & I915_DEPENDENCY_ALLOC)
+			i915_dependency_free(i915, dep);
+	}
+
+	/* Remove ourselves from everyone who depends upon us */
+	list_for_each_entry_safe(dep, next, &pt->waiters_list, wait_link) {
+		list_del(&dep->signal_link);
+		if (dep->flags & I915_DEPENDENCY_ALLOC)
+			i915_dependency_free(i915, dep);
+	}
+}
+
+static void
+i915_priotree_init(struct i915_priotree *pt)
+{
+	INIT_LIST_HEAD(&pt->signalers_list);
+	INIT_LIST_HEAD(&pt->waiters_list);
+}
+
 void i915_gem_retire_noop(struct i915_gem_active *active,
 			  struct drm_i915_gem_request *request)
 {
@@ -182,6 +253,8 @@ static void i915_gem_request_retire(struct drm_i915_gem_request *request)
 	i915_gem_context_put(request->ctx);
 
 	dma_fence_signal(&request->fence);
+
+	i915_priotree_fini(request->i915, &request->priotree);
 	i915_gem_request_put(request);
 }
 
@@ -467,6 +540,8 @@ i915_gem_request_alloc(struct intel_engine_cs *engine,
 	 */
 	i915_sw_fence_await_sw_fence(&req->execute, &req->submit, &req->execq);
 
+	i915_priotree_init(&req->priotree);
+
 	INIT_LIST_HEAD(&req->active_list);
 	req->i915 = dev_priv;
 	req->engine = engine;
@@ -520,6 +595,14 @@ i915_gem_request_await_request(struct drm_i915_gem_request *to,
 
 	GEM_BUG_ON(to == from);
 
+	if (to->engine->schedule) {
+		ret = i915_priotree_add_dependency(to->i915,
+						   &to->priotree,
+						   &from->priotree);
+		if (ret < 0)
+			return ret;
+	}
+
 	if (to->timeline == from->timeline)
 		return 0;
 
@@ -743,9 +826,15 @@ void __i915_add_request(struct drm_i915_gem_request *request, bool flush_caches)
 
 	prev = i915_gem_active_raw(&timeline->last_request,
 				   &request->i915->drm.struct_mutex);
-	if (prev)
+	if (prev) {
 		i915_sw_fence_await_sw_fence(&request->submit, &prev->submit,
 					     &request->submitq);
+		if (engine->schedule)
+			__i915_priotree_add_dependency(&request->priotree,
+						       &prev->priotree,
+						       &request->dep,
+						       0);
+	}
 
 	spin_lock_irq(&timeline->lock);
 	list_add_tail(&request->link, &timeline->requests);

drivers/gpu/drm/i915/i915_gem_request.h

@@ -44,6 +44,28 @@ struct intel_signal_node {
 	struct intel_wait wait;
 };
 
+struct i915_dependency {
+	struct i915_priotree *signaler;
+	struct list_head signal_link;
+	struct list_head wait_link;
+	unsigned long flags;
+#define I915_DEPENDENCY_ALLOC BIT(0)
+};
+
+/* Requests exist in a complex web of interdependencies. Each request
+ * has to wait for some other request to complete before it is ready to be run
+ * (e.g. we have to wait until the pixels have been rendering into a texture
+ * before we can copy from it). We track the readiness of a request in terms
+ * of fences, but we also need to keep the dependency tree for the lifetime
+ * of the request (beyond the life of an individual fence). We use the tree
+ * at various points to reorder the requests whilst keeping the requests
+ * in order with respect to their various dependencies.
+ */
+struct i915_priotree {
+	struct list_head signalers_list; /* those before us, we depend upon */
+	struct list_head waiters_list; /* those after us, they depend upon us */
+};
+
 /**
  * Request queue structure.
  *
@@ -105,6 +127,17 @@ struct drm_i915_gem_request {
 	wait_queue_t submitq;
 	wait_queue_t execq;
 
+	/* A list of everyone we wait upon, and everyone who waits upon us.
+	 * Even though we will not be submitted to the hardware before the
+	 * submit fence is signaled (it waits for all external events as well
+	 * as our own requests), the scheduler still needs to know the
+	 * dependency tree for the lifetime of the request (from execbuf
+	 * to retirement), i.e. bidirectional dependency information for the
+	 * request not tied to individual fences.
+	 */
+	struct i915_priotree priotree;
+	struct i915_dependency dep;
+
 	u32 global_seqno;
 
 	/** GEM sequence number associated with the previous request,