Skip to contentSkip to conversation

drm: add syncobj timeline support v4

RFC PATCH In review Chunming Zhou 6 participants 21 messages Sep 6, 2018 – Sep 14, 2018 15 messages not in the archive

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

RFC PATCH 1/3 drm: add syncobj timeline support v4

VK_KHR_timeline_semaphore:
This extension introduces a new type of semaphore that has an integer payload identifying a point in a timeline. Such timeline semaphores support the following operations:
* CPU query - A host operation that allows querying the payload of the timeline semaphore.
* CPU wait - A host operation that allows a blocking wait for a timeline semaphore to reach a specified value.
* Device wait - A device operation that allows waiting for a timeline semaphore to reach a specified value.
* Device signal - A device operation that allows advancing the timeline semaphore to a specified value.

Since it's a timeline, that means the front time point(PT) always is signaled before the late PT. a. signal PT design:
Signal PT fence N depends on PT[N-1] fence and signal opertion fence, when PT[N] fence is signaled, the timeline will increase to value of PT[N].
b. wait PT design:
Wait PT fence is signaled by reaching timeline point value, when timeline is increasing, will compare wait PTs value with new timeline value, if PT value is lower than timeline value, then wait PT will be signaled, otherwise keep in list. semaphore wait operation can wait on any point of timeline, so need a RB tree to order them. And wait PT could ahead of signal PT, we need a sumission fence to perform that.

v2:
1. remove unused DRM_SYNCOBJ_CREATE_TYPE_NORMAL. (Christian)
2. move unexposed denitions to .c file. (Daniel Vetter)
3. split up the change to drm_syncobj_find_fence() in a separate patch. (Christian)
4. split up the change to drm_syncobj_replace_fence() in a separate patch.
5. drop the submission_fence implementation and instead use wait_event() for that. (Christian)
6. WARN_ON(point != 0) for NORMAL type syncobj case. (Daniel Vetter)

v3:
1. replace normal syncobj with timeline implemenation. (Vetter and Christian)
a. normal syncobj signal op will create a signal PT to tail of signal pt list.
b. normal syncobj wait op will create a wait pt with last signal point, and this wait PT is only signaled by related signal point PT.
2. many bug fix and clean up
3. stub fence moving is moved to other patch.

v4:
1. fix RB tree loop with while(node=rb_first(...)). (Christian)
2. fix syncobj lifecycle. (Christian)
3. only enable_signaling when there is wait_pt. (Christian)

Tested by ./deqp-vk -n dEQP-VK*semaphore* for normal syncobj

Cc Christian Konig <[email protected]>
Cc Dave Airlie <[email protected]>
Cc Daniel Rakos <[email protected]>
Cc Daniel Vetter <[email protected]>
---
 drivers/gpu/drm/drm_syncobj.c              | 516 +++++++++++++++++----
 drivers/gpu/drm/i915/i915_gem_execbuffer.c |   4 +-
 include/drm/drm_syncobj.h                  |  78 ++--
 include/uapi/drm/drm.h                     |   1 +
 4 files changed, 448 insertions(+), 151 deletions(-)
drivers/gpu/drm/drm_syncobj.c +417 -99
index e9ce623d049e..94b31de23858 100644
@@ -56,6 +56,9 @@
 #include "drm_internal.h"
 #include <drm/drm_syncobj.h>
 
+/* merge normal syncobj to timeline syncobj, the point interval is 1 */
+#define DRM_SYNCOBJ_NORMAL_POINT 1
+
 struct drm_syncobj_stub_fence {
 	struct dma_fence base;
 	spinlock_t lock;
@@ -82,6 +85,18 @@ static const struct dma_fence_ops drm_syncobj_stub_fence_ops = {
 	.release = drm_syncobj_stub_fence_release,
 };
 
+struct drm_syncobj_signal_pt {
+	struct dma_fence_array *base;
+	u64    value;
+	struct list_head list;
+};
+struct drm_syncobj_wait_pt {
+	struct drm_syncobj_stub_fence base;
+	struct dma_fence *signal_pt_fence;
+	struct dma_fence_cb wait_cb;
+	struct rb_node   node;
+	u64    value;
+};
 
 /**
  * drm_syncobj_find - lookup and reference a sync object.
@@ -109,61 +124,238 @@ struct drm_syncobj *drm_syncobj_find(struct drm_file *file_private,
 }
 EXPORT_SYMBOL(drm_syncobj_find);
 
-static void drm_syncobj_add_callback_locked(struct drm_syncobj *syncobj,
-					    struct drm_syncobj_cb *cb,
-					    drm_syncobj_func_t func)
+static void drm_syncobj_timeline_init(struct drm_syncobj *syncobj,
+				      struct drm_syncobj_timeline *syncobj_timeline)
 {
-	cb->func = func;
-	list_add_tail(&cb->node, &syncobj->cb_list);
+	spin_lock(&syncobj->lock);
+	syncobj_timeline->timeline_context = dma_fence_context_alloc(1);
+	syncobj_timeline->timeline = 0;
+	syncobj_timeline->signal_point = 0;
+	init_waitqueue_head(&syncobj_timeline->wq);
+
+	syncobj_timeline->wait_pt_tree = RB_ROOT;
+	INIT_LIST_HEAD(&syncobj_timeline->signal_pt_list);
+	spin_unlock(&syncobj->lock);
 }
 
-static int drm_syncobj_fence_get_or_add_callback(struct drm_syncobj *syncobj,
-						 struct dma_fence **fence,
-						 struct drm_syncobj_cb *cb,
-						 drm_syncobj_func_t func)
+static void drm_syncobj_timeline_fini(struct drm_syncobj *syncobj,
+				      struct drm_syncobj_timeline *syncobj_timeline)
 {
-	int ret;
-
-	*fence = drm_syncobj_fence_get(syncobj);
-	if (*fence)
-		return 1;
+	struct rb_node *node = NULL;
+	struct drm_syncobj_wait_pt *wait_pt = NULL;
+	struct drm_syncobj_signal_pt *signal_pt = NULL, *tmp;
 
 	spin_lock(&syncobj->lock);
-	/* We've already tried once to get a fence and failed.  Now that we
-	 * have the lock, try one more time just to be sure we don't add a
-	 * callback when a fence has already been set.
-	 */
-	if (syncobj->fence) {
-		*fence = dma_fence_get(rcu_dereference_protected(syncobj->fence,
-								 lockdep_is_held(&syncobj->lock)));
-		ret = 1;
-	} else {
-		*fence = NULL;
-		drm_syncobj_add_callback_locked(syncobj, cb, func);
-		ret = 0;
+	if (!RB_EMPTY_ROOT(&syncobj_timeline->wait_pt_tree))
+		DRM_ERROR("There still are advanced wait_pts! pls check!");
+	while ((node = rb_first(&syncobj_timeline->wait_pt_tree))) {
+		wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node);
+		rb_erase(&wait_pt->node,
+			 &syncobj_timeline->wait_pt_tree);
+		RB_CLEAR_NODE(&wait_pt->node);
+		if (!wait_pt->signal_pt_fence) {
+			/* if wait_pt still is advanced, fake signal it.
+			 * generally shouldn't happens */
+			dma_fence_set_error(&wait_pt->base.base, -ENOENT);
+			dma_fence_signal(&wait_pt->base.base);
+		}
+		/* kfree(wait_pt) is excuted by fence put */
+		dma_fence_put(&wait_pt->base.base);
+	}
+	list_for_each_entry_safe(signal_pt, tmp,
+				 &syncobj_timeline->signal_pt_list, list) {
+		list_del(&signal_pt->list);
+		dma_fence_put(&signal_pt->base->base);
+		kfree(signal_pt);
 	}
 	spin_unlock(&syncobj->lock);
+}
 
-	return ret;
+static void wait_pt_func(struct dma_fence *fence, struct dma_fence_cb *cb)
+{
+	struct drm_syncobj_wait_pt *wait_pt =
+		container_of(cb, struct drm_syncobj_wait_pt, wait_cb);
+
+	dma_fence_signal(&wait_pt->base.base);
+	dma_fence_put(wait_pt->signal_pt_fence);
+	wait_pt->signal_pt_fence = NULL;
+	dma_fence_put(&wait_pt->base.base);
+}
+
+static void drm_syncobj_find_signal_pt_for_wait_pt(struct drm_syncobj *syncobj,
+						   struct drm_syncobj_wait_pt *wait_pt)
+{
+	struct drm_syncobj_timeline *timeline = &syncobj->syncobj_timeline;
+	struct drm_syncobj_signal_pt *signal_pt;
+	int ret;
+
+	if (wait_pt->signal_pt_fence) {
+		return;
+	} else if ((syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) &&
+		   (wait_pt->value <= timeline->timeline)) {
+		dma_fence_signal(&wait_pt->base.base);
+		rb_erase(&wait_pt->node,
+			 &timeline->wait_pt_tree);
+		RB_CLEAR_NODE(&wait_pt->node);
+		dma_fence_put(&wait_pt->base.base);
+		return;
+	}
+
+	list_for_each_entry(signal_pt, &timeline->signal_pt_list, list) {
+		if (wait_pt->value < signal_pt->value)
+			continue;
+		if ((syncobj->type == DRM_SYNCOBJ_TYPE_NORMAL) &&
+		    (wait_pt->value != signal_pt->value))
+			continue;
+		wait_pt->signal_pt_fence = dma_fence_get(&signal_pt->base->base);
+		ret = dma_fence_add_callback(wait_pt->signal_pt_fence,
+					     &wait_pt->wait_cb,
+					     wait_pt_func);
+		if (ret == -ENOENT) {
+			dma_fence_signal(&wait_pt->base.base);
+			dma_fence_put(wait_pt->signal_pt_fence);
+			wait_pt->signal_pt_fence = NULL;
+			rb_erase(&wait_pt->node,
+				 &timeline->wait_pt_tree);
+			RB_CLEAR_NODE(&wait_pt->node);
+			dma_fence_put(&wait_pt->base.base);
+		} else if (ret < 0) {
+			dma_fence_put(wait_pt->signal_pt_fence);
+			DRM_ERROR("add callback error!");
+		} else {
+			/* after adding callback, remove from rb tree */
+			rb_erase(&wait_pt->node,
+				 &timeline->wait_pt_tree);
+			RB_CLEAR_NODE(&wait_pt->node);
+		}
+		return;
+	}
+	/* signaled pt was released */
+	if (!wait_pt->signal_pt_fence && (wait_pt->value <=
+					  timeline->signal_point)) {
+		dma_fence_signal(&wait_pt->base.base);
+		rb_erase(&wait_pt->node,
+			 &timeline->wait_pt_tree);
+		RB_CLEAR_NODE(&wait_pt->node);
+		dma_fence_put(&wait_pt->base.base);
+	}
 }
 
-void drm_syncobj_add_callback(struct drm_syncobj *syncobj,
-			      struct drm_syncobj_cb *cb,
-			      drm_syncobj_func_t func)
+static int drm_syncobj_timeline_create_signal_pt(struct drm_syncobj *syncobj,
+						 struct dma_fence *fence,
+						 u64 point)
 {
+	struct drm_syncobj_signal_pt *signal_pt =
+		kzalloc(sizeof(struct drm_syncobj_signal_pt), GFP_KERNEL);
+	struct drm_syncobj_signal_pt *tail_pt;
+	struct dma_fence **fences;
+	struct rb_node *node;
+	struct drm_syncobj_wait_pt *tail_wait_pt = NULL;
+	int num_fences = 0;
+	int ret = 0, i;
+
+	if (!signal_pt)
+		return -ENOMEM;
+	if (syncobj->syncobj_timeline.signal_point >= point) {
+		DRM_WARN("A later signal is ready!");
+		goto out;
+	}
+	if (!fence)
+		goto out;
+
+	fences = kmalloc_array(sizeof(void *), 2, GFP_KERNEL);
+	if (!fences)
+		goto out;
+	fences[num_fences++] = dma_fence_get(fence);
+	/* timeline syncobj must take this dependency */
+	if (syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+		spin_lock(&syncobj->lock);
+		if (!list_empty(&syncobj->syncobj_timeline.signal_pt_list)) {
+			tail_pt = list_last_entry(&syncobj->syncobj_timeline.signal_pt_list,
+						  struct drm_syncobj_signal_pt, list);
+			fences[num_fences++] = dma_fence_get(&tail_pt->base->base);
+		}
+		spin_unlock(&syncobj->lock);
+	}
+	signal_pt->base = dma_fence_array_create(num_fences, fences,
+						 syncobj->syncobj_timeline.timeline_context,
+						 point, false);
+	if (!signal_pt->base)
+		goto fail;
+
+	signal_pt->value = point;
 	spin_lock(&syncobj->lock);
-	drm_syncobj_add_callback_locked(syncobj, cb, func);
+	syncobj->syncobj_timeline.signal_point = point;
+	INIT_LIST_HEAD(&signal_pt->list);
+	list_add_tail(&signal_pt->list, &syncobj->syncobj_timeline.signal_pt_list);
+
+	/* check if there is advanced wait */
+	node = rb_last(&syncobj->syncobj_timeline.wait_pt_tree);
+	if (node)
+		tail_wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node);
+	if (tail_wait_pt && !tail_wait_pt->signal_pt_fence) {
+		for (node = rb_first(&syncobj->syncobj_timeline.wait_pt_tree);
+		     node != NULL; node = rb_next(node)) {
+			struct drm_syncobj_wait_pt *wait_pt =
+				rb_entry(node, struct drm_syncobj_wait_pt,
+					 node);
+
+			drm_syncobj_find_signal_pt_for_wait_pt(syncobj,
+							       wait_pt);
+		}
+	}
+
 	spin_unlock(&syncobj->lock);
+	wake_up_all(&syncobj->syncobj_timeline.wq);
+
+	return 0;
+fail:
+	for (i = 0; i < num_fences; i++)
+		dma_fence_put(fences[i]);
+	kfree(fences);
+out:
+	kfree(signal_pt);
+	return ret;
 }
 
-void drm_syncobj_remove_callback(struct drm_syncobj *syncobj,
-				 struct drm_syncobj_cb *cb)
+static void drm_syncobj_timeline_garbage_collection(struct drm_syncobj *syncobj)
 {
+	struct drm_syncobj_timeline *timeline = &syncobj->syncobj_timeline;
+	struct rb_node *node;
+	struct drm_syncobj_wait_pt *wait_pt;
+	struct drm_syncobj_signal_pt *signal_pt, *tmp;
+
 	spin_lock(&syncobj->lock);
-	list_del_init(&cb->node);
+	node = rb_first(&timeline->wait_pt_tree);
+	while (node) {
+		wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node);
+		if (!dma_fence_is_signaled(&wait_pt->base.base)) {
+			node = rb_next(node);
+			continue;
+		}
+		rb_erase(&wait_pt->node,
+			 &timeline->wait_pt_tree);
+		RB_CLEAR_NODE(&wait_pt->node);
+		/* kfree(wait_pt) is excuted by fence put */
+		dma_fence_put(&wait_pt->base.base);
+		node = rb_first(&timeline->wait_pt_tree);
+	}
+	list_for_each_entry_safe(signal_pt, tmp,
+				 &timeline->signal_pt_list, list) {
+		if (dma_fence_is_signaled(&signal_pt->base->base)) {
+			timeline->timeline = signal_pt->value;
+			list_del(&signal_pt->list);
+			dma_fence_put(&signal_pt->base->base);
+			kfree(signal_pt);
+		} else {
+			/*signal_pt is in order in list, from small to big, so
+			 * the later must not be signal either */
+			break;
+		}
+	}
+
 	spin_unlock(&syncobj->lock);
 }
-
 /**
  * drm_syncobj_replace_fence - replace fence in a sync object.
  * @syncobj: Sync object to replace fence in
@@ -176,28 +368,29 @@ void drm_syncobj_replace_fence(struct drm_syncobj *syncobj,
 			       u64 point,
 			       struct dma_fence *fence)
 {
-	struct dma_fence *old_fence;
-	struct drm_syncobj_cb *cur, *tmp;
-
-	if (fence)
-		dma_fence_get(fence);
-
-	spin_lock(&syncobj->lock);
-
-	old_fence = rcu_dereference_protected(syncobj->fence,
-					      lockdep_is_held(&syncobj->lock));
-	rcu_assign_pointer(syncobj->fence, fence);
-
-	if (fence != old_fence) {
-		list_for_each_entry_safe(cur, tmp, &syncobj->cb_list, node) {
-			list_del_init(&cur->node);
-			cur->func(syncobj, cur);
+	drm_syncobj_timeline_garbage_collection(syncobj);
+	if (syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+		if (fence)
+			drm_syncobj_timeline_create_signal_pt(syncobj, fence,
+							      point);
+		return;
+	} else if (syncobj->type == DRM_SYNCOBJ_TYPE_NORMAL) {
+		u64 pt_value;
+
+		if (!fence) {
+			drm_syncobj_timeline_fini(syncobj,
+						  &syncobj->syncobj_timeline);
+			drm_syncobj_timeline_init(syncobj,
+						  &syncobj->syncobj_timeline);
+			return;
 		}
+		pt_value = syncobj->syncobj_timeline.signal_point +
+			DRM_SYNCOBJ_NORMAL_POINT;
+		drm_syncobj_timeline_create_signal_pt(syncobj, fence, pt_value);
+		return;
+	} else {
+		DRM_ERROR("the syncobj type isn't support\n");
 	}
-
-	spin_unlock(&syncobj->lock);
-
-	dma_fence_put(old_fence);
 }
 EXPORT_SYMBOL(drm_syncobj_replace_fence);
 
@@ -220,6 +413,120 @@ static int drm_syncobj_assign_null_handle(struct drm_syncobj *syncobj)
 	return 0;
 }
 
+static struct drm_syncobj_wait_pt *
+drm_syncobj_timeline_lookup_wait_pt_fence(struct drm_syncobj *syncobj,
+					  u64 point,
+					  struct dma_fence **fence)
+{
+	struct rb_node *node = syncobj->syncobj_timeline.wait_pt_tree.rb_node;
+	struct drm_syncobj_wait_pt *wait_pt = NULL;
+
+	spin_lock(&syncobj->lock);
+	while(node) {
+		int result;
+
+		wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node);
+		result = point - wait_pt->value;
+		if (result < 0) {
+			node = node->rb_left;
+		} else if (result > 0) {
+			node = node->rb_right;
+		} else {
+			if (fence)
+				*fence = dma_fence_get(&wait_pt->base.base);
+			break;
+		}
+	}
+	spin_unlock(&syncobj->lock);
+
+	return wait_pt;
+}
+
+static struct drm_syncobj_wait_pt *
+drm_syncobj_timeline_create_wait_pt_and_return_fence(struct drm_syncobj *syncobj,
+						     u64 point,
+						     struct dma_fence **fence)
+{
+	struct drm_syncobj_wait_pt *wait_pt = kzalloc(sizeof(*wait_pt),
+						      GFP_KERNEL);
+	struct drm_syncobj_stub_fence *base;
+	struct rb_node **new = &(syncobj->syncobj_timeline.wait_pt_tree.rb_node), *parent = NULL;
+
+	if (!wait_pt)
+		return NULL;
+	base = &wait_pt->base;
+	spin_lock_init(&base->lock);
+	dma_fence_init(&base->base,
+		       &drm_syncobj_stub_fence_ops,
+		       &base->lock,
+		       syncobj->syncobj_timeline.timeline_context, point);
+	wait_pt->value = point;
+	wait_pt->signal_pt_fence = NULL;
+
+	/* wait pt must be in an order, so that we can easily lookup and signal
+	 * it */
+	spin_lock(&syncobj->lock);
+	while(*new) {
+		struct drm_syncobj_wait_pt *this =
+			rb_entry(*new, struct drm_syncobj_wait_pt, node);
+		int result = wait_pt->value - this->value;
+
+		parent = *new;
+		if (result < 0)
+			new = &((*new)->rb_left);
+		else if (result > 0)
+			new = &((*new)->rb_right);
+		else
+			goto exist;
+	}
+	if (fence)
+		*fence = dma_fence_get(&wait_pt->base.base);
+	rb_link_node(&wait_pt->node, parent, new);
+	rb_insert_color(&wait_pt->node, &syncobj->syncobj_timeline.wait_pt_tree);
+	spin_unlock(&syncobj->lock);
+	return wait_pt;
+exist:
+	spin_unlock(&syncobj->lock);
+	dma_fence_put(&wait_pt->base.base);
+	wait_pt = drm_syncobj_timeline_lookup_wait_pt_fence(syncobj, point,
+							    fence);
+	return wait_pt;
+}
+
+static struct dma_fence *
+drm_syncobj_timeline_point_get(struct drm_syncobj *syncobj, u64 point, u64 flags)
+{
+	struct drm_syncobj_wait_pt *wait_pt;
+	struct dma_fence *fence = NULL;
+
+	/* check if the wait pt exists */
+	wait_pt = drm_syncobj_timeline_lookup_wait_pt_fence(syncobj, point,
+							    &fence);
+	if (!fence) {
+		/* This is a new wait pt, so create it */
+		wait_pt = drm_syncobj_timeline_create_wait_pt_and_return_fence(syncobj, point,
+							      &fence);
+		if (!fence)
+			return NULL;
+	}
+	if (wait_pt) {
+		int ret = 0;
+
+		if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) {
+			ret = wait_event_interruptible_timeout(syncobj->syncobj_timeline.wq,
+							       wait_pt->value <= syncobj->syncobj_timeline.signal_point,
+							       msecs_to_jiffies(10000)); /* wait 10s */
+			if (ret <= 0)
+				return NULL;
+		}
+		spin_lock(&syncobj->lock);
+		drm_syncobj_find_signal_pt_for_wait_pt(syncobj, wait_pt);
+		spin_unlock(&syncobj->lock);
+		return fence;
+	}
+	return NULL;
+}
+
 /**
  * drm_syncobj_find_fence - lookup and reference the fence in a sync object
  * @file_private: drm file private pointer
@@ -234,20 +541,46 @@ static int drm_syncobj_assign_null_handle(struct drm_syncobj *syncobj)
  * contains a reference to the fence, which must be released by calling
  * dma_fence_put().
  */
-int drm_syncobj_find_fence(struct drm_file *file_private,
-			   u32 handle, u64 point,
-			   struct dma_fence **fence)
+int drm_syncobj_search_fence(struct drm_syncobj *syncobj, u64 point,
+			     u64 flags, struct dma_fence **fence)
 {
-	struct drm_syncobj *syncobj = drm_syncobj_find(file_private, handle);
 	int ret = 0;
 
 	if (!syncobj)
 		return -ENOENT;
 
-	*fence = drm_syncobj_fence_get(syncobj);
+	drm_syncobj_timeline_garbage_collection(syncobj);
+	if (syncobj->type == DRM_SYNCOBJ_TYPE_NORMAL) {
+		/*NORMAL syncobj always wait on last pt */
+		u64 tail_pt_value = syncobj->syncobj_timeline.signal_point;
+
+		if (tail_pt_value == 0)
+			tail_pt_value += DRM_SYNCOBJ_NORMAL_POINT;
+		/* NORMAL syncobj doesn't care point value */
+		WARN_ON(point != 0);
+		*fence = drm_syncobj_timeline_point_get(syncobj, tail_pt_value,
+							flags);
+	} else if (syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+		*fence = drm_syncobj_timeline_point_get(syncobj, point,
+							flags);
+	} else {
+		DRM_ERROR("Don't support this type syncobj\n");
+		*fence = NULL;
+	}
 	if (!*fence) {
 		ret = -EINVAL;
 	}
+	return ret;
+}
+EXPORT_SYMBOL(drm_syncobj_search_fence);
+int drm_syncobj_find_fence(struct drm_file *file_private,
+			   u32 handle, u64 point,
+			   struct dma_fence **fence) {
+	struct drm_syncobj *syncobj = drm_syncobj_find(file_private, handle);
+
+	int ret = drm_syncobj_search_fence(syncobj, point,
+					DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT,
+					fence);
 	drm_syncobj_put(syncobj);
 	return ret;
 }
@@ -264,7 +597,7 @@ void drm_syncobj_free(struct kref *kref)
 	struct drm_syncobj *syncobj = container_of(kref,
 						   struct drm_syncobj,
 						   refcount);
-	drm_syncobj_replace_fence(syncobj, 0, NULL);
+	drm_syncobj_timeline_fini(syncobj, &syncobj->syncobj_timeline);
 	kfree(syncobj);
 }
 EXPORT_SYMBOL(drm_syncobj_free);
@@ -292,8 +625,12 @@ int drm_syncobj_create(struct drm_syncobj **out_syncobj, uint32_t flags,
 		return -ENOMEM;
 
 	kref_init(&syncobj->refcount);
-	INIT_LIST_HEAD(&syncobj->cb_list);
 	spin_lock_init(&syncobj->lock);
+	if (flags & DRM_SYNCOBJ_CREATE_TYPE_TIMELINE)
+		syncobj->type = DRM_SYNCOBJ_TYPE_TIMELINE;
+	else
+		syncobj->type = DRM_SYNCOBJ_TYPE_NORMAL;
+	drm_syncobj_timeline_init(syncobj, &syncobj->syncobj_timeline);
 
 	if (flags & DRM_SYNCOBJ_CREATE_SIGNALED) {
 		ret = drm_syncobj_assign_null_handle(syncobj);
@@ -651,7 +988,6 @@ struct syncobj_wait_entry {
 	struct task_struct *task;
 	struct dma_fence *fence;
 	struct dma_fence_cb fence_cb;
-	struct drm_syncobj_cb syncobj_cb;
 };
 
 static void syncobj_wait_fence_func(struct dma_fence *fence,
@@ -663,18 +999,6 @@ static void syncobj_wait_fence_func(struct dma_fence *fence,
 	wake_up_process(wait->task);
 }
 
-static void syncobj_wait_syncobj_func(struct drm_syncobj *syncobj,
-				      struct drm_syncobj_cb *cb)
-{
-	struct syncobj_wait_entry *wait =
-		container_of(cb, struct syncobj_wait_entry, syncobj_cb);
-
-	/* This happens inside the syncobj lock */
-	wait->fence = dma_fence_get(rcu_dereference_protected(syncobj->fence,
-							      lockdep_is_held(&syncobj->lock)));
-	wake_up_process(wait->task);
-}
-
 static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
 						  uint32_t count,
 						  uint32_t flags,
@@ -698,14 +1022,11 @@ static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
 	signaled_count = 0;
 	for (i = 0; i < count; ++i) {
 		entries[i].task = current;
-		entries[i].fence = drm_syncobj_fence_get(syncobjs[i]);
+		ret = drm_syncobj_search_fence(syncobjs[i], 0, 0,
+					       &entries[i].fence);
 		if (!entries[i].fence) {
-			if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) {
-				continue;
-			} else {
-				ret = -EINVAL;
-				goto cleanup_entries;
-			}
+			ret = -EINVAL;
+			goto cleanup_entries;
 		}
 
 		if (dma_fence_is_signaled(entries[i].fence)) {
@@ -733,15 +1054,6 @@ static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
 	 * fallthough and try a 0 timeout wait!
 	 */
 
-	if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) {
-		for (i = 0; i < count; ++i) {
-			drm_syncobj_fence_get_or_add_callback(syncobjs[i],
-							      &entries[i].fence,
-							      &entries[i].syncobj_cb,
-							      syncobj_wait_syncobj_func);
-		}
-	}
-
 	do {
 		set_current_state(TASK_INTERRUPTIBLE);
 
@@ -789,13 +1101,10 @@ static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
 
 cleanup_entries:
 	for (i = 0; i < count; ++i) {
-		if (entries[i].syncobj_cb.func)
-			drm_syncobj_remove_callback(syncobjs[i],
-						    &entries[i].syncobj_cb);
+		dma_fence_put(entries[i].fence);
 		if (entries[i].fence_cb.func)
 			dma_fence_remove_callback(entries[i].fence,
 						  &entries[i].fence_cb);
-		dma_fence_put(entries[i].fence);
 	}
 	kfree(entries);
 
@@ -970,12 +1279,21 @@ drm_syncobj_reset_ioctl(struct drm_device *dev, void *data,
 	if (ret < 0)
 		return ret;
 
-	for (i = 0; i < args->count_handles; i++)
-		drm_syncobj_replace_fence(syncobjs[i], 0, NULL);
-
+	for (i = 0; i < args->count_handles; i++) {
+		if (syncobjs[i]->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+			DRM_ERROR("timeline syncobj cannot reset!\n");
+			ret = -EINVAL;
+			goto out;
+		}
+		drm_syncobj_timeline_fini(syncobjs[i],
+					  &syncobjs[i]->syncobj_timeline);
+		drm_syncobj_timeline_init(syncobjs[i],
+					  &syncobjs[i]->syncobj_timeline);
+	}
+out:
 	drm_syncobj_array_free(syncobjs, args->count_handles);
 
-	return 0;
+	return ret;
 }
 
 int
drivers/gpu/drm/i915/i915_gem_execbuffer.c +3 -1
index 0a8d2d64f380..579e91a5858b 100644
@@ -2137,7 +2137,9 @@ await_fence_array(struct i915_execbuffer *eb,
 		if (!(flags & I915_EXEC_FENCE_WAIT))
 			continue;
 
-		fence = drm_syncobj_fence_get(syncobj);
+		drm_syncobj_search_fence(syncobj, 0,
+					 DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT,
+					 &fence);
 		if (!fence)
 			return -EINVAL;
 
include/drm/drm_syncobj.h +27 -51

111 lines not on this page · raw message

include/uapi/drm/drm.h +1 -0
index 300f336633f2..cebdb2541eb7 100644
@@ -717,6 +717,7 @@ struct drm_prime_handle {
 struct drm_syncobj_create {
 	__u32 handle;
 #define DRM_SYNCOBJ_CREATE_SIGNALED (1 << 0)
+#define DRM_SYNCOBJ_CREATE_TYPE_TIMELINE (1 << 1)
 	__u32 flags;
 };
 
Signature · 6 lines
2.17.1

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
#1 lore raw

PATCH 2/3 drm: add support of syncobj timeline point wait

points array is one-to-one match with syncobjs array.

---
 drivers/gpu/drm/drm_syncobj.c | 23 ++++++++++++++++++++---
 include/uapi/drm/drm.h        |  2 ++
 2 files changed, 22 insertions(+), 3 deletions(-)
drivers/gpu/drm/drm_syncobj.c +20 -3
index 94b31de23858..4e43c62ab432 100644
@@ -1000,6 +1000,7 @@ static void syncobj_wait_fence_func(struct dma_fence *fence,
 }
 
 static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
+						  void __user *user_points,
 						  uint32_t count,
 						  uint32_t flags,
 						  signed long timeout,
@@ -1007,13 +1008,25 @@ static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
 {
 	struct syncobj_wait_entry *entries;
 	struct dma_fence *fence;
+	uint64_t *points;
 	signed long ret;
 	uint32_t signaled_count, i;
 
-	entries = kcalloc(count, sizeof(*entries), GFP_KERNEL);
-	if (!entries)
+	points = kmalloc_array(count, sizeof(*points), GFP_KERNEL);
+	if (points == NULL)
 		return -ENOMEM;
 
+	if (copy_from_user(points, user_points, sizeof(uint64_t) * count)) {
+		ret = -EFAULT;
+		goto err_free_points;
+	}
+
+
+	entries = kcalloc(count, sizeof(*entries), GFP_KERNEL);
+	if (!entries) {
+		ret = -ENOMEM;
+		goto err_free_points;
+	}
 	/* Walk the list of sync objects and initialize entries.  We do
 	 * this up-front so that we can properly return -EINVAL if there is
 	 * a syncobj with a missing fence and then never have the chance of
@@ -1022,7 +1035,7 @@ static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
 	signaled_count = 0;
 	for (i = 0; i < count; ++i) {
 		entries[i].task = current;
-		ret = drm_syncobj_search_fence(syncobjs[i], 0, 0,
+		ret = drm_syncobj_search_fence(syncobjs[i], points[i], 0,
 					       &entries[i].fence);
 		if (!entries[i].fence) {
 			ret = -EINVAL;
@@ -1108,6 +1121,9 @@ static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
 	}
 	kfree(entries);
 
+err_free_points:
+	kfree(points);
+
 	return ret;
 }
 
@@ -1153,6 +1169,7 @@ static int drm_syncobj_array_wait(struct drm_device *dev,
 	uint32_t first = ~0;
 
 	ret = drm_syncobj_array_wait_timeout(syncobjs,
+					     u64_to_user_ptr(wait->points),
 					     wait->count_handles,
 					     wait->flags,
 					     timeout, &first);
include/uapi/drm/drm.h +2 -0
index cebdb2541eb7..4420c7604690 100644
@@ -740,6 +740,8 @@ struct drm_syncobj_handle {
 #define DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT (1 << 1)
 struct drm_syncobj_wait {
 	__u64 handles;
+	/* wait on specific timeline point for every handles*/
+	__u64 points;
 	/* absolute timeout */
 	__s64 timeout_nsec;
 	__u32 count_handles;
Signature · 6 lines
2.17.1

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

A message that is not in the archive (sent off-list or not yet received): [email protected]

#3 lore raw

Re: [PATCH 2/3] drm: add support of syncobj timeline point wait

Am 06.09.2018 um 08:25 schrieb Chunming Zhou:

Quoted 80 lines

points array is one-to-one match with syncobjs array.

---
  drivers/gpu/drm/drm_syncobj.c | 23 ++++++++++++++++++++---
  include/uapi/drm/drm.h        |  2 ++
  2 files changed, 22 insertions(+), 3 deletions(-)
drivers/gpu/drm/drm_syncobj.c +20 -3
index 94b31de23858..4e43c62ab432 100644
@@ -1000,6 +1000,7 @@ static void syncobj_wait_fence_func(struct dma_fence *fence,
  }
 
  static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
+						  void __user *user_points,
  						  uint32_t count,
  						  uint32_t flags,
  						  signed long timeout,
@@ -1007,13 +1008,25 @@ static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
  {
  	struct syncobj_wait_entry *entries;
  	struct dma_fence *fence;
+	uint64_t *points;
  	signed long ret;
  	uint32_t signaled_count, i;
 
-	entries = kcalloc(count, sizeof(*entries), GFP_KERNEL);
-	if (!entries)
+	points = kmalloc_array(count, sizeof(*points), GFP_KERNEL);
+	if (points == NULL)
  		return -ENOMEM;
 
+	if (copy_from_user(points, user_points, sizeof(uint64_t) * count)) {
+		ret = -EFAULT;
+		goto err_free_points;
+	}
+
+
+	entries = kcalloc(count, sizeof(*entries), GFP_KERNEL);
+	if (!entries) {
+		ret = -ENOMEM;
+		goto err_free_points;
+	}
  	/* Walk the list of sync objects and initialize entries.  We do
  	 * this up-front so that we can properly return -EINVAL if there is
  	 * a syncobj with a missing fence and then never have the chance of
@@ -1022,7 +1035,7 @@ static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
  	signaled_count = 0;
  	for (i = 0; i < count; ++i) {
  		entries[i].task = current;
-		ret = drm_syncobj_search_fence(syncobjs[i], 0, 0,
+		ret = drm_syncobj_search_fence(syncobjs[i], points[i], 0,
  					       &entries[i].fence);
  		if (!entries[i].fence) {
  			ret = -EINVAL;
@@ -1108,6 +1121,9 @@ static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
  	}
  	kfree(entries);
 
+err_free_points:
+	kfree(points);
+
  	return ret;
  }
 
@@ -1153,6 +1169,7 @@ static int drm_syncobj_array_wait(struct drm_device *dev,
  	uint32_t first = ~0;
 
  	ret = drm_syncobj_array_wait_timeout(syncobjs,
+					     u64_to_user_ptr(wait->points),
  					     wait->count_handles,
  					     wait->flags,
  					     timeout, &first);
include/uapi/drm/drm.h +0 -0
index cebdb2541eb7..4420c7604690 100644

include/uapi/drm/drm.h struct drm_syncobj_handle { · line 744

Quoted 5 lines
include/uapi/drm/drm.h +2 -0
@@ -740,6 +740,8 @@ struct drm_syncobj_handle {
  #define DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT (1 << 1)
  struct drm_syncobj_wait {
  	__u64 handles;
+	/* wait on specific timeline point for every handles*/
+	__u64 points;

That would break UAPI and isn't allowed like this.

You need to add a new operation to wait for multiple sync objects and their sequence number or add a flag or something like this here.

Regards,
Christian.

include/uapi/drm/drm.h struct drm_syncobj_handle { · line 747

Quoted 3 lines
include/uapi/drm/drm.h +0 -0
  	/* absolute timeout */
  	__s64 timeout_nsec;
  	__u32 count_handles;

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): [email protected]

#5 lore raw

PATCH 3/3 drm: add timeline syncobj payload query ioctl

user mode can query timeline payload.

---
 drivers/gpu/drm/drm_internal.h |  2 ++
 drivers/gpu/drm/drm_ioctl.c    |  2 ++
 drivers/gpu/drm/drm_syncobj.c  | 53 ++++++++++++++++++++++++++++++++++
 include/uapi/drm/drm.h         |  3 ++
 4 files changed, 60 insertions(+)
drivers/gpu/drm/drm_internal.h +2 -0
index 40179c5fc6b8..cd57a7ba8bf6 100644
@@ -182,6 +182,8 @@ int drm_syncobj_reset_ioctl(struct drm_device *dev, void *data,
 			    struct drm_file *file_private);
 int drm_syncobj_signal_ioctl(struct drm_device *dev, void *data,
 			     struct drm_file *file_private);
+int drm_syncobj_query_ioctl(struct drm_device *dev, void *data,
+			    struct drm_file *file_private);
 
 /* drm_framebuffer.c */
 void drm_framebuffer_print_info(struct drm_printer *p, unsigned int indent,
drivers/gpu/drm/drm_ioctl.c +2 -0
index ea10e9a26aad..af98663ea252 100644
@@ -673,6 +673,8 @@ static const struct drm_ioctl_desc drm_ioctls[] = {
 		      DRM_UNLOCKED|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF(DRM_IOCTL_SYNCOBJ_SIGNAL, drm_syncobj_signal_ioctl,
 		      DRM_UNLOCKED|DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF(DRM_IOCTL_SYNCOBJ_QUERY, drm_syncobj_query_ioctl,
+		      DRM_UNLOCKED|DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF(DRM_IOCTL_CRTC_GET_SEQUENCE, drm_crtc_get_sequence_ioctl, DRM_UNLOCKED),
 	DRM_IOCTL_DEF(DRM_IOCTL_CRTC_QUEUE_SEQUENCE, drm_crtc_queue_sequence_ioctl, DRM_UNLOCKED),
 	DRM_IOCTL_DEF(DRM_IOCTL_MODE_CREATE_LEASE, drm_mode_create_lease_ioctl, DRM_MASTER|DRM_UNLOCKED),
drivers/gpu/drm/drm_syncobj.c +53 -0

56 lines not on this page · raw message

include/uapi/drm/drm.h +3 -0
index 4420c7604690..6b64ce1fdfe5 100644
@@ -752,6 +752,8 @@ struct drm_syncobj_wait {
 
 struct drm_syncobj_array {
 	__u64 handles;
+	/* points are timeline syncobjs payloads returned by query ioctl */
+	__u64 points;
 	__u32 count_handles;
 	__u32 pad;
 };
@@ -912,6 +914,7 @@ extern "C" {
 #define DRM_IOCTL_MODE_GET_LEASE	DRM_IOWR(0xC8, struct drm_mode_get_lease)
 #define DRM_IOCTL_MODE_REVOKE_LEASE	DRM_IOWR(0xC9, struct drm_mode_revoke_lease)
 
+#define DRM_IOCTL_SYNCOBJ_QUERY		DRM_IOWR(0xCA, struct drm_syncobj_array)
 /**
  * Device specific ioctls should only be in their respective headers
  * The device specific ioctl range is from 0x40 to 0x9f.
Signature · 6 lines
2.17.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
#6 lore raw
In reply to a message not in the archive

Am 06.09.2018 um 08:25 schrieb Chunming Zhou:

Quoted 3 lines

VK_KHR_timeline_semaphore:
This extension introduces a new type of semaphore that has an integer payload identifying a point in a timeline. Such timeline semaphores support the

Drop the term semaphore here, better use syncobj.

Quoted 89 lines

following operations:
* CPU query - A host operation that allows querying the payload of the timeline semaphore.
* CPU wait - A host operation that allows a blocking wait for a timeline semaphore to reach a specified value.
* Device wait - A device operation that allows waiting for a timeline semaphore to reach a specified value.
* Device signal - A device operation that allows advancing the timeline semaphore to a specified value.

Since it's a timeline, that means the front time point(PT) always is signaled before the late PT. a. signal PT design:
Signal PT fence N depends on PT[N-1] fence and signal opertion fence, when PT[N] fence is signaled, the timeline will increase to value of PT[N].
b. wait PT design:
Wait PT fence is signaled by reaching timeline point value, when timeline is increasing, will compare wait PTs value with new timeline value, if PT value is lower than timeline value, then wait PT will be signaled, otherwise keep in list. semaphore wait operation can wait on any point of timeline, so need a RB tree to order them. And wait PT could ahead of signal PT, we need a sumission fence to perform that.

v2:
1. remove unused DRM_SYNCOBJ_CREATE_TYPE_NORMAL. (Christian)
2. move unexposed denitions to .c file. (Daniel Vetter)
3. split up the change to drm_syncobj_find_fence() in a separate patch. (Christian)
4. split up the change to drm_syncobj_replace_fence() in a separate patch.
5. drop the submission_fence implementation and instead use wait_event() for that. (Christian)
6. WARN_ON(point != 0) for NORMAL type syncobj case. (Daniel Vetter)

v3:
1. replace normal syncobj with timeline implemenation. (Vetter and Christian)
a. normal syncobj signal op will create a signal PT to tail of signal pt list.
b. normal syncobj wait op will create a wait pt with last signal point, and this wait PT is only signaled by related signal point PT.
2. many bug fix and clean up
3. stub fence moving is moved to other patch.

v4:
1. fix RB tree loop with while(node=rb_first(...)). (Christian)
2. fix syncobj lifecycle. (Christian)
3. only enable_signaling when there is wait_pt. (Christian)

Tested by ./deqp-vk -n dEQP-VK*semaphore* for normal syncobj

Cc Christian Konig <[email protected]>
Cc Dave Airlie <[email protected]>
Cc Daniel Rakos <[email protected]>
Cc Daniel Vetter <[email protected]>
---
  drivers/gpu/drm/drm_syncobj.c              | 516 +++++++++++++++++----
  drivers/gpu/drm/i915/i915_gem_execbuffer.c |   4 +-
  include/drm/drm_syncobj.h                  |  78 ++--
  include/uapi/drm/drm.h                     |   1 +
  4 files changed, 448 insertions(+), 151 deletions(-)
drivers/gpu/drm/drm_syncobj.c +15 -0
index e9ce623d049e..94b31de23858 100644
@@ -56,6 +56,9 @@
  #include "drm_internal.h"
  #include <drm/drm_syncobj.h>
 
+/* merge normal syncobj to timeline syncobj, the point interval is 1 */
+#define DRM_SYNCOBJ_NORMAL_POINT 1
+
  struct drm_syncobj_stub_fence {
  	struct dma_fence base;
  	spinlock_t lock;
@@ -82,6 +85,18 @@ static const struct dma_fence_ops drm_syncobj_stub_fence_ops = {
  	.release = drm_syncobj_stub_fence_release,
  };
 
+struct drm_syncobj_signal_pt {
+	struct dma_fence_array *base;
+	u64    value;
+	struct list_head list;
+};
+struct drm_syncobj_wait_pt {
+	struct drm_syncobj_stub_fence base;
+	struct dma_fence *signal_pt_fence;
+	struct dma_fence_cb wait_cb;
+	struct rb_node   node;
+	u64    value;
+};
 
  /**
   * drm_syncobj_find - lookup and reference a sync object.

drivers/gpu/drm/drm_syncobj.c struct drm_syncobj *drm_syncobj_find(struct drm_file *file_private, · line 128

Quoted 8 lines
drivers/gpu/drm/drm_syncobj.c +2 -3
@@ -109,61 +124,238 @@ struct drm_syncobj *drm_syncobj_find(struct drm_file *file_private,
  }
  EXPORT_SYMBOL(drm_syncobj_find);
 
-static void drm_syncobj_add_callback_locked(struct drm_syncobj *syncobj,
-					    struct drm_syncobj_cb *cb,
-					    drm_syncobj_func_t func)
+static void drm_syncobj_timeline_init(struct drm_syncobj *syncobj,
+				      struct drm_syncobj_timeline *syncobj_timeline)

Since we merged timeline and singleton syncobj you can drop the extra _timeline_ part in the function name I think.

Quoted 73 lines
drivers/gpu/drm/drm_syncobj.c +40 -24
  {
-	cb->func = func;
-	list_add_tail(&cb->node, &syncobj->cb_list);
+	spin_lock(&syncobj->lock);
+	syncobj_timeline->timeline_context = dma_fence_context_alloc(1);
+	syncobj_timeline->timeline = 0;
+	syncobj_timeline->signal_point = 0;
+	init_waitqueue_head(&syncobj_timeline->wq);
+
+	syncobj_timeline->wait_pt_tree = RB_ROOT;
+	INIT_LIST_HEAD(&syncobj_timeline->signal_pt_list);
+	spin_unlock(&syncobj->lock);
  }
 
-static int drm_syncobj_fence_get_or_add_callback(struct drm_syncobj *syncobj,
-						 struct dma_fence **fence,
-						 struct drm_syncobj_cb *cb,
-						 drm_syncobj_func_t func)
+static void drm_syncobj_timeline_fini(struct drm_syncobj *syncobj,
+				      struct drm_syncobj_timeline *syncobj_timeline)
  {
-	int ret;
-
-	*fence = drm_syncobj_fence_get(syncobj);
-	if (*fence)
-		return 1;
+	struct rb_node *node = NULL;
+	struct drm_syncobj_wait_pt *wait_pt = NULL;
+	struct drm_syncobj_signal_pt *signal_pt = NULL, *tmp;
 
  	spin_lock(&syncobj->lock);
-	/* We've already tried once to get a fence and failed.  Now that we
-	 * have the lock, try one more time just to be sure we don't add a
-	 * callback when a fence has already been set.
-	 */
-	if (syncobj->fence) {
-		*fence = dma_fence_get(rcu_dereference_protected(syncobj->fence,
-								 lockdep_is_held(&syncobj->lock)));
-		ret = 1;
-	} else {
-		*fence = NULL;
-		drm_syncobj_add_callback_locked(syncobj, cb, func);
-		ret = 0;
+	if (!RB_EMPTY_ROOT(&syncobj_timeline->wait_pt_tree))
+		DRM_ERROR("There still are advanced wait_pts! pls check!");
+	while ((node = rb_first(&syncobj_timeline->wait_pt_tree))) {
+		wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node);
+		rb_erase(&wait_pt->node,
+			 &syncobj_timeline->wait_pt_tree);
+		RB_CLEAR_NODE(&wait_pt->node);
+		if (!wait_pt->signal_pt_fence) {
+			/* if wait_pt still is advanced, fake signal it.
+			 * generally shouldn't happens */
+			dma_fence_set_error(&wait_pt->base.base, -ENOENT);
+			dma_fence_signal(&wait_pt->base.base);
+		}
+		/* kfree(wait_pt) is excuted by fence put */
+		dma_fence_put(&wait_pt->base.base);
+	}
+	list_for_each_entry_safe(signal_pt, tmp,
+				 &syncobj_timeline->signal_pt_list, list) {
+		list_del(&signal_pt->list);
+		dma_fence_put(&signal_pt->base->base);
+		kfree(signal_pt);
  	}
  	spin_unlock(&syncobj->lock);
+}
 
-	return ret;
+static void wait_pt_func(struct dma_fence *fence, struct dma_fence_cb *cb)
+{
+	struct drm_syncobj_wait_pt *wait_pt =
+		container_of(cb, struct drm_syncobj_wait_pt, wait_cb);

drivers/gpu/drm/drm_syncobj.c struct drm_syncobj *drm_syncobj_find(struct drm_file *file_private, · line 187

Quoted 10 lines
drivers/gpu/drm/drm_syncobj.c +10 -0
+
+	dma_fence_signal(&wait_pt->base.base);
+	dma_fence_put(wait_pt->signal_pt_fence);
+	wait_pt->signal_pt_fence = NULL;
+	dma_fence_put(&wait_pt->base.base);
+}
+
+static void drm_syncobj_find_signal_pt_for_wait_pt(struct drm_syncobj *syncobj,
+						   struct drm_syncobj_wait_pt *wait_pt)
+{

That whole approach still looks horrible complicated to me.

Especially the separation of signal and wait pt is completely unnecessary as far as I can see.

When a wait pt is requested we just need to search for the signal point which it will trigger.

Regards,
Christian.

Quoted 671 lines
drivers/gpu/drm/drm_syncobj.c +350 -72
+	struct drm_syncobj_timeline *timeline = &syncobj->syncobj_timeline;
+	struct drm_syncobj_signal_pt *signal_pt;
+	int ret;
+
+	if (wait_pt->signal_pt_fence) {
+		return;
+	} else if ((syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) &&
+		   (wait_pt->value <= timeline->timeline)) {
+		dma_fence_signal(&wait_pt->base.base);
+		rb_erase(&wait_pt->node,
+			 &timeline->wait_pt_tree);
+		RB_CLEAR_NODE(&wait_pt->node);
+		dma_fence_put(&wait_pt->base.base);
+		return;
+	}
+
+	list_for_each_entry(signal_pt, &timeline->signal_pt_list, list) {
+		if (wait_pt->value < signal_pt->value)
+			continue;
+		if ((syncobj->type == DRM_SYNCOBJ_TYPE_NORMAL) &&
+		    (wait_pt->value != signal_pt->value))
+			continue;
+		wait_pt->signal_pt_fence = dma_fence_get(&signal_pt->base->base);
+		ret = dma_fence_add_callback(wait_pt->signal_pt_fence,
+					     &wait_pt->wait_cb,
+					     wait_pt_func);
+		if (ret == -ENOENT) {
+			dma_fence_signal(&wait_pt->base.base);
+			dma_fence_put(wait_pt->signal_pt_fence);
+			wait_pt->signal_pt_fence = NULL;
+			rb_erase(&wait_pt->node,
+				 &timeline->wait_pt_tree);
+			RB_CLEAR_NODE(&wait_pt->node);
+			dma_fence_put(&wait_pt->base.base);
+		} else if (ret < 0) {
+			dma_fence_put(wait_pt->signal_pt_fence);
+			DRM_ERROR("add callback error!");
+		} else {
+			/* after adding callback, remove from rb tree */
+			rb_erase(&wait_pt->node,
+				 &timeline->wait_pt_tree);
+			RB_CLEAR_NODE(&wait_pt->node);
+		}
+		return;
+	}
+	/* signaled pt was released */
+	if (!wait_pt->signal_pt_fence && (wait_pt->value <=
+					  timeline->signal_point)) {
+		dma_fence_signal(&wait_pt->base.base);
+		rb_erase(&wait_pt->node,
+			 &timeline->wait_pt_tree);
+		RB_CLEAR_NODE(&wait_pt->node);
+		dma_fence_put(&wait_pt->base.base);
+	}
  }
 
-void drm_syncobj_add_callback(struct drm_syncobj *syncobj,
-			      struct drm_syncobj_cb *cb,
-			      drm_syncobj_func_t func)
+static int drm_syncobj_timeline_create_signal_pt(struct drm_syncobj *syncobj,
+						 struct dma_fence *fence,
+						 u64 point)
  {
+	struct drm_syncobj_signal_pt *signal_pt =
+		kzalloc(sizeof(struct drm_syncobj_signal_pt), GFP_KERNEL);
+	struct drm_syncobj_signal_pt *tail_pt;
+	struct dma_fence **fences;
+	struct rb_node *node;
+	struct drm_syncobj_wait_pt *tail_wait_pt = NULL;
+	int num_fences = 0;
+	int ret = 0, i;
+
+	if (!signal_pt)
+		return -ENOMEM;
+	if (syncobj->syncobj_timeline.signal_point >= point) {
+		DRM_WARN("A later signal is ready!");
+		goto out;
+	}
+	if (!fence)
+		goto out;
+
+	fences = kmalloc_array(sizeof(void *), 2, GFP_KERNEL);
+	if (!fences)
+		goto out;
+	fences[num_fences++] = dma_fence_get(fence);
+	/* timeline syncobj must take this dependency */
+	if (syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+		spin_lock(&syncobj->lock);
+		if (!list_empty(&syncobj->syncobj_timeline.signal_pt_list)) {
+			tail_pt = list_last_entry(&syncobj->syncobj_timeline.signal_pt_list,
+						  struct drm_syncobj_signal_pt, list);
+			fences[num_fences++] = dma_fence_get(&tail_pt->base->base);
+		}
+		spin_unlock(&syncobj->lock);
+	}
+	signal_pt->base = dma_fence_array_create(num_fences, fences,
+						 syncobj->syncobj_timeline.timeline_context,
+						 point, false);
+	if (!signal_pt->base)
+		goto fail;
+
+	signal_pt->value = point;
  	spin_lock(&syncobj->lock);
-	drm_syncobj_add_callback_locked(syncobj, cb, func);
+	syncobj->syncobj_timeline.signal_point = point;
+	INIT_LIST_HEAD(&signal_pt->list);
+	list_add_tail(&signal_pt->list, &syncobj->syncobj_timeline.signal_pt_list);
+
+	/* check if there is advanced wait */
+	node = rb_last(&syncobj->syncobj_timeline.wait_pt_tree);
+	if (node)
+		tail_wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node);
+	if (tail_wait_pt && !tail_wait_pt->signal_pt_fence) {
+		for (node = rb_first(&syncobj->syncobj_timeline.wait_pt_tree);
+		     node != NULL; node = rb_next(node)) {
+			struct drm_syncobj_wait_pt *wait_pt =
+				rb_entry(node, struct drm_syncobj_wait_pt,
+					 node);
+
+			drm_syncobj_find_signal_pt_for_wait_pt(syncobj,
+							       wait_pt);
+		}
+	}
+
  	spin_unlock(&syncobj->lock);
+	wake_up_all(&syncobj->syncobj_timeline.wq);
+
+	return 0;
+fail:
+	for (i = 0; i < num_fences; i++)
+		dma_fence_put(fences[i]);
+	kfree(fences);
+out:
+	kfree(signal_pt);
+	return ret;
  }
 
-void drm_syncobj_remove_callback(struct drm_syncobj *syncobj,
-				 struct drm_syncobj_cb *cb)
+static void drm_syncobj_timeline_garbage_collection(struct drm_syncobj *syncobj)
  {
+	struct drm_syncobj_timeline *timeline = &syncobj->syncobj_timeline;
+	struct rb_node *node;
+	struct drm_syncobj_wait_pt *wait_pt;
+	struct drm_syncobj_signal_pt *signal_pt, *tmp;
+
  	spin_lock(&syncobj->lock);
-	list_del_init(&cb->node);
+	node = rb_first(&timeline->wait_pt_tree);
+	while (node) {
+		wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node);
+		if (!dma_fence_is_signaled(&wait_pt->base.base)) {
+			node = rb_next(node);
+			continue;
+		}
+		rb_erase(&wait_pt->node,
+			 &timeline->wait_pt_tree);
+		RB_CLEAR_NODE(&wait_pt->node);
+		/* kfree(wait_pt) is excuted by fence put */
+		dma_fence_put(&wait_pt->base.base);
+		node = rb_first(&timeline->wait_pt_tree);
+	}
+	list_for_each_entry_safe(signal_pt, tmp,
+				 &timeline->signal_pt_list, list) {
+		if (dma_fence_is_signaled(&signal_pt->base->base)) {
+			timeline->timeline = signal_pt->value;
+			list_del(&signal_pt->list);
+			dma_fence_put(&signal_pt->base->base);
+			kfree(signal_pt);
+		} else {
+			/*signal_pt is in order in list, from small to big, so
+			 * the later must not be signal either */
+			break;
+		}
+	}
+
  	spin_unlock(&syncobj->lock);
  }
-
  /**
   * drm_syncobj_replace_fence - replace fence in a sync object.
   * @syncobj: Sync object to replace fence in
@@ -176,28 +368,29 @@ void drm_syncobj_replace_fence(struct drm_syncobj *syncobj,
  			       u64 point,
  			       struct dma_fence *fence)
  {
-	struct dma_fence *old_fence;
-	struct drm_syncobj_cb *cur, *tmp;
-
-	if (fence)
-		dma_fence_get(fence);
-
-	spin_lock(&syncobj->lock);
-
-	old_fence = rcu_dereference_protected(syncobj->fence,
-					      lockdep_is_held(&syncobj->lock));
-	rcu_assign_pointer(syncobj->fence, fence);
-
-	if (fence != old_fence) {
-		list_for_each_entry_safe(cur, tmp, &syncobj->cb_list, node) {
-			list_del_init(&cur->node);
-			cur->func(syncobj, cur);
+	drm_syncobj_timeline_garbage_collection(syncobj);
+	if (syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+		if (fence)
+			drm_syncobj_timeline_create_signal_pt(syncobj, fence,
+							      point);
+		return;
+	} else if (syncobj->type == DRM_SYNCOBJ_TYPE_NORMAL) {
+		u64 pt_value;
+
+		if (!fence) {
+			drm_syncobj_timeline_fini(syncobj,
+						  &syncobj->syncobj_timeline);
+			drm_syncobj_timeline_init(syncobj,
+						  &syncobj->syncobj_timeline);
+			return;
  		}
+		pt_value = syncobj->syncobj_timeline.signal_point +
+			DRM_SYNCOBJ_NORMAL_POINT;
+		drm_syncobj_timeline_create_signal_pt(syncobj, fence, pt_value);
+		return;
+	} else {
+		DRM_ERROR("the syncobj type isn't support\n");
  	}
-
-	spin_unlock(&syncobj->lock);
-
-	dma_fence_put(old_fence);
  }
  EXPORT_SYMBOL(drm_syncobj_replace_fence);
 
@@ -220,6 +413,120 @@ static int drm_syncobj_assign_null_handle(struct drm_syncobj *syncobj)
  	return 0;
  }
 
+static struct drm_syncobj_wait_pt *
+drm_syncobj_timeline_lookup_wait_pt_fence(struct drm_syncobj *syncobj,
+					  u64 point,
+					  struct dma_fence **fence)
+{
+	struct rb_node *node = syncobj->syncobj_timeline.wait_pt_tree.rb_node;
+	struct drm_syncobj_wait_pt *wait_pt = NULL;
+
+	spin_lock(&syncobj->lock);
+	while(node) {
+		int result;
+
+		wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node);
+		result = point - wait_pt->value;
+		if (result < 0) {
+			node = node->rb_left;
+		} else if (result > 0) {
+			node = node->rb_right;
+		} else {
+			if (fence)
+				*fence = dma_fence_get(&wait_pt->base.base);
+			break;
+		}
+	}
+	spin_unlock(&syncobj->lock);
+
+	return wait_pt;
+}
+
+static struct drm_syncobj_wait_pt *
+drm_syncobj_timeline_create_wait_pt_and_return_fence(struct drm_syncobj *syncobj,
+						     u64 point,
+						     struct dma_fence **fence)
+{
+	struct drm_syncobj_wait_pt *wait_pt = kzalloc(sizeof(*wait_pt),
+						      GFP_KERNEL);
+	struct drm_syncobj_stub_fence *base;
+	struct rb_node **new = &(syncobj->syncobj_timeline.wait_pt_tree.rb_node), *parent = NULL;
+
+	if (!wait_pt)
+		return NULL;
+	base = &wait_pt->base;
+	spin_lock_init(&base->lock);
+	dma_fence_init(&base->base,
+		       &drm_syncobj_stub_fence_ops,
+		       &base->lock,
+		       syncobj->syncobj_timeline.timeline_context, point);
+	wait_pt->value = point;
+	wait_pt->signal_pt_fence = NULL;
+
+	/* wait pt must be in an order, so that we can easily lookup and signal
+	 * it */
+	spin_lock(&syncobj->lock);
+	while(*new) {
+		struct drm_syncobj_wait_pt *this =
+			rb_entry(*new, struct drm_syncobj_wait_pt, node);
+		int result = wait_pt->value - this->value;
+
+		parent = *new;
+		if (result < 0)
+			new = &((*new)->rb_left);
+		else if (result > 0)
+			new = &((*new)->rb_right);
+		else
+			goto exist;
+	}
+	if (fence)
+		*fence = dma_fence_get(&wait_pt->base.base);
+	rb_link_node(&wait_pt->node, parent, new);
+	rb_insert_color(&wait_pt->node, &syncobj->syncobj_timeline.wait_pt_tree);
+	spin_unlock(&syncobj->lock);
+	return wait_pt;
+exist:
+	spin_unlock(&syncobj->lock);
+	dma_fence_put(&wait_pt->base.base);
+	wait_pt = drm_syncobj_timeline_lookup_wait_pt_fence(syncobj, point,
+							    fence);
+	return wait_pt;
+}
+
+static struct dma_fence *
+drm_syncobj_timeline_point_get(struct drm_syncobj *syncobj, u64 point, u64 flags)
+{
+	struct drm_syncobj_wait_pt *wait_pt;
+	struct dma_fence *fence = NULL;
+
+	/* check if the wait pt exists */
+	wait_pt = drm_syncobj_timeline_lookup_wait_pt_fence(syncobj, point,
+							    &fence);
+	if (!fence) {
+		/* This is a new wait pt, so create it */
+		wait_pt = drm_syncobj_timeline_create_wait_pt_and_return_fence(syncobj, point,
+							      &fence);
+		if (!fence)
+			return NULL;
+	}
+	if (wait_pt) {
+		int ret = 0;
+
+		if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) {
+			ret = wait_event_interruptible_timeout(syncobj->syncobj_timeline.wq,
+							       wait_pt->value <= syncobj->syncobj_timeline.signal_point,
+							       msecs_to_jiffies(10000)); /* wait 10s */
+			if (ret <= 0)
+				return NULL;
+		}
+		spin_lock(&syncobj->lock);
+		drm_syncobj_find_signal_pt_for_wait_pt(syncobj, wait_pt);
+		spin_unlock(&syncobj->lock);
+		return fence;
+	}
+	return NULL;
+}
+
  /**
   * drm_syncobj_find_fence - lookup and reference the fence in a sync object
   * @file_private: drm file private pointer
@@ -234,20 +541,46 @@ static int drm_syncobj_assign_null_handle(struct drm_syncobj *syncobj)
   * contains a reference to the fence, which must be released by calling
   * dma_fence_put().
   */
-int drm_syncobj_find_fence(struct drm_file *file_private,
-			   u32 handle, u64 point,
-			   struct dma_fence **fence)
+int drm_syncobj_search_fence(struct drm_syncobj *syncobj, u64 point,
+			     u64 flags, struct dma_fence **fence)
  {
-	struct drm_syncobj *syncobj = drm_syncobj_find(file_private, handle);
  	int ret = 0;
 
  	if (!syncobj)
  		return -ENOENT;
 
-	*fence = drm_syncobj_fence_get(syncobj);
+	drm_syncobj_timeline_garbage_collection(syncobj);
+	if (syncobj->type == DRM_SYNCOBJ_TYPE_NORMAL) {
+		/*NORMAL syncobj always wait on last pt */
+		u64 tail_pt_value = syncobj->syncobj_timeline.signal_point;
+
+		if (tail_pt_value == 0)
+			tail_pt_value += DRM_SYNCOBJ_NORMAL_POINT;
+		/* NORMAL syncobj doesn't care point value */
+		WARN_ON(point != 0);
+		*fence = drm_syncobj_timeline_point_get(syncobj, tail_pt_value,
+							flags);
+	} else if (syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+		*fence = drm_syncobj_timeline_point_get(syncobj, point,
+							flags);
+	} else {
+		DRM_ERROR("Don't support this type syncobj\n");
+		*fence = NULL;
+	}
  	if (!*fence) {
  		ret = -EINVAL;
  	}
+	return ret;
+}
+EXPORT_SYMBOL(drm_syncobj_search_fence);
+int drm_syncobj_find_fence(struct drm_file *file_private,
+			   u32 handle, u64 point,
+			   struct dma_fence **fence) {
+	struct drm_syncobj *syncobj = drm_syncobj_find(file_private, handle);
+
+	int ret = drm_syncobj_search_fence(syncobj, point,
+					DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT,
+					fence);
  	drm_syncobj_put(syncobj);
  	return ret;
  }
@@ -264,7 +597,7 @@ void drm_syncobj_free(struct kref *kref)
  	struct drm_syncobj *syncobj = container_of(kref,
  						   struct drm_syncobj,
  						   refcount);
-	drm_syncobj_replace_fence(syncobj, 0, NULL);
+	drm_syncobj_timeline_fini(syncobj, &syncobj->syncobj_timeline);
  	kfree(syncobj);
  }
  EXPORT_SYMBOL(drm_syncobj_free);
@@ -292,8 +625,12 @@ int drm_syncobj_create(struct drm_syncobj **out_syncobj, uint32_t flags,
  		return -ENOMEM;
 
  	kref_init(&syncobj->refcount);
-	INIT_LIST_HEAD(&syncobj->cb_list);
  	spin_lock_init(&syncobj->lock);
+	if (flags & DRM_SYNCOBJ_CREATE_TYPE_TIMELINE)
+		syncobj->type = DRM_SYNCOBJ_TYPE_TIMELINE;
+	else
+		syncobj->type = DRM_SYNCOBJ_TYPE_NORMAL;
+	drm_syncobj_timeline_init(syncobj, &syncobj->syncobj_timeline);
 
  	if (flags & DRM_SYNCOBJ_CREATE_SIGNALED) {
  		ret = drm_syncobj_assign_null_handle(syncobj);
@@ -651,7 +988,6 @@ struct syncobj_wait_entry {
  	struct task_struct *task;
  	struct dma_fence *fence;
  	struct dma_fence_cb fence_cb;
-	struct drm_syncobj_cb syncobj_cb;
  };
 
  static void syncobj_wait_fence_func(struct dma_fence *fence,
@@ -663,18 +999,6 @@ static void syncobj_wait_fence_func(struct dma_fence *fence,
  	wake_up_process(wait->task);
  }
 
-static void syncobj_wait_syncobj_func(struct drm_syncobj *syncobj,
-				      struct drm_syncobj_cb *cb)
-{
-	struct syncobj_wait_entry *wait =
-		container_of(cb, struct syncobj_wait_entry, syncobj_cb);
-
-	/* This happens inside the syncobj lock */
-	wait->fence = dma_fence_get(rcu_dereference_protected(syncobj->fence,
-							      lockdep_is_held(&syncobj->lock)));
-	wake_up_process(wait->task);
-}
-
  static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
  						  uint32_t count,
  						  uint32_t flags,
@@ -698,14 +1022,11 @@ static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
  	signaled_count = 0;
  	for (i = 0; i < count; ++i) {
  		entries[i].task = current;
-		entries[i].fence = drm_syncobj_fence_get(syncobjs[i]);
+		ret = drm_syncobj_search_fence(syncobjs[i], 0, 0,
+					       &entries[i].fence);
  		if (!entries[i].fence) {
-			if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) {
-				continue;
-			} else {
-				ret = -EINVAL;
-				goto cleanup_entries;
-			}
+			ret = -EINVAL;
+			goto cleanup_entries;
  		}
 
  		if (dma_fence_is_signaled(entries[i].fence)) {
@@ -733,15 +1054,6 @@ static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
  	 * fallthough and try a 0 timeout wait!
  	 */
 
-	if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) {
-		for (i = 0; i < count; ++i) {
-			drm_syncobj_fence_get_or_add_callback(syncobjs[i],
-							      &entries[i].fence,
-							      &entries[i].syncobj_cb,
-							      syncobj_wait_syncobj_func);
-		}
-	}
-
  	do {
  		set_current_state(TASK_INTERRUPTIBLE);
 
@@ -789,13 +1101,10 @@ static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
 
  cleanup_entries:
  	for (i = 0; i < count; ++i) {
-		if (entries[i].syncobj_cb.func)
-			drm_syncobj_remove_callback(syncobjs[i],
-						    &entries[i].syncobj_cb);
+		dma_fence_put(entries[i].fence);
  		if (entries[i].fence_cb.func)
  			dma_fence_remove_callback(entries[i].fence,
  						  &entries[i].fence_cb);
-		dma_fence_put(entries[i].fence);
  	}
  	kfree(entries);
 
@@ -970,12 +1279,21 @@ drm_syncobj_reset_ioctl(struct drm_device *dev, void *data,
  	if (ret < 0)
  		return ret;
 
-	for (i = 0; i < args->count_handles; i++)
-		drm_syncobj_replace_fence(syncobjs[i], 0, NULL);
-
+	for (i = 0; i < args->count_handles; i++) {
+		if (syncobjs[i]->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+			DRM_ERROR("timeline syncobj cannot reset!\n");
+			ret = -EINVAL;
+			goto out;
+		}
+		drm_syncobj_timeline_fini(syncobjs[i],
+					  &syncobjs[i]->syncobj_timeline);
+		drm_syncobj_timeline_init(syncobjs[i],
+					  &syncobjs[i]->syncobj_timeline);
+	}
+out:
  	drm_syncobj_array_free(syncobjs, args->count_handles);
 
-	return 0;
+	return ret;
  }
 
  int
drivers/gpu/drm/i915/i915_gem_execbuffer.c +3 -1
index 0a8d2d64f380..579e91a5858b 100644
@@ -2137,7 +2137,9 @@ await_fence_array(struct i915_execbuffer *eb,
  		if (!(flags & I915_EXEC_FENCE_WAIT))
  			continue;
 
-		fence = drm_syncobj_fence_get(syncobj);
+		drm_syncobj_search_fence(syncobj, 0,
+					 DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT,
+					 &fence);
  		if (!fence)
  			return -EINVAL;
 
include/drm/drm_syncobj.h +27 -51
index 425432b85a87..657c97dc25ec 100644
@@ -30,6 +30,25 @@
 
  struct drm_syncobj_cb;
 
+enum drm_syncobj_type {
+	DRM_SYNCOBJ_TYPE_NORMAL,
+	DRM_SYNCOBJ_TYPE_TIMELINE
+};
+
+struct drm_syncobj_timeline {
+	wait_queue_head_t	wq;
+	u64 timeline_context;
+	/**
+	 * @timeline: syncobj timeline
+	 */
+	u64 timeline;
+	u64 signal_point;
+
+
+	struct rb_root wait_pt_tree;
+	struct list_head signal_pt_list;
+};
+
  /**
   * struct drm_syncobj - sync object.
   *
@@ -41,19 +60,16 @@ struct drm_syncobj {
  	 */
  	struct kref refcount;
  	/**
-	 * @fence:
-	 * NULL or a pointer to the fence bound to this object.
-	 *
-	 * This field should not be used directly. Use drm_syncobj_fence_get()
-	 * and drm_syncobj_replace_fence() instead.
+	 * @type: indicate syncobj type
  	 */
-	struct dma_fence __rcu *fence;
+	enum drm_syncobj_type type;
  	/**
-	 * @cb_list: List of callbacks to call when the &fence gets replaced.
+	 * @syncobj_timeline: timeline
  	 */
-	struct list_head cb_list;
+	struct drm_syncobj_timeline syncobj_timeline;
+
  	/**
-	 * @lock: Protects &cb_list and write-locks &fence.
+	 * @lock: Protects timeline list and write-locks &fence.
  	 */
  	spinlock_t lock;
  	/**
@@ -62,25 +78,6 @@ struct drm_syncobj {
  	struct file *file;
  };
 
-typedef void (*drm_syncobj_func_t)(struct drm_syncobj *syncobj,
-				   struct drm_syncobj_cb *cb);
-
-/**
- * struct drm_syncobj_cb - callback for drm_syncobj_add_callback
- * @node: used by drm_syncob_add_callback to append this struct to
- *	  &drm_syncobj.cb_list
- * @func: drm_syncobj_func_t to call
- *
- * This struct will be initialized by drm_syncobj_add_callback, additional
- * data can be passed along by embedding drm_syncobj_cb in another struct.
- * The callback will get called the next time drm_syncobj_replace_fence is
- * called.
- */
-struct drm_syncobj_cb {
-	struct list_head node;
-	drm_syncobj_func_t func;
-};
-
  void drm_syncobj_free(struct kref *kref);
 
  /**
@@ -106,29 +103,6 @@ drm_syncobj_put(struct drm_syncobj *obj)
  	kref_put(&obj->refcount, drm_syncobj_free);
  }
 
-/**
- * drm_syncobj_fence_get - get a reference to a fence in a sync object
- * @syncobj: sync object.
- *
- * This acquires additional reference to &drm_syncobj.fence contained in @obj,
- * if not NULL. It is illegal to call this without already holding a reference.
- * No locks required.
- *
- * Returns:
- * Either the fence of @obj or NULL if there's none.
- */
-static inline struct dma_fence *
-drm_syncobj_fence_get(struct drm_syncobj *syncobj)
-{
-	struct dma_fence *fence;
-
-	rcu_read_lock();
-	fence = dma_fence_get_rcu_safe(&syncobj->fence);
-	rcu_read_unlock();
-
-	return fence;
-}
-
  struct drm_syncobj *drm_syncobj_find(struct drm_file *file_private,
  				     u32 handle);
  void drm_syncobj_replace_fence(struct drm_syncobj *syncobj, u64 point,
@@ -142,5 +116,7 @@ int drm_syncobj_create(struct drm_syncobj **out_syncobj, uint32_t flags,
  int drm_syncobj_get_handle(struct drm_file *file_private,
  			   struct drm_syncobj *syncobj, u32 *handle);
  int drm_syncobj_get_fd(struct drm_syncobj *syncobj, int *p_fd);
+int drm_syncobj_search_fence(struct drm_syncobj *syncobj, u64 point,
+			     u64 flags, struct dma_fence **fence);
 
  #endif
include/uapi/drm/drm.h +0 -0
index 300f336633f2..cebdb2541eb7 100644

include/uapi/drm/drm.h struct drm_prime_handle { · line 723

Quoted 7 lines
include/uapi/drm/drm.h +1 -0
@@ -717,6 +717,7 @@ struct drm_prime_handle {
  struct drm_syncobj_create {
  	__u32 handle;
  #define DRM_SYNCOBJ_CREATE_SIGNALED (1 << 0)
+#define DRM_SYNCOBJ_CREATE_TYPE_TIMELINE (1 << 1)
  	__u32 flags;
  };
 

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 4374aa17-c659-0c22-a487-c5e236690108-5C7GfCeVMHo@public.gmane.org

#8 lore raw

Ping? Have you seen my comments here?

Looks like you haven't addressed any of them in your last mail.

Christian.

Am 06.09.2018 um 09:25 schrieb Christian König:

Quoted 875 lines

Am 06.09.2018 um 08:25 schrieb Chunming Zhou:

Quoted 3 lines

VK_KHR_timeline_semaphore:
This extension introduces a new type of semaphore that has an integer payload identifying a point in a timeline. Such timeline semaphores support the

Drop the term semaphore here, better use syncobj.

Quoted 87 lines

following operations:
    * CPU query - A host operation that allows querying the payload of the timeline semaphore.
    * CPU wait - A host operation that allows a blocking wait for a timeline semaphore to reach a specified value.
    * Device wait - A device operation that allows waiting for a timeline semaphore to reach a specified value.
    * Device signal - A device operation that allows advancing the timeline semaphore to a specified value.

Since it's a timeline, that means the front time point(PT) always is signaled before the late PT. a. signal PT design:
Signal PT fence N depends on PT[N-1] fence and signal opertion fence, when PT[N] fence is signaled, the timeline will increase to value of PT[N].
b. wait PT design:
Wait PT fence is signaled by reaching timeline point value, when timeline is increasing, will compare wait PTs value with new timeline value, if PT value is lower than timeline value, then wait PT will be signaled, otherwise keep in list. semaphore wait operation can wait on any point of timeline, so need a RB tree to order them. And wait PT could ahead of signal PT, we need a sumission fence to perform that.

v2:
1. remove unused DRM_SYNCOBJ_CREATE_TYPE_NORMAL. (Christian)
2. move unexposed denitions to .c file. (Daniel Vetter)
3. split up the change to drm_syncobj_find_fence() in a separate patch. (Christian)
4. split up the change to drm_syncobj_replace_fence() in a separate patch.
5. drop the submission_fence implementation and instead use wait_event() for that. (Christian)
6. WARN_ON(point != 0) for NORMAL type syncobj case. (Daniel Vetter)

v3:
1. replace normal syncobj with timeline implemenation. (Vetter and Christian)
     a. normal syncobj signal op will create a signal PT to tail of signal pt list.
     b. normal syncobj wait op will create a wait pt with last signal point, and this wait PT is only signaled by related signal point PT.
2. many bug fix and clean up
3. stub fence moving is moved to other patch.

v4:
1. fix RB tree loop with while(node=rb_first(...)). (Christian)
2. fix syncobj lifecycle. (Christian)
3. only enable_signaling when there is wait_pt. (Christian)

Tested by ./deqp-vk -n dEQP-VK*semaphore* for normal syncobj

Cc Christian Konig <[email protected]>
Cc Dave Airlie <[email protected]>
Cc Daniel Rakos <[email protected]>
Cc Daniel Vetter <[email protected]>

---
  drivers/gpu/drm/drm_syncobj.c              | 516 +++++++++++++++++---- drivers/gpu/drm/i915/i915_gem_execbuffer.c |   4 +- include/drm/drm_syncobj.h                  |  78 ++-- include/uapi/drm/drm.h                     |   1 +

  4 files changed, 448 insertions(+), 151 deletions(-)
drivers/gpu/drm/drm_syncobj.c +0 -0
index e9ce623d049e..94b31de23858 100644
@@ -56,6 +56,9 @@
   #include "drm_internal.h"
   #include <drm/drm_syncobj.h>

  +/* merge normal syncobj to timeline syncobj, the point interval is 1 */

drivers/gpu/drm/drm_syncobj.c +2 -0
+#define DRM_SYNCOBJ_NORMAL_POINT 1
+

  struct drm_syncobj_stub_fence {
      struct dma_fence base;
      spinlock_t lock;

drivers/gpu/drm/drm_syncobj.c +11 -0
@@ -82,6 +85,18 @@ static const struct dma_fence_ops 
 drm_syncobj_stub_fence_ops = {
       .release = drm_syncobj_stub_fence_release,
   };
   +struct drm_syncobj_signal_pt {
+    struct dma_fence_array *base;
+    u64    value;
+    struct list_head list;
+};
+struct drm_syncobj_wait_pt {
+    struct drm_syncobj_stub_fence base;
+    struct dma_fence *signal_pt_fence;
+    struct dma_fence_cb wait_cb;
+    struct rb_node   node;
+    u64    value;
+};
     /**
    * drm_syncobj_find - lookup and reference a sync object.

drivers/gpu/drm/drm_syncobj.c struct drm_syncobj *drm_syncobj_find(struct · line 129

Quoted 8 lines
drivers/gpu/drm/drm_syncobj.c +2 -2
@@ -109,61 +124,238 @@ struct drm_syncobj *drm_syncobj_find(struct 
 drm_file *file_private,
   }
   EXPORT_SYMBOL(drm_syncobj_find);
   -static void drm_syncobj_add_callback_locked(struct drm_syncobj *syncobj,
-                        struct drm_syncobj_cb *cb,
-                        drm_syncobj_func_t func)
+static void drm_syncobj_timeline_init(struct drm_syncobj *syncobj,
+                      struct drm_syncobj_timeline *syncobj_timeline)

Since we merged timeline and singleton syncobj you can drop the extra _timeline_ part in the function name I think.

Quoted 73 lines
drivers/gpu/drm/drm_syncobj.c +41 -22
   {
-    cb->func = func;
-    list_add_tail(&cb->node, &syncobj->cb_list);
+    spin_lock(&syncobj->lock);
+    syncobj_timeline->timeline_context = dma_fence_context_alloc(1);
+    syncobj_timeline->timeline = 0;
+    syncobj_timeline->signal_point = 0;
+    init_waitqueue_head(&syncobj_timeline->wq);
+
+    syncobj_timeline->wait_pt_tree = RB_ROOT;
+    INIT_LIST_HEAD(&syncobj_timeline->signal_pt_list);
+    spin_unlock(&syncobj->lock);
   }
   -static int drm_syncobj_fence_get_or_add_callback(struct drm_syncobj *syncobj,
-                         struct dma_fence **fence,
-                         struct drm_syncobj_cb *cb,
-                         drm_syncobj_func_t func)
+static void drm_syncobj_timeline_fini(struct drm_syncobj *syncobj,
+                      struct drm_syncobj_timeline *syncobj_timeline)
   {
-    int ret;
-
-    *fence = drm_syncobj_fence_get(syncobj);
-    if (*fence)
-        return 1;
+    struct rb_node *node = NULL;
+    struct drm_syncobj_wait_pt *wait_pt = NULL;
+    struct drm_syncobj_signal_pt *signal_pt = NULL, *tmp;
         spin_lock(&syncobj->lock);
-    /* We've already tried once to get a fence and failed.  Now that we
-     * have the lock, try one more time just to be sure we don't add a
-     * callback when a fence has already been set.
-     */
-    if (syncobj->fence) {
-        *fence = 
 dma_fence_get(rcu_dereference_protected(syncobj->fence,
- lockdep_is_held(&syncobj->lock)));
-        ret = 1;
-    } else {
-        *fence = NULL;
-        drm_syncobj_add_callback_locked(syncobj, cb, func);
-        ret = 0;
+    if (!RB_EMPTY_ROOT(&syncobj_timeline->wait_pt_tree))
+        DRM_ERROR("There still are advanced wait_pts! pls check!");
+    while ((node = rb_first(&syncobj_timeline->wait_pt_tree))) {
+        wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node);
+        rb_erase(&wait_pt->node,
+             &syncobj_timeline->wait_pt_tree);
+        RB_CLEAR_NODE(&wait_pt->node);
+        if (!wait_pt->signal_pt_fence) {
+            /* if wait_pt still is advanced, fake signal it.
+             * generally shouldn't happens */
+            dma_fence_set_error(&wait_pt->base.base, -ENOENT);
+            dma_fence_signal(&wait_pt->base.base);
+        }
+        /* kfree(wait_pt) is excuted by fence put */
+        dma_fence_put(&wait_pt->base.base);
+    }
+    list_for_each_entry_safe(signal_pt, tmp,
+                 &syncobj_timeline->signal_pt_list, list) {
+        list_del(&signal_pt->list);
+        dma_fence_put(&signal_pt->base->base);
+        kfree(signal_pt);
       }
       spin_unlock(&syncobj->lock);
+}
   -    return ret;
+static void wait_pt_func(struct dma_fence *fence, struct 
 dma_fence_cb *cb)
+{
+    struct drm_syncobj_wait_pt *wait_pt =
+        container_of(cb, struct drm_syncobj_wait_pt, wait_cb);
+

drivers/gpu/drm/drm_syncobj.c struct drm_syncobj *drm_syncobj_find(struct drm_file *file_private, · line 187

Quoted 10 lines
drivers/gpu/drm/drm_syncobj.c +9 -0
+    dma_fence_signal(&wait_pt->base.base);
+    dma_fence_put(wait_pt->signal_pt_fence);
+    wait_pt->signal_pt_fence = NULL;
+    dma_fence_put(&wait_pt->base.base);
+}
+
+static void drm_syncobj_find_signal_pt_for_wait_pt(struct 
 drm_syncobj *syncobj,
+                           struct drm_syncobj_wait_pt *wait_pt)
+{

That whole approach still looks horrible complicated to me.

Especially the separation of signal and wait pt is completely unnecessary as far as I can see.

When a wait pt is requested we just need to search for the signal point which it will trigger.

Regards,
Christian.

Quoted 677 lines
drivers/gpu/drm/drm_syncobj.c +160 -5
+    struct drm_syncobj_timeline *timeline = &syncobj->syncobj_timeline;
+    struct drm_syncobj_signal_pt *signal_pt;
+    int ret;
+
+    if (wait_pt->signal_pt_fence) {
+        return;
+    } else if ((syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) &&
+           (wait_pt->value <= timeline->timeline)) {
+        dma_fence_signal(&wait_pt->base.base);
+        rb_erase(&wait_pt->node,
+             &timeline->wait_pt_tree);
+        RB_CLEAR_NODE(&wait_pt->node);
+        dma_fence_put(&wait_pt->base.base);
+        return;
+    }
+
+    list_for_each_entry(signal_pt, &timeline->signal_pt_list, list) {
+        if (wait_pt->value < signal_pt->value)
+            continue;
+        if ((syncobj->type == DRM_SYNCOBJ_TYPE_NORMAL) &&
+            (wait_pt->value != signal_pt->value))
+            continue;
+        wait_pt->signal_pt_fence = 
 dma_fence_get(&signal_pt->base->base);
+        ret = dma_fence_add_callback(wait_pt->signal_pt_fence,
+                         &wait_pt->wait_cb,
+                         wait_pt_func);
+        if (ret == -ENOENT) {
+            dma_fence_signal(&wait_pt->base.base);
+            dma_fence_put(wait_pt->signal_pt_fence);
+            wait_pt->signal_pt_fence = NULL;
+            rb_erase(&wait_pt->node,
+                 &timeline->wait_pt_tree);
+            RB_CLEAR_NODE(&wait_pt->node);
+            dma_fence_put(&wait_pt->base.base);
+        } else if (ret < 0) {
+            dma_fence_put(wait_pt->signal_pt_fence);
+            DRM_ERROR("add callback error!");
+        } else {
+            /* after adding callback, remove from rb tree */
+            rb_erase(&wait_pt->node,
+                 &timeline->wait_pt_tree);
+            RB_CLEAR_NODE(&wait_pt->node);
+        }
+        return;
+    }
+    /* signaled pt was released */
+    if (!wait_pt->signal_pt_fence && (wait_pt->value <=
+                      timeline->signal_point)) {
+        dma_fence_signal(&wait_pt->base.base);
+        rb_erase(&wait_pt->node,
+             &timeline->wait_pt_tree);
+        RB_CLEAR_NODE(&wait_pt->node);
+        dma_fence_put(&wait_pt->base.base);
+    }
   }
   -void drm_syncobj_add_callback(struct drm_syncobj *syncobj,
-                  struct drm_syncobj_cb *cb,
-                  drm_syncobj_func_t func)
+static int drm_syncobj_timeline_create_signal_pt(struct drm_syncobj 
 *syncobj,
+                         struct dma_fence *fence,
+                         u64 point)
   {
+    struct drm_syncobj_signal_pt *signal_pt =
+        kzalloc(sizeof(struct drm_syncobj_signal_pt), GFP_KERNEL);
+    struct drm_syncobj_signal_pt *tail_pt;
+    struct dma_fence **fences;
+    struct rb_node *node;
+    struct drm_syncobj_wait_pt *tail_wait_pt = NULL;
+    int num_fences = 0;
+    int ret = 0, i;
+
+    if (!signal_pt)
+        return -ENOMEM;
+    if (syncobj->syncobj_timeline.signal_point >= point) {
+        DRM_WARN("A later signal is ready!");
+        goto out;
+    }
+    if (!fence)
+        goto out;
+
+    fences = kmalloc_array(sizeof(void *), 2, GFP_KERNEL);
+    if (!fences)
+        goto out;
+    fences[num_fences++] = dma_fence_get(fence);
+    /* timeline syncobj must take this dependency */
+    if (syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+        spin_lock(&syncobj->lock);
+        if (!list_empty(&syncobj->syncobj_timeline.signal_pt_list)) {
+            tail_pt = 
 list_last_entry(&syncobj->syncobj_timeline.signal_pt_list,
+                          struct drm_syncobj_signal_pt, list);
+            fences[num_fences++] = dma_fence_get(&tail_pt->base->base);
+        }
+        spin_unlock(&syncobj->lock);
+    }
+    signal_pt->base = dma_fence_array_create(num_fences, fences,
+ syncobj->syncobj_timeline.timeline_context,
+                         point, false);
+    if (!signal_pt->base)
+        goto fail;
+
+    signal_pt->value = point;
       spin_lock(&syncobj->lock);
-    drm_syncobj_add_callback_locked(syncobj, cb, func);
+    syncobj->syncobj_timeline.signal_point = point;
+    INIT_LIST_HEAD(&signal_pt->list);
+    list_add_tail(&signal_pt->list, 
 &syncobj->syncobj_timeline.signal_pt_list);
+
+    /* check if there is advanced wait */
+    node = rb_last(&syncobj->syncobj_timeline.wait_pt_tree);
+    if (node)
+        tail_wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, 
 node);
+    if (tail_wait_pt && !tail_wait_pt->signal_pt_fence) {
+        for (node = rb_first(&syncobj->syncobj_timeline.wait_pt_tree);
+             node != NULL; node = rb_next(node)) {
+            struct drm_syncobj_wait_pt *wait_pt =
+                rb_entry(node, struct drm_syncobj_wait_pt,
+                     node);
+
+            drm_syncobj_find_signal_pt_for_wait_pt(syncobj,
+                                   wait_pt);
+        }
+    }
+
       spin_unlock(&syncobj->lock);
+    wake_up_all(&syncobj->syncobj_timeline.wq);
+
+    return 0;
+fail:
+    for (i = 0; i < num_fences; i++)
+        dma_fence_put(fences[i]);
+    kfree(fences);
+out:
+    kfree(signal_pt);
+    return ret;
   }
   -void drm_syncobj_remove_callback(struct drm_syncobj *syncobj,
-                 struct drm_syncobj_cb *cb)
+static void drm_syncobj_timeline_garbage_collection(struct 
 drm_syncobj *syncobj)
   {
+    struct drm_syncobj_timeline *timeline = &syncobj->syncobj_timeline;
+    struct rb_node *node;
+    struct drm_syncobj_wait_pt *wait_pt;
+    struct drm_syncobj_signal_pt *signal_pt, *tmp;
+
       spin_lock(&syncobj->lock);
-    list_del_init(&cb->node);
+    node = rb_first(&timeline->wait_pt_tree);
+    while (node) {
+        wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node);
+        if (!dma_fence_is_signaled(&wait_pt->base.base)) {
+            node = rb_next(node);
+            continue;
+        }
+        rb_erase(&wait_pt->node,
+             &timeline->wait_pt_tree);
+        RB_CLEAR_NODE(&wait_pt->node);
+        /* kfree(wait_pt) is excuted by fence put */
+        dma_fence_put(&wait_pt->base.base);
+        node = rb_first(&timeline->wait_pt_tree);
+    }
+    list_for_each_entry_safe(signal_pt, tmp,
+                 &timeline->signal_pt_list, list) {
+        if (dma_fence_is_signaled(&signal_pt->base->base)) {
+            timeline->timeline = signal_pt->value;
+            list_del(&signal_pt->list);
+            dma_fence_put(&signal_pt->base->base);
+            kfree(signal_pt);
+        } else {
+            /*signal_pt is in order in list, from small to big, so
+             * the later must not be signal either */
+            break;
+        }
+    }
+

      spin_unlock(&syncobj->lock);
  }
-
  /**
   * drm_syncobj_replace_fence - replace fence in a sync object.
   * @syncobj: Sync object to replace fence in

drivers/gpu/drm/drm_syncobj.c +21 -20
@@ -176,28 +368,29 @@ void drm_syncobj_replace_fence(struct 
 drm_syncobj *syncobj,
                      u64 point,
                      struct dma_fence *fence)
   {
-    struct dma_fence *old_fence;
-    struct drm_syncobj_cb *cur, *tmp;
-
-    if (fence)
-        dma_fence_get(fence);
-
-    spin_lock(&syncobj->lock);
-
-    old_fence = rcu_dereference_protected(syncobj->fence,
- lockdep_is_held(&syncobj->lock));
-    rcu_assign_pointer(syncobj->fence, fence);
-
-    if (fence != old_fence) {
-        list_for_each_entry_safe(cur, tmp, &syncobj->cb_list, node) {
-            list_del_init(&cur->node);
-            cur->func(syncobj, cur);
+    drm_syncobj_timeline_garbage_collection(syncobj);
+    if (syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+        if (fence)
+            drm_syncobj_timeline_create_signal_pt(syncobj, fence,
+                                  point);
+        return;
+    } else if (syncobj->type == DRM_SYNCOBJ_TYPE_NORMAL) {
+        u64 pt_value;
+
+        if (!fence) {
+            drm_syncobj_timeline_fini(syncobj,
+                          &syncobj->syncobj_timeline);
+            drm_syncobj_timeline_init(syncobj,
+                          &syncobj->syncobj_timeline);
+            return;
           }
+        pt_value = syncobj->syncobj_timeline.signal_point +
+            DRM_SYNCOBJ_NORMAL_POINT;
+        drm_syncobj_timeline_create_signal_pt(syncobj, fence, 
 pt_value);
+        return;
+    } else {
+        DRM_ERROR("the syncobj type isn't support\n");
       }
-
-    spin_unlock(&syncobj->lock);
-
-    dma_fence_put(old_fence);
   }

  EXPORT_SYMBOL(drm_syncobj_replace_fence);
  @@ -220,6 +413,120 @@ static int drm_syncobj_assign_null_handle(struct drm_syncobj *syncobj) return 0;
  }
  +static struct drm_syncobj_wait_pt *

drivers/gpu/drm/drm_syncobj.c +5 -0
+drm_syncobj_timeline_lookup_wait_pt_fence(struct drm_syncobj *syncobj,
+                      u64 point,
+                      struct dma_fence **fence)
+{
+    struct rb_node *node = 

syncobj->syncobj_timeline.wait_pt_tree.rb_node;
+    struct drm_syncobj_wait_pt *wait_pt = NULL;

drivers/gpu/drm/drm_syncobj.c +24 -0
+
+    spin_lock(&syncobj->lock);
+    while(node) {
+        int result;
+
+        wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node);
+        result = point - wait_pt->value;
+        if (result < 0) {
+            node = node->rb_left;
+        } else if (result > 0) {
+            node = node->rb_right;
+        } else {
+            if (fence)
+                *fence = dma_fence_get(&wait_pt->base.base);
+            break;
+        }
+    }
+    spin_unlock(&syncobj->lock);
+
+    return wait_pt;
+}
+
+static struct drm_syncobj_wait_pt *
+drm_syncobj_timeline_create_wait_pt_and_return_fence(struct 

drm_syncobj *syncobj,
+                             u64 point,
+                             struct dma_fence **fence) +{
+    struct drm_syncobj_wait_pt *wait_pt = kzalloc(sizeof(*wait_pt), +                              GFP_KERNEL);
+    struct drm_syncobj_stub_fence *base;
+    struct rb_node **new = &(syncobj->syncobj_timeline.wait_pt_tree.rb_node), *parent = NULL;

drivers/gpu/drm/drm_syncobj.c +13 -0
+
+    if (!wait_pt)
+        return NULL;
+    base = &wait_pt->base;
+    spin_lock_init(&base->lock);
+    dma_fence_init(&base->base,
+               &drm_syncobj_stub_fence_ops,
+               &base->lock,
+               syncobj->syncobj_timeline.timeline_context, point);
+    wait_pt->value = point;
+    wait_pt->signal_pt_fence = NULL;
+
+    /* wait pt must be in an order, so that we can easily lookup and 

signal
+     * it */
+    spin_lock(&syncobj->lock);
+    while(*new) {
+        struct drm_syncobj_wait_pt *this =
+            rb_entry(*new, struct drm_syncobj_wait_pt, node); +        int result = wait_pt->value - this->value; +
+        parent = *new;
+        if (result < 0)
+            new = &((*new)->rb_left);
+        else if (result > 0)
+            new = &((*new)->rb_right);
+        else
+            goto exist;
+    }
+    if (fence)
+        *fence = dma_fence_get(&wait_pt->base.base); +    rb_link_node(&wait_pt->node, parent, new);
+    rb_insert_color(&wait_pt->node, &syncobj->syncobj_timeline.wait_pt_tree);
+    spin_unlock(&syncobj->lock);
+    return wait_pt;

drivers/gpu/drm/drm_syncobj.c +10 -0
+exist:
+    spin_unlock(&syncobj->lock);
+    dma_fence_put(&wait_pt->base.base);
+    wait_pt = drm_syncobj_timeline_lookup_wait_pt_fence(syncobj, point,
+                                fence);
+    return wait_pt;
+}
+
+static struct dma_fence *
+drm_syncobj_timeline_point_get(struct drm_syncobj *syncobj, u64 

point, u64 flags)

drivers/gpu/drm/drm_syncobj.c +10 -0
+{
+    struct drm_syncobj_wait_pt *wait_pt;
+    struct dma_fence *fence = NULL;
+
+    /* check if the wait pt exists */
+    wait_pt = drm_syncobj_timeline_lookup_wait_pt_fence(syncobj, point,
+                                &fence);
+    if (!fence) {
+        /* This is a new wait pt, so create it */
+        wait_pt = 

drm_syncobj_timeline_create_wait_pt_and_return_fence(syncobj, point, +                                  &fence);
+        if (!fence)
+            return NULL;
+    }
+    if (wait_pt) {
+        int ret = 0;
+
+        if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) { +            ret = wait_event_interruptible_timeout(syncobj->syncobj_timeline.wq, +                                   wait_pt->value <= syncobj->syncobj_timeline.signal_point,
+                                   msecs_to_jiffies(10000)); /* wait 10s */
+            if (ret <= 0)
+                return NULL;
+        }
+        spin_lock(&syncobj->lock);
+        drm_syncobj_find_signal_pt_for_wait_pt(syncobj, wait_pt); +        spin_unlock(&syncobj->lock);
+        return fence;
+    }
+    return NULL;

drivers/gpu/drm/drm_syncobj.c +2 -0
+}
+

  /**
   * drm_syncobj_find_fence - lookup and reference the fence in a sync object
   * @file_private: drm file private pointer

drivers/gpu/drm/drm_syncobj.c +31 -4
@@ -234,20 +541,46 @@ static int 
 drm_syncobj_assign_null_handle(struct drm_syncobj *syncobj)
    * contains a reference to the fence, which must be released by calling
    * dma_fence_put().
    */
-int drm_syncobj_find_fence(struct drm_file *file_private,
-               u32 handle, u64 point,
-               struct dma_fence **fence)
+int drm_syncobj_search_fence(struct drm_syncobj *syncobj, u64 point,
+                 u64 flags, struct dma_fence **fence)
   {
-    struct drm_syncobj *syncobj = drm_syncobj_find(file_private, 
 handle);
       int ret = 0;
         if (!syncobj)
           return -ENOENT;
   -    *fence = drm_syncobj_fence_get(syncobj);
+    drm_syncobj_timeline_garbage_collection(syncobj);
+    if (syncobj->type == DRM_SYNCOBJ_TYPE_NORMAL) {
+        /*NORMAL syncobj always wait on last pt */
+        u64 tail_pt_value = syncobj->syncobj_timeline.signal_point;
+
+        if (tail_pt_value == 0)
+            tail_pt_value += DRM_SYNCOBJ_NORMAL_POINT;
+        /* NORMAL syncobj doesn't care point value */
+        WARN_ON(point != 0);
+        *fence = drm_syncobj_timeline_point_get(syncobj, tail_pt_value,
+                            flags);
+    } else if (syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+        *fence = drm_syncobj_timeline_point_get(syncobj, point,
+                            flags);
+    } else {
+        DRM_ERROR("Don't support this type syncobj\n");
+        *fence = NULL;
+    }
       if (!*fence) {
           ret = -EINVAL;
       }
+    return ret;
+}
+EXPORT_SYMBOL(drm_syncobj_search_fence);
+int drm_syncobj_find_fence(struct drm_file *file_private,
+               u32 handle, u64 point,
+               struct dma_fence **fence) {
+    struct drm_syncobj *syncobj = drm_syncobj_find(file_private, 
 handle);
+
+    int ret = drm_syncobj_search_fence(syncobj, point,
+                    DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT,
+                    fence);
       drm_syncobj_put(syncobj);

      return ret;
  }

drivers/gpu/drm/drm_syncobj.c +6 -14
@@ -264,7 +597,7 @@ void drm_syncobj_free(struct kref *kref)
       struct drm_syncobj *syncobj = container_of(kref,
                              struct drm_syncobj,
                              refcount);
-    drm_syncobj_replace_fence(syncobj, 0, NULL);
+    drm_syncobj_timeline_fini(syncobj, &syncobj->syncobj_timeline);
       kfree(syncobj);
   }
   EXPORT_SYMBOL(drm_syncobj_free);
@@ -292,8 +625,12 @@ int drm_syncobj_create(struct drm_syncobj 
 **out_syncobj, uint32_t flags,
           return -ENOMEM;
         kref_init(&syncobj->refcount);
-    INIT_LIST_HEAD(&syncobj->cb_list);
       spin_lock_init(&syncobj->lock);
+    if (flags & DRM_SYNCOBJ_CREATE_TYPE_TIMELINE)
+        syncobj->type = DRM_SYNCOBJ_TYPE_TIMELINE;
+    else
+        syncobj->type = DRM_SYNCOBJ_TYPE_NORMAL;
+    drm_syncobj_timeline_init(syncobj, &syncobj->syncobj_timeline);
         if (flags & DRM_SYNCOBJ_CREATE_SIGNALED) {
           ret = drm_syncobj_assign_null_handle(syncobj);
@@ -651,7 +988,6 @@ struct syncobj_wait_entry {
       struct task_struct *task;
       struct dma_fence *fence;
       struct dma_fence_cb fence_cb;
-    struct drm_syncobj_cb syncobj_cb;
   };
     static void syncobj_wait_fence_func(struct dma_fence *fence,
@@ -663,18 +999,6 @@ static void syncobj_wait_fence_func(struct 
 dma_fence *fence,
       wake_up_process(wait->task);
   }
   -static void syncobj_wait_syncobj_func(struct drm_syncobj *syncobj,
-                      struct drm_syncobj_cb *cb)
-{
-    struct syncobj_wait_entry *wait =
-        container_of(cb, struct syncobj_wait_entry, syncobj_cb);
-
-    /* This happens inside the syncobj lock */
-    wait->fence = 
 dma_fence_get(rcu_dereference_protected(syncobj->fence,
- lockdep_is_held(&syncobj->lock)));
-    wake_up_process(wait->task);
-}
-
   static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,

                            uint32_t count,
                            uint32_t flags,

drivers/gpu/drm/drm_syncobj.c +4 -15
@@ -698,14 +1022,11 @@ static signed long 
 drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
       signaled_count = 0;
       for (i = 0; i < count; ++i) {
           entries[i].task = current;
-        entries[i].fence = drm_syncobj_fence_get(syncobjs[i]);
+        ret = drm_syncobj_search_fence(syncobjs[i], 0, 0,
+                           &entries[i].fence);
           if (!entries[i].fence) {
-            if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) {
-                continue;
-            } else {
-                ret = -EINVAL;
-                goto cleanup_entries;
-            }
+            ret = -EINVAL;
+            goto cleanup_entries;
           }
             if (dma_fence_is_signaled(entries[i].fence)) {
@@ -733,15 +1054,6 @@ static signed long 
 drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
        * fallthough and try a 0 timeout wait!
        */
   -    if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) {
-        for (i = 0; i < count; ++i) {
-            drm_syncobj_fence_get_or_add_callback(syncobjs[i],
-                                  &entries[i].fence,
-                                  &entries[i].syncobj_cb,
-                                  syncobj_wait_syncobj_func);
-        }
-    }
-
       do {
           set_current_state(TASK_INTERRUPTIBLE);

  @@ -789,13 +1101,10 @@ static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs, cleanup_entries:
      for (i = 0; i < count; ++i) {
-        if (entries[i].syncobj_cb.func)
-            drm_syncobj_remove_callback(syncobjs[i],
-                            &entries[i].syncobj_cb);
+        dma_fence_put(entries[i].fence);
          if (entries[i].fence_cb.func)
              dma_fence_remove_callback(entries[i].fence,
                            &entries[i].fence_cb);
-        dma_fence_put(entries[i].fence);
      }
      kfree(entries);
  @@ -970,12 +1279,21 @@ drm_syncobj_reset_ioctl(struct drm_device *dev, void *data,
      if (ret < 0)
          return ret;
  -    for (i = 0; i < args->count_handles; i++)
-        drm_syncobj_replace_fence(syncobjs[i], 0, NULL);

drivers/gpu/drm/drm_syncobj.c +12 -1
-
+    for (i = 0; i < args->count_handles; i++) {
+        if (syncobjs[i]->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+            DRM_ERROR("timeline syncobj cannot reset!\n");
+            ret = -EINVAL;
+            goto out;
+        }
+        drm_syncobj_timeline_fini(syncobjs[i],
+                      &syncobjs[i]->syncobj_timeline);
+        drm_syncobj_timeline_init(syncobjs[i],
+                      &syncobjs[i]->syncobj_timeline);
+    }
+out:

      drm_syncobj_array_free(syncobjs, args->count_handles); -    return 0;
+    return ret;
  }
    int

drivers/gpu/drm/i915/i915_gem_execbuffer.c +3 -0
index 0a8d2d64f380..579e91a5858b 100644
@@ -2137,7 +2137,9 @@ await_fence_array(struct i915_execbuffer *eb,
           if (!(flags & I915_EXEC_FENCE_WAIT))
               continue;
   -        fence = drm_syncobj_fence_get(syncobj);
+        drm_syncobj_search_fence(syncobj, 0,
+                     DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT,
+                     &fence);
           if (!fence)
               return -EINVAL;

  diff --git a/include/drm/drm_syncobj.h b/include/drm/drm_syncobj.h index 425432b85a87..657c97dc25ec 100644

include/drm/drm_syncobj.h +20 -7
@@ -30,6 +30,25 @@
     struct drm_syncobj_cb;
   +enum drm_syncobj_type {
+    DRM_SYNCOBJ_TYPE_NORMAL,
+    DRM_SYNCOBJ_TYPE_TIMELINE
+};
+
+struct drm_syncobj_timeline {
+    wait_queue_head_t    wq;
+    u64 timeline_context;
+    /**
+     * @timeline: syncobj timeline
+     */
+    u64 timeline;
+    u64 signal_point;
+
+
+    struct rb_root wait_pt_tree;
+    struct list_head signal_pt_list;
+};
+
   /**
    * struct drm_syncobj - sync object.
    *
@@ -41,19 +60,16 @@ struct drm_syncobj {
        */
       struct kref refcount;
       /**
-     * @fence:
-     * NULL or a pointer to the fence bound to this object.
-     *
-     * This field should not be used directly. Use 
 drm_syncobj_fence_get()
-     * and drm_syncobj_replace_fence() instead.
+     * @type: indicate syncobj type
        */
-    struct dma_fence __rcu *fence;
+    enum drm_syncobj_type type;
       /**
-     * @cb_list: List of callbacks to call when the &fence gets 

replaced.
+     * @syncobj_timeline: timeline
       */
-    struct list_head cb_list;
+    struct drm_syncobj_timeline syncobj_timeline; +
      /**
-     * @lock: Protects &cb_list and write-locks &fence. +     * @lock: Protects timeline list and write-locks &fence. */
      spinlock_t lock;
      /**

include/drm/drm_syncobj.h +0 -9
@@ -62,25 +78,6 @@ struct drm_syncobj {
       struct file *file;
   };
   -typedef void (*drm_syncobj_func_t)(struct drm_syncobj *syncobj,
-                   struct drm_syncobj_cb *cb);
-
-/**
- * struct drm_syncobj_cb - callback for drm_syncobj_add_callback
- * @node: used by drm_syncob_add_callback to append this struct to
- *      &drm_syncobj.cb_list
- * @func: drm_syncobj_func_t to call
- *
- * This struct will be initialized by drm_syncobj_add_callback, 

additional
- * data can be passed along by embedding drm_syncobj_cb in another struct.
- * The callback will get called the next time drm_syncobj_replace_fence is
- * called.
- */

include/drm/drm_syncobj.h +0 -5
-struct drm_syncobj_cb {
-    struct list_head node;
-    drm_syncobj_func_t func;
-};
-

  void drm_syncobj_free(struct kref *kref);
    /**

include/drm/drm_syncobj.h +0 -0
@@ -106,29 +103,6 @@ drm_syncobj_put(struct drm_syncobj *obj)
       kref_put(&obj->refcount, drm_syncobj_free);
   }

  -/**
- * drm_syncobj_fence_get - get a reference to a fence in a sync object
- * @syncobj: sync object.
- *
- * This acquires additional reference to &drm_syncobj.fence contained in @obj,
- * if not NULL. It is illegal to call this without already holding a reference.
- * No locks required.
- *
- * Returns:
- * Either the fence of @obj or NULL if there's none.
- */

include/drm/drm_syncobj.h +0 -12
-static inline struct dma_fence *
-drm_syncobj_fence_get(struct drm_syncobj *syncobj)
-{
-    struct dma_fence *fence;
-
-    rcu_read_lock();
-    fence = dma_fence_get_rcu_safe(&syncobj->fence);
-    rcu_read_unlock();
-
-    return fence;
-}
-

  struct drm_syncobj *drm_syncobj_find(struct drm_file *file_private, u32 handle);
  void drm_syncobj_replace_fence(struct drm_syncobj *syncobj, u64 point,

include/drm/drm_syncobj.h +2 -0
@@ -142,5 +116,7 @@ int drm_syncobj_create(struct drm_syncobj 
 **out_syncobj, uint32_t flags,
   int drm_syncobj_get_handle(struct drm_file *file_private,
                  struct drm_syncobj *syncobj, u32 *handle);
   int drm_syncobj_get_fd(struct drm_syncobj *syncobj, int *p_fd);
+int drm_syncobj_search_fence(struct drm_syncobj *syncobj, u64 point,
+                 u64 flags, struct dma_fence **fence);
     #endif
include/uapi/drm/drm.h +0 -1
index 300f336633f2..cebdb2541eb7 100644
--- a/include/uapi/drm/drm.h
Quoted 9 lines

include/uapi/drm/drm.h struct drm_prime_handle { · line 722

Quoted 8 lines

+++ b/include/uapi/drm/drm.h

include/uapi/drm/drm.h +1 -0
@@ -717,6 +717,7 @@ struct drm_prime_handle {
   struct drm_syncobj_create {
       __u32 handle;
   #define DRM_SYNCOBJ_CREATE_SIGNALED (1 << 0)
+#define DRM_SYNCOBJ_CREATE_TYPE_TIMELINE (1 << 1)
       __u32 flags;
   };

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): b69c0d93-5e46-fc25-2545-e2e1de18295e-5C7GfCeVMHo@public.gmane.org

#10 lore raw
Quoted 1 line · Christian König

Ping? Have you seen my comments here?

Sorry, I didn't see this reply.  inline...

Quoted 12 lines

Looks like you haven't addressed any of them in your last mail.

Christian.

Am 06.09.2018 um 09:25 schrieb Christian König:

Quoted 6 lines

Am 06.09.2018 um 08:25 schrieb Chunming Zhou:

Quoted 3 lines

VK_KHR_timeline_semaphore:
This extension introduces a new type of semaphore that has an integer payload identifying a point in a timeline. Such timeline semaphores support the

Drop the term semaphore here, better use syncobj.

This is from VK_KHR_timeline_semaphore extension describe, not my invention, I just quote it. In kernel side, we call syncobj, in UMD, they still call semaphore.

Quoted 91 lines
Quoted 91 lines
Quoted 91 lines

following operations:
    * CPU query - A host operation that allows querying the payload of the timeline semaphore.
    * CPU wait - A host operation that allows a blocking wait for a timeline semaphore to reach a specified value.
    * Device wait - A device operation that allows waiting for a timeline semaphore to reach a specified value.
    * Device signal - A device operation that allows advancing the timeline semaphore to a specified value.

Since it's a timeline, that means the front time point(PT) always is signaled before the late PT. a. signal PT design:
Signal PT fence N depends on PT[N-1] fence and signal opertion fence, when PT[N] fence is signaled, the timeline will increase to value of PT[N].
b. wait PT design:
Wait PT fence is signaled by reaching timeline point value, when timeline is increasing, will compare wait PTs value with new timeline value, if PT value is lower than timeline value, then wait PT will be signaled, otherwise keep in list. semaphore wait operation can wait on any point of timeline, so need a RB tree to order them. And wait PT could ahead of signal PT, we need a sumission fence to perform that.

v2:
1. remove unused DRM_SYNCOBJ_CREATE_TYPE_NORMAL. (Christian)
2. move unexposed denitions to .c file. (Daniel Vetter)
3. split up the change to drm_syncobj_find_fence() in a separate patch. (Christian)
4. split up the change to drm_syncobj_replace_fence() in a separate patch.
5. drop the submission_fence implementation and instead use wait_event() for that. (Christian)
6. WARN_ON(point != 0) for NORMAL type syncobj case. (Daniel Vetter)

v3:
1. replace normal syncobj with timeline implemenation. (Vetter and Christian)
     a. normal syncobj signal op will create a signal PT to tail of signal pt list.
     b. normal syncobj wait op will create a wait pt with last signal point, and this wait PT is only signaled by related signal point PT.
2. many bug fix and clean up
3. stub fence moving is moved to other patch.

v4:
1. fix RB tree loop with while(node=rb_first(...)). (Christian)
2. fix syncobj lifecycle. (Christian)
3. only enable_signaling when there is wait_pt. (Christian)

Tested by ./deqp-vk -n dEQP-VK*semaphore* for normal syncobj

Cc Christian Konig <[email protected]>
Cc Dave Airlie <[email protected]>
Cc Daniel Rakos <[email protected]>
Cc Daniel Vetter <[email protected]>

---
  drivers/gpu/drm/drm_syncobj.c              | 516
+++++++++++++++++----
  drivers/gpu/drm/i915/i915_gem_execbuffer.c |   4 +- include/drm/drm_syncobj.h                  |  78 ++-- include/uapi/drm/drm.h                     |   1 +

  4 files changed, 448 insertions(+), 151 deletions(-)
drivers/gpu/drm/drm_syncobj.c +0 -0
index e9ce623d049e..94b31de23858 100644
@@ -56,6 +56,9 @@
   #include "drm_internal.h"
   #include <drm/drm_syncobj.h>

  +/* merge normal syncobj to timeline syncobj, the point interval is 1 */

drivers/gpu/drm/drm_syncobj.c +2 -0
+#define DRM_SYNCOBJ_NORMAL_POINT 1
+

  struct drm_syncobj_stub_fence {
      struct dma_fence base;
      spinlock_t lock;

drivers/gpu/drm/drm_syncobj.c +11 -0
@@ -82,6 +85,18 @@ static const struct dma_fence_ops 
 drm_syncobj_stub_fence_ops = {
       .release = drm_syncobj_stub_fence_release,
   };
   +struct drm_syncobj_signal_pt {
+    struct dma_fence_array *base;
+    u64    value;
+    struct list_head list;
+};
+struct drm_syncobj_wait_pt {
+    struct drm_syncobj_stub_fence base;
+    struct dma_fence *signal_pt_fence;
+    struct dma_fence_cb wait_cb;
+    struct rb_node   node;
+    u64    value;
+};
     /**
    * drm_syncobj_find - lookup and reference a sync object.
@@ -109,61 +124,238 @@ struct drm_syncobj *drm_syncobj_find(struct 
 drm_file *file_private,
   }
Quoted 8 lines
Quoted 8 lines

drivers/gpu/drm/drm_syncobj.c struct drm_syncobj *drm_syncobj_find(struct drm_file *file_private, · line 128

Quoted 6 lines
drivers/gpu/drm/drm_syncobj.c +2 -2
   EXPORT_SYMBOL(drm_syncobj_find);
   -static void drm_syncobj_add_callback_locked(struct drm_syncobj *syncobj,
-                        struct drm_syncobj_cb *cb,
-                        drm_syncobj_func_t func)
+static void drm_syncobj_timeline_init(struct drm_syncobj *syncobj,
+                      struct drm_syncobj_timeline *syncobj_timeline)

Since we merged timeline and singleton syncobj you can drop the extra _timeline_ part in the function name I think.

Will try in v5.

Quoted 78 lines
Quoted 78 lines
Quoted 78 lines
drivers/gpu/drm/drm_syncobj.c +45 -22
   {
-    cb->func = func;
-    list_add_tail(&cb->node, &syncobj->cb_list);
+    spin_lock(&syncobj->lock);
+    syncobj_timeline->timeline_context = dma_fence_context_alloc(1);
+    syncobj_timeline->timeline = 0;
+    syncobj_timeline->signal_point = 0;
+    init_waitqueue_head(&syncobj_timeline->wq);
+
+    syncobj_timeline->wait_pt_tree = RB_ROOT;
+    INIT_LIST_HEAD(&syncobj_timeline->signal_pt_list);
+    spin_unlock(&syncobj->lock);
   }
   -static int drm_syncobj_fence_get_or_add_callback(struct drm_syncobj *syncobj,
-                         struct dma_fence **fence,
-                         struct drm_syncobj_cb *cb,
-                         drm_syncobj_func_t func)
+static void drm_syncobj_timeline_fini(struct drm_syncobj *syncobj,
+                      struct drm_syncobj_timeline *syncobj_timeline)
   {
-    int ret;
-
-    *fence = drm_syncobj_fence_get(syncobj);
-    if (*fence)
-        return 1;
+    struct rb_node *node = NULL;
+    struct drm_syncobj_wait_pt *wait_pt = NULL;
+    struct drm_syncobj_signal_pt *signal_pt = NULL, *tmp;
         spin_lock(&syncobj->lock);
-    /* We've already tried once to get a fence and failed. Now that we
-     * have the lock, try one more time just to be sure we don't add a
-     * callback when a fence has already been set.
-     */
-    if (syncobj->fence) {
-        *fence = 
 dma_fence_get(rcu_dereference_protected(syncobj->fence,
- lockdep_is_held(&syncobj->lock)));
-        ret = 1;
-    } else {
-        *fence = NULL;
-        drm_syncobj_add_callback_locked(syncobj, cb, func);
-        ret = 0;
+    if (!RB_EMPTY_ROOT(&syncobj_timeline->wait_pt_tree))
+        DRM_ERROR("There still are advanced wait_pts! pls check!");
+    while ((node = rb_first(&syncobj_timeline->wait_pt_tree))) {
+        wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node);
+        rb_erase(&wait_pt->node,
+             &syncobj_timeline->wait_pt_tree);
+        RB_CLEAR_NODE(&wait_pt->node);
+        if (!wait_pt->signal_pt_fence) {
+            /* if wait_pt still is advanced, fake signal it.
+             * generally shouldn't happens */
+            dma_fence_set_error(&wait_pt->base.base, -ENOENT);
+            dma_fence_signal(&wait_pt->base.base);
+        }
+        /* kfree(wait_pt) is excuted by fence put */
+        dma_fence_put(&wait_pt->base.base);
+    }
+    list_for_each_entry_safe(signal_pt, tmp,
+                 &syncobj_timeline->signal_pt_list, list) {
+        list_del(&signal_pt->list);
+        dma_fence_put(&signal_pt->base->base);
+        kfree(signal_pt);
       }
       spin_unlock(&syncobj->lock);
+}
   -    return ret;
+static void wait_pt_func(struct dma_fence *fence, struct 
 dma_fence_cb *cb)
+{
+    struct drm_syncobj_wait_pt *wait_pt =
+        container_of(cb, struct drm_syncobj_wait_pt, wait_cb);
+
+    dma_fence_signal(&wait_pt->base.base);
+    dma_fence_put(wait_pt->signal_pt_fence);
+    wait_pt->signal_pt_fence = NULL;
+    dma_fence_put(&wait_pt->base.base);
Quoted 8 lines
Quoted 8 lines

drivers/gpu/drm/drm_syncobj.c struct drm_syncobj *drm_syncobj_find(struct drm_file *file_private, · line 187

Quoted 6 lines
drivers/gpu/drm/drm_syncobj.c +5 -0
+}
+
+static void drm_syncobj_find_signal_pt_for_wait_pt(struct 
 drm_syncobj *syncobj,
+                           struct drm_syncobj_wait_pt *wait_pt)
+{

That whole approach still looks horrible complicated to me.

It's already very close to what you said before.

Quoted 3 lines
Quoted 3 lines

Especially the separation of signal and wait pt is completely unnecessary as far as I can see. When a wait pt is requested we just need to search for the signal point which it will trigger.

Yeah, I tried this, but when I implement cpu wait ioctl on specific point, we need a advanced wait pt fence, otherwise, we could still need old syncobj cb.

Thanks,
David Zhou

Quoted 697 lines
Quoted 687 lines

Regards,
Christian.

Quoted 683 lines
drivers/gpu/drm/drm_syncobj.c +160 -5
+    struct drm_syncobj_timeline *timeline = 
 &syncobj->syncobj_timeline;
+    struct drm_syncobj_signal_pt *signal_pt;
+    int ret;
+
+    if (wait_pt->signal_pt_fence) {
+        return;
+    } else if ((syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) &&
+           (wait_pt->value <= timeline->timeline)) {
+        dma_fence_signal(&wait_pt->base.base);
+        rb_erase(&wait_pt->node,
+             &timeline->wait_pt_tree);
+        RB_CLEAR_NODE(&wait_pt->node);
+        dma_fence_put(&wait_pt->base.base);
+        return;
+    }
+
+    list_for_each_entry(signal_pt, &timeline->signal_pt_list, list) {
+        if (wait_pt->value < signal_pt->value)
+            continue;
+        if ((syncobj->type == DRM_SYNCOBJ_TYPE_NORMAL) &&
+            (wait_pt->value != signal_pt->value))
+            continue;
+        wait_pt->signal_pt_fence = 
 dma_fence_get(&signal_pt->base->base);
+        ret = dma_fence_add_callback(wait_pt->signal_pt_fence,
+                         &wait_pt->wait_cb,
+                         wait_pt_func);
+        if (ret == -ENOENT) {
+            dma_fence_signal(&wait_pt->base.base);
+            dma_fence_put(wait_pt->signal_pt_fence);
+            wait_pt->signal_pt_fence = NULL;
+            rb_erase(&wait_pt->node,
+                 &timeline->wait_pt_tree);
+            RB_CLEAR_NODE(&wait_pt->node);
+            dma_fence_put(&wait_pt->base.base);
+        } else if (ret < 0) {
+            dma_fence_put(wait_pt->signal_pt_fence);
+            DRM_ERROR("add callback error!");
+        } else {
+            /* after adding callback, remove from rb tree */
+            rb_erase(&wait_pt->node,
+                 &timeline->wait_pt_tree);
+            RB_CLEAR_NODE(&wait_pt->node);
+        }
+        return;
+    }
+    /* signaled pt was released */
+    if (!wait_pt->signal_pt_fence && (wait_pt->value <=
+                      timeline->signal_point)) {
+        dma_fence_signal(&wait_pt->base.base);
+        rb_erase(&wait_pt->node,
+             &timeline->wait_pt_tree);
+        RB_CLEAR_NODE(&wait_pt->node);
+        dma_fence_put(&wait_pt->base.base);
+    }
   }
   -void drm_syncobj_add_callback(struct drm_syncobj *syncobj,
-                  struct drm_syncobj_cb *cb,
-                  drm_syncobj_func_t func)
+static int drm_syncobj_timeline_create_signal_pt(struct drm_syncobj 
 *syncobj,
+                         struct dma_fence *fence,
+                         u64 point)
   {
+    struct drm_syncobj_signal_pt *signal_pt =
+        kzalloc(sizeof(struct drm_syncobj_signal_pt), GFP_KERNEL);
+    struct drm_syncobj_signal_pt *tail_pt;
+    struct dma_fence **fences;
+    struct rb_node *node;
+    struct drm_syncobj_wait_pt *tail_wait_pt = NULL;
+    int num_fences = 0;
+    int ret = 0, i;
+
+    if (!signal_pt)
+        return -ENOMEM;
+    if (syncobj->syncobj_timeline.signal_point >= point) {
+        DRM_WARN("A later signal is ready!");
+        goto out;
+    }
+    if (!fence)
+        goto out;
+
+    fences = kmalloc_array(sizeof(void *), 2, GFP_KERNEL);
+    if (!fences)
+        goto out;
+    fences[num_fences++] = dma_fence_get(fence);
+    /* timeline syncobj must take this dependency */
+    if (syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+        spin_lock(&syncobj->lock);
+        if (!list_empty(&syncobj->syncobj_timeline.signal_pt_list)) {
+            tail_pt = 
 list_last_entry(&syncobj->syncobj_timeline.signal_pt_list,
+                          struct drm_syncobj_signal_pt, list);
+            fences[num_fences++] = 
 dma_fence_get(&tail_pt->base->base);
+        }
+        spin_unlock(&syncobj->lock);
+    }
+    signal_pt->base = dma_fence_array_create(num_fences, fences,
+ syncobj->syncobj_timeline.timeline_context,
+                         point, false);
+    if (!signal_pt->base)
+        goto fail;
+
+    signal_pt->value = point;
       spin_lock(&syncobj->lock);
-    drm_syncobj_add_callback_locked(syncobj, cb, func);
+    syncobj->syncobj_timeline.signal_point = point;
+    INIT_LIST_HEAD(&signal_pt->list);
+    list_add_tail(&signal_pt->list, 
 &syncobj->syncobj_timeline.signal_pt_list);
+
+    /* check if there is advanced wait */
+    node = rb_last(&syncobj->syncobj_timeline.wait_pt_tree);
+    if (node)
+        tail_wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, 
 node);
+    if (tail_wait_pt && !tail_wait_pt->signal_pt_fence) {
+        for (node = rb_first(&syncobj->syncobj_timeline.wait_pt_tree);
+             node != NULL; node = rb_next(node)) {
+            struct drm_syncobj_wait_pt *wait_pt =
+                rb_entry(node, struct drm_syncobj_wait_pt,
+                     node);
+
+            drm_syncobj_find_signal_pt_for_wait_pt(syncobj,
+                                   wait_pt);
+        }
+    }
+
       spin_unlock(&syncobj->lock);
+    wake_up_all(&syncobj->syncobj_timeline.wq);
+
+    return 0;
+fail:
+    for (i = 0; i < num_fences; i++)
+        dma_fence_put(fences[i]);
+    kfree(fences);
+out:
+    kfree(signal_pt);
+    return ret;
   }
   -void drm_syncobj_remove_callback(struct drm_syncobj *syncobj,
-                 struct drm_syncobj_cb *cb)
+static void drm_syncobj_timeline_garbage_collection(struct 
 drm_syncobj *syncobj)
   {
+    struct drm_syncobj_timeline *timeline = 
 &syncobj->syncobj_timeline;
+    struct rb_node *node;
+    struct drm_syncobj_wait_pt *wait_pt;
+    struct drm_syncobj_signal_pt *signal_pt, *tmp;
+
       spin_lock(&syncobj->lock);
-    list_del_init(&cb->node);
+    node = rb_first(&timeline->wait_pt_tree);
+    while (node) {
+        wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node);
+        if (!dma_fence_is_signaled(&wait_pt->base.base)) {
+            node = rb_next(node);
+            continue;
+        }
+        rb_erase(&wait_pt->node,
+             &timeline->wait_pt_tree);
+        RB_CLEAR_NODE(&wait_pt->node);
+        /* kfree(wait_pt) is excuted by fence put */
+        dma_fence_put(&wait_pt->base.base);
+        node = rb_first(&timeline->wait_pt_tree);
+    }
+    list_for_each_entry_safe(signal_pt, tmp,
+                 &timeline->signal_pt_list, list) {
+        if (dma_fence_is_signaled(&signal_pt->base->base)) {
+            timeline->timeline = signal_pt->value;
+            list_del(&signal_pt->list);
+            dma_fence_put(&signal_pt->base->base);
+            kfree(signal_pt);
+        } else {
+            /*signal_pt is in order in list, from small to big, so
+             * the later must not be signal either */
+            break;
+        }
+    }
+

      spin_unlock(&syncobj->lock);
  }
-
  /**
   * drm_syncobj_replace_fence - replace fence in a sync object.
   * @syncobj: Sync object to replace fence in

drivers/gpu/drm/drm_syncobj.c +21 -20
@@ -176,28 +368,29 @@ void drm_syncobj_replace_fence(struct 
 drm_syncobj *syncobj,
                      u64 point,
                      struct dma_fence *fence)
   {
-    struct dma_fence *old_fence;
-    struct drm_syncobj_cb *cur, *tmp;
-
-    if (fence)
-        dma_fence_get(fence);
-
-    spin_lock(&syncobj->lock);
-
-    old_fence = rcu_dereference_protected(syncobj->fence,
- lockdep_is_held(&syncobj->lock));
-    rcu_assign_pointer(syncobj->fence, fence);
-
-    if (fence != old_fence) {
-        list_for_each_entry_safe(cur, tmp, &syncobj->cb_list, node) {
-            list_del_init(&cur->node);
-            cur->func(syncobj, cur);
+    drm_syncobj_timeline_garbage_collection(syncobj);
+    if (syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+        if (fence)
+            drm_syncobj_timeline_create_signal_pt(syncobj, fence,
+                                  point);
+        return;
+    } else if (syncobj->type == DRM_SYNCOBJ_TYPE_NORMAL) {
+        u64 pt_value;
+
+        if (!fence) {
+            drm_syncobj_timeline_fini(syncobj,
+                          &syncobj->syncobj_timeline);
+            drm_syncobj_timeline_init(syncobj,
+                          &syncobj->syncobj_timeline);
+            return;
           }
+        pt_value = syncobj->syncobj_timeline.signal_point +
+            DRM_SYNCOBJ_NORMAL_POINT;
+        drm_syncobj_timeline_create_signal_pt(syncobj, fence, 
 pt_value);
+        return;
+    } else {
+        DRM_ERROR("the syncobj type isn't support\n");
       }
-
-    spin_unlock(&syncobj->lock);
-
-    dma_fence_put(old_fence);
   }

  EXPORT_SYMBOL(drm_syncobj_replace_fence);
  @@ -220,6 +413,120 @@ static int drm_syncobj_assign_null_handle(struct drm_syncobj *syncobj) return 0;
  }
  +static struct drm_syncobj_wait_pt *

drivers/gpu/drm/drm_syncobj.c +5 -0
+drm_syncobj_timeline_lookup_wait_pt_fence(struct drm_syncobj *syncobj,
+                      u64 point,
+                      struct dma_fence **fence)
+{
+    struct rb_node *node = 

syncobj->syncobj_timeline.wait_pt_tree.rb_node;
+    struct drm_syncobj_wait_pt *wait_pt = NULL;

drivers/gpu/drm/drm_syncobj.c +24 -0
+
+    spin_lock(&syncobj->lock);
+    while(node) {
+        int result;
+
+        wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node);
+        result = point - wait_pt->value;
+        if (result < 0) {
+            node = node->rb_left;
+        } else if (result > 0) {
+            node = node->rb_right;
+        } else {
+            if (fence)
+                *fence = dma_fence_get(&wait_pt->base.base);
+            break;
+        }
+    }
+    spin_unlock(&syncobj->lock);
+
+    return wait_pt;
+}
+
+static struct drm_syncobj_wait_pt *
+drm_syncobj_timeline_create_wait_pt_and_return_fence(struct 

drm_syncobj *syncobj,
+                             u64 point,
+                             struct dma_fence **fence) +{
+    struct drm_syncobj_wait_pt *wait_pt = kzalloc(sizeof(*wait_pt), +                              GFP_KERNEL);
+    struct drm_syncobj_stub_fence *base;
+    struct rb_node **new = &(syncobj->syncobj_timeline.wait_pt_tree.rb_node), *parent = NULL;

drivers/gpu/drm/drm_syncobj.c +13 -0
+
+    if (!wait_pt)
+        return NULL;
+    base = &wait_pt->base;
+    spin_lock_init(&base->lock);
+    dma_fence_init(&base->base,
+               &drm_syncobj_stub_fence_ops,
+               &base->lock,
+               syncobj->syncobj_timeline.timeline_context, point);
+    wait_pt->value = point;
+    wait_pt->signal_pt_fence = NULL;
+
+    /* wait pt must be in an order, so that we can easily lookup 

and signal
+     * it */
+    spin_lock(&syncobj->lock);
+    while(*new) {
+        struct drm_syncobj_wait_pt *this =
+            rb_entry(*new, struct drm_syncobj_wait_pt, node); +        int result = wait_pt->value - this->value;
+
+        parent = *new;
+        if (result < 0)
+            new = &((*new)->rb_left);
+        else if (result > 0)
+            new = &((*new)->rb_right);
+        else
+            goto exist;
+    }
+    if (fence)
+        *fence = dma_fence_get(&wait_pt->base.base);
+    rb_link_node(&wait_pt->node, parent, new);
+    rb_insert_color(&wait_pt->node, &syncobj->syncobj_timeline.wait_pt_tree);
+    spin_unlock(&syncobj->lock);
+    return wait_pt;
+exist:
+    spin_unlock(&syncobj->lock);
+    dma_fence_put(&wait_pt->base.base);
+    wait_pt = drm_syncobj_timeline_lookup_wait_pt_fence(syncobj, point,
+                                fence);
+    return wait_pt;

drivers/gpu/drm/drm_syncobj.c +4 -0
+}
+
+static struct dma_fence *
+drm_syncobj_timeline_point_get(struct drm_syncobj *syncobj, u64 

point, u64 flags)

drivers/gpu/drm/drm_syncobj.c +6 -0
+{
+    struct drm_syncobj_wait_pt *wait_pt;
+    struct dma_fence *fence = NULL;
+
+    /* check if the wait pt exists */
+    wait_pt = drm_syncobj_timeline_lookup_wait_pt_fence(syncobj, 

point,
+                                &fence);
+    if (!fence) {
+        /* This is a new wait pt, so create it */
+        wait_pt = drm_syncobj_timeline_create_wait_pt_and_return_fence(syncobj, point, +                                  &fence);
+        if (!fence)
+            return NULL;
+    }
+    if (wait_pt) {
+        int ret = 0;
+
+        if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) { +            ret = wait_event_interruptible_timeout(syncobj->syncobj_timeline.wq, +                                   wait_pt->value <= syncobj->syncobj_timeline.signal_point,
+                                   msecs_to_jiffies(10000)); /* wait 10s */
+            if (ret <= 0)
+                return NULL;
+        }
+        spin_lock(&syncobj->lock);
+        drm_syncobj_find_signal_pt_for_wait_pt(syncobj, wait_pt); +        spin_unlock(&syncobj->lock);
+        return fence;
+    }
+    return NULL;

drivers/gpu/drm/drm_syncobj.c +2 -0
+}
+

  /**
   * drm_syncobj_find_fence - lookup and reference the fence in a sync object
   * @file_private: drm file private pointer

drivers/gpu/drm/drm_syncobj.c +31 -4
@@ -234,20 +541,46 @@ static int 
 drm_syncobj_assign_null_handle(struct drm_syncobj *syncobj)
    * contains a reference to the fence, which must be released by calling
    * dma_fence_put().
    */
-int drm_syncobj_find_fence(struct drm_file *file_private,
-               u32 handle, u64 point,
-               struct dma_fence **fence)
+int drm_syncobj_search_fence(struct drm_syncobj *syncobj, u64 point,
+                 u64 flags, struct dma_fence **fence)
   {
-    struct drm_syncobj *syncobj = drm_syncobj_find(file_private, 
 handle);
       int ret = 0;
         if (!syncobj)
           return -ENOENT;
   -    *fence = drm_syncobj_fence_get(syncobj);
+    drm_syncobj_timeline_garbage_collection(syncobj);
+    if (syncobj->type == DRM_SYNCOBJ_TYPE_NORMAL) {
+        /*NORMAL syncobj always wait on last pt */
+        u64 tail_pt_value = syncobj->syncobj_timeline.signal_point;
+
+        if (tail_pt_value == 0)
+            tail_pt_value += DRM_SYNCOBJ_NORMAL_POINT;
+        /* NORMAL syncobj doesn't care point value */
+        WARN_ON(point != 0);
+        *fence = drm_syncobj_timeline_point_get(syncobj, 
 tail_pt_value,
+                            flags);
+    } else if (syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+        *fence = drm_syncobj_timeline_point_get(syncobj, point,
+                            flags);
+    } else {
+        DRM_ERROR("Don't support this type syncobj\n");
+        *fence = NULL;
+    }
       if (!*fence) {
           ret = -EINVAL;
       }
+    return ret;
+}
+EXPORT_SYMBOL(drm_syncobj_search_fence);
+int drm_syncobj_find_fence(struct drm_file *file_private,
+               u32 handle, u64 point,
+               struct dma_fence **fence) {
+    struct drm_syncobj *syncobj = drm_syncobj_find(file_private, 
 handle);
+
+    int ret = drm_syncobj_search_fence(syncobj, point,
+                    DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT,
+                    fence);

      drm_syncobj_put(syncobj);
      return ret;
  }

drivers/gpu/drm/drm_syncobj.c +6 -14
@@ -264,7 +597,7 @@ void drm_syncobj_free(struct kref *kref)
       struct drm_syncobj *syncobj = container_of(kref,
                              struct drm_syncobj,
                              refcount);
-    drm_syncobj_replace_fence(syncobj, 0, NULL);
+    drm_syncobj_timeline_fini(syncobj, &syncobj->syncobj_timeline);
       kfree(syncobj);
   }
   EXPORT_SYMBOL(drm_syncobj_free);
@@ -292,8 +625,12 @@ int drm_syncobj_create(struct drm_syncobj 
 **out_syncobj, uint32_t flags,
           return -ENOMEM;
         kref_init(&syncobj->refcount);
-    INIT_LIST_HEAD(&syncobj->cb_list);
       spin_lock_init(&syncobj->lock);
+    if (flags & DRM_SYNCOBJ_CREATE_TYPE_TIMELINE)
+        syncobj->type = DRM_SYNCOBJ_TYPE_TIMELINE;
+    else
+        syncobj->type = DRM_SYNCOBJ_TYPE_NORMAL;
+    drm_syncobj_timeline_init(syncobj, &syncobj->syncobj_timeline);
         if (flags & DRM_SYNCOBJ_CREATE_SIGNALED) {
           ret = drm_syncobj_assign_null_handle(syncobj);
@@ -651,7 +988,6 @@ struct syncobj_wait_entry {
       struct task_struct *task;
       struct dma_fence *fence;
       struct dma_fence_cb fence_cb;
-    struct drm_syncobj_cb syncobj_cb;
   };
     static void syncobj_wait_fence_func(struct dma_fence *fence,
@@ -663,18 +999,6 @@ static void syncobj_wait_fence_func(struct 
 dma_fence *fence,
       wake_up_process(wait->task);
   }
   -static void syncobj_wait_syncobj_func(struct drm_syncobj *syncobj,
-                      struct drm_syncobj_cb *cb)
-{
-    struct syncobj_wait_entry *wait =
-        container_of(cb, struct syncobj_wait_entry, syncobj_cb);
-
-    /* This happens inside the syncobj lock */
-    wait->fence = 
 dma_fence_get(rcu_dereference_protected(syncobj->fence,
- lockdep_is_held(&syncobj->lock)));
-    wake_up_process(wait->task);
-}
-
   static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,

                            uint32_t count,
                            uint32_t flags,

drivers/gpu/drm/drm_syncobj.c +4 -15
@@ -698,14 +1022,11 @@ static signed long 
 drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
       signaled_count = 0;
       for (i = 0; i < count; ++i) {
           entries[i].task = current;
-        entries[i].fence = drm_syncobj_fence_get(syncobjs[i]);
+        ret = drm_syncobj_search_fence(syncobjs[i], 0, 0,
+                           &entries[i].fence);
           if (!entries[i].fence) {
-            if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) {
-                continue;
-            } else {
-                ret = -EINVAL;
-                goto cleanup_entries;
-            }
+            ret = -EINVAL;
+            goto cleanup_entries;
           }
             if (dma_fence_is_signaled(entries[i].fence)) {
@@ -733,15 +1054,6 @@ static signed long 
 drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
        * fallthough and try a 0 timeout wait!
        */
   -    if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) {
-        for (i = 0; i < count; ++i) {
- drm_syncobj_fence_get_or_add_callback(syncobjs[i],
-                                  &entries[i].fence,
-                                  &entries[i].syncobj_cb,
-                                  syncobj_wait_syncobj_func);
-        }
-    }
-
       do {
           set_current_state(TASK_INTERRUPTIBLE);

  @@ -789,13 +1101,10 @@ static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs, cleanup_entries:
      for (i = 0; i < count; ++i) {
-        if (entries[i].syncobj_cb.func)
-            drm_syncobj_remove_callback(syncobjs[i],
-                            &entries[i].syncobj_cb);
+        dma_fence_put(entries[i].fence);
          if (entries[i].fence_cb.func)
              dma_fence_remove_callback(entries[i].fence,
                            &entries[i].fence_cb);
-        dma_fence_put(entries[i].fence);
      }
      kfree(entries);
  @@ -970,12 +1279,21 @@ drm_syncobj_reset_ioctl(struct drm_device *dev, void *data,
      if (ret < 0)
          return ret;
  -    for (i = 0; i < args->count_handles; i++)
-        drm_syncobj_replace_fence(syncobjs[i], 0, NULL);

drivers/gpu/drm/drm_syncobj.c +12 -1
-
+    for (i = 0; i < args->count_handles; i++) {
+        if (syncobjs[i]->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+            DRM_ERROR("timeline syncobj cannot reset!\n");
+            ret = -EINVAL;
+            goto out;
+        }
+        drm_syncobj_timeline_fini(syncobjs[i],
+                      &syncobjs[i]->syncobj_timeline);
+        drm_syncobj_timeline_init(syncobjs[i],
+                      &syncobjs[i]->syncobj_timeline);
+    }
+out:

      drm_syncobj_array_free(syncobjs, args->count_handles); -    return 0;
+    return ret;
  }
    int

drivers/gpu/drm/i915/i915_gem_execbuffer.c +3 -0
index 0a8d2d64f380..579e91a5858b 100644
@@ -2137,7 +2137,9 @@ await_fence_array(struct i915_execbuffer *eb,
           if (!(flags & I915_EXEC_FENCE_WAIT))
               continue;
   -        fence = drm_syncobj_fence_get(syncobj);
+        drm_syncobj_search_fence(syncobj, 0,
+                     DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT,
+                     &fence);
           if (!fence)
               return -EINVAL;

  diff --git a/include/drm/drm_syncobj.h b/include/drm/drm_syncobj.h index 425432b85a87..657c97dc25ec 100644

include/drm/drm_syncobj.h +20 -7
@@ -30,6 +30,25 @@
     struct drm_syncobj_cb;
   +enum drm_syncobj_type {
+    DRM_SYNCOBJ_TYPE_NORMAL,
+    DRM_SYNCOBJ_TYPE_TIMELINE
+};
+
+struct drm_syncobj_timeline {
+    wait_queue_head_t    wq;
+    u64 timeline_context;
+    /**
+     * @timeline: syncobj timeline
+     */
+    u64 timeline;
+    u64 signal_point;
+
+
+    struct rb_root wait_pt_tree;
+    struct list_head signal_pt_list;
+};
+
   /**
    * struct drm_syncobj - sync object.
    *
@@ -41,19 +60,16 @@ struct drm_syncobj {
        */
       struct kref refcount;
       /**
-     * @fence:
-     * NULL or a pointer to the fence bound to this object.
-     *
-     * This field should not be used directly. Use 
 drm_syncobj_fence_get()
-     * and drm_syncobj_replace_fence() instead.
+     * @type: indicate syncobj type
        */
-    struct dma_fence __rcu *fence;
+    enum drm_syncobj_type type;
       /**
-     * @cb_list: List of callbacks to call when the &fence gets 

replaced.
+     * @syncobj_timeline: timeline
       */
-    struct list_head cb_list;
+    struct drm_syncobj_timeline syncobj_timeline; +
      /**
-     * @lock: Protects &cb_list and write-locks &fence. +     * @lock: Protects timeline list and write-locks &fence. */
      spinlock_t lock;
      /**

include/drm/drm_syncobj.h +0 -9
@@ -62,25 +78,6 @@ struct drm_syncobj {
       struct file *file;
   };
   -typedef void (*drm_syncobj_func_t)(struct drm_syncobj *syncobj,
-                   struct drm_syncobj_cb *cb);
-
-/**
- * struct drm_syncobj_cb - callback for drm_syncobj_add_callback
- * @node: used by drm_syncob_add_callback to append this struct to
- *      &drm_syncobj.cb_list
- * @func: drm_syncobj_func_t to call
- *
- * This struct will be initialized by drm_syncobj_add_callback, 

additional
- * data can be passed along by embedding drm_syncobj_cb in another struct.
- * The callback will get called the next time drm_syncobj_replace_fence is
- * called.
- */

include/drm/drm_syncobj.h +0 -5
-struct drm_syncobj_cb {
-    struct list_head node;
-    drm_syncobj_func_t func;
-};
-

  void drm_syncobj_free(struct kref *kref);
    /**

include/drm/drm_syncobj.h +0 -0
@@ -106,29 +103,6 @@ drm_syncobj_put(struct drm_syncobj *obj)
       kref_put(&obj->refcount, drm_syncobj_free);
   }

  -/**
- * drm_syncobj_fence_get - get a reference to a fence in a sync object
- * @syncobj: sync object.
- *
- * This acquires additional reference to &drm_syncobj.fence contained in @obj,
- * if not NULL. It is illegal to call this without already holding a reference.
- * No locks required.
- *
- * Returns:
- * Either the fence of @obj or NULL if there's none.
- */

include/drm/drm_syncobj.h +0 -12
-static inline struct dma_fence *
-drm_syncobj_fence_get(struct drm_syncobj *syncobj)
-{
-    struct dma_fence *fence;
-
-    rcu_read_lock();
-    fence = dma_fence_get_rcu_safe(&syncobj->fence);
-    rcu_read_unlock();
-
-    return fence;
-}
-

  struct drm_syncobj *drm_syncobj_find(struct drm_file *file_private, u32 handle);
  void drm_syncobj_replace_fence(struct drm_syncobj *syncobj, u64 point,

include/drm/drm_syncobj.h +2 -0
@@ -142,5 +116,7 @@ int drm_syncobj_create(struct drm_syncobj 
 **out_syncobj, uint32_t flags,
   int drm_syncobj_get_handle(struct drm_file *file_private,
                  struct drm_syncobj *syncobj, u32 *handle);
   int drm_syncobj_get_fd(struct drm_syncobj *syncobj, int *p_fd);
+int drm_syncobj_search_fence(struct drm_syncobj *syncobj, u64 point,
+                 u64 flags, struct dma_fence **fence);
     #endif
include/uapi/drm/drm.h +0 -1
index 300f336633f2..cebdb2541eb7 100644
--- a/include/uapi/drm/drm.h
Quoted 9 lines

include/uapi/drm/drm.h struct drm_prime_handle { · line 722

Quoted 8 lines

+++ b/include/uapi/drm/drm.h

include/uapi/drm/drm.h +1 -0
@@ -717,6 +717,7 @@ struct drm_prime_handle {
   struct drm_syncobj_create {
       __u32 handle;
   #define DRM_SYNCOBJ_CREATE_SIGNALED (1 << 0)
+#define DRM_SYNCOBJ_CREATE_TYPE_TIMELINE (1 << 1)
       __u32 flags;
   };
Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): e90fbb9b-f201-9887-c722-fa91c300441b-5C7GfCeVMHo@public.gmane.org

#12 lore raw

Am 12.09.2018 um 12:20 schrieb zhoucm1:

Quoted 3 lines

[SNIP]

Quoted 1 line
Quoted 1 line

Drop the term semaphore here, better use syncobj.

This is from VK_KHR_timeline_semaphore extension describe, not my invention, I just quote it. In kernel side, we call syncobj, in UMD, they still call semaphore.

Yeah, but we don't care about close source UMD names in the kernel and the open source UMD calls it syncobj as well.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 6 lines

[SNIP]
+static void drm_syncobj_find_signal_pt_for_wait_pt(struct drm_syncobj *syncobj,
+                           struct drm_syncobj_wait_pt *wait_pt) +{

That whole approach still looks horrible complicated to me.

Quoted 6 lines

It's already very close to what you said before.

Quoted 3 lines
Quoted 3 lines

Especially the separation of signal and wait pt is completely unnecessary as far as I can see. When a wait pt is requested we just need to search for the signal point which it will trigger.

Yeah, I tried this, but when I implement cpu wait ioctl on specific point, we need a advanced wait pt fence, otherwise, we could still need old syncobj cb.

Why? I mean you just need to call drm_syncobj_find_fence() and when that one returns NULL you use wait_event_*() to wait for a signal point >= your wait point to appear and try again.

Regards,
Christian.

Quoted 708 lines

Thanks,
David Zhou

Quoted 699 lines
Quoted 689 lines

Regards,
Christian.

Quoted 685 lines

+    struct drm_syncobj_timeline *timeline =
&syncobj->syncobj_timeline;
+    struct drm_syncobj_signal_pt *signal_pt;
+    int ret;

(quoted diff) +20 -0
+
+    if (wait_pt->signal_pt_fence) {
+        return;
+    } else if ((syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) &&
+           (wait_pt->value <= timeline->timeline)) {
+        dma_fence_signal(&wait_pt->base.base);
+        rb_erase(&wait_pt->node,
+             &timeline->wait_pt_tree);
+        RB_CLEAR_NODE(&wait_pt->node);
+        dma_fence_put(&wait_pt->base.base);
+        return;
+    }
+
+    list_for_each_entry(signal_pt, &timeline->signal_pt_list, list) {
+        if (wait_pt->value < signal_pt->value)
+            continue;
+        if ((syncobj->type == DRM_SYNCOBJ_TYPE_NORMAL) &&
+            (wait_pt->value != signal_pt->value))
+            continue;
+        wait_pt->signal_pt_fence = 

dma_fence_get(&signal_pt->base->base);
+        ret = dma_fence_add_callback(wait_pt->signal_pt_fence, +                         &wait_pt->wait_cb,
+                         wait_pt_func);
+        if (ret == -ENOENT) {
+            dma_fence_signal(&wait_pt->base.base);
+            dma_fence_put(wait_pt->signal_pt_fence);
+            wait_pt->signal_pt_fence = NULL;
+            rb_erase(&wait_pt->node,
+                 &timeline->wait_pt_tree);
+            RB_CLEAR_NODE(&wait_pt->node);
+            dma_fence_put(&wait_pt->base.base);
+        } else if (ret < 0) {
+            dma_fence_put(wait_pt->signal_pt_fence);
+            DRM_ERROR("add callback error!");
+        } else {
+            /* after adding callback, remove from rb tree */ +            rb_erase(&wait_pt->node,
+                 &timeline->wait_pt_tree);
+            RB_CLEAR_NODE(&wait_pt->node);
+        }
+        return;
+    }
+    /* signaled pt was released */
+    if (!wait_pt->signal_pt_fence && (wait_pt->value <= +                      timeline->signal_point)) {
+        dma_fence_signal(&wait_pt->base.base);
+        rb_erase(&wait_pt->node,
+             &timeline->wait_pt_tree);
+        RB_CLEAR_NODE(&wait_pt->node);
+        dma_fence_put(&wait_pt->base.base);
+    }
  }
  -void drm_syncobj_add_callback(struct drm_syncobj *syncobj, -                  struct drm_syncobj_cb *cb,
-                  drm_syncobj_func_t func)
+static int drm_syncobj_timeline_create_signal_pt(struct drm_syncobj *syncobj,
+                         struct dma_fence *fence,
+                         u64 point)
  {
+    struct drm_syncobj_signal_pt *signal_pt =
+        kzalloc(sizeof(struct drm_syncobj_signal_pt), GFP_KERNEL); +    struct drm_syncobj_signal_pt *tail_pt;
+    struct dma_fence **fences;
+    struct rb_node *node;
+    struct drm_syncobj_wait_pt *tail_wait_pt = NULL;
+    int num_fences = 0;
+    int ret = 0, i;

(quoted diff) +19 -0
+
+    if (!signal_pt)
+        return -ENOMEM;
+    if (syncobj->syncobj_timeline.signal_point >= point) {
+        DRM_WARN("A later signal is ready!");
+        goto out;
+    }
+    if (!fence)
+        goto out;
+
+    fences = kmalloc_array(sizeof(void *), 2, GFP_KERNEL);
+    if (!fences)
+        goto out;
+    fences[num_fences++] = dma_fence_get(fence);
+    /* timeline syncobj must take this dependency */
+    if (syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+        spin_lock(&syncobj->lock);
+        if (!list_empty(&syncobj->syncobj_timeline.signal_pt_list)) {
+            tail_pt = 

list_last_entry(&syncobj->syncobj_timeline.signal_pt_list, +                          struct drm_syncobj_signal_pt, list); +            fences[num_fences++] = dma_fence_get(&tail_pt->base->base);
+        }
+        spin_unlock(&syncobj->lock);
+    }
+    signal_pt->base = dma_fence_array_create(num_fences, fences,
+ syncobj->syncobj_timeline.timeline_context,
+                         point, false);
+    if (!signal_pt->base)
+        goto fail;
+
+    signal_pt->value = point; spin_lock(&syncobj->lock);
-    drm_syncobj_add_callback_locked(syncobj, cb, func); +    syncobj->syncobj_timeline.signal_point = point;
+    INIT_LIST_HEAD(&signal_pt->list);
+    list_add_tail(&signal_pt->list, &syncobj->syncobj_timeline.signal_pt_list);
+
+    /* check if there is advanced wait */
+    node = rb_last(&syncobj->syncobj_timeline.wait_pt_tree); +    if (node)
+        tail_wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node);
+    if (tail_wait_pt && !tail_wait_pt->signal_pt_fence) { +        for (node = rb_first(&syncobj->syncobj_timeline.wait_pt_tree);
+             node != NULL; node = rb_next(node)) {
+            struct drm_syncobj_wait_pt *wait_pt =
+                rb_entry(node, struct drm_syncobj_wait_pt, +                     node);

(quoted diff) +6 -0
+
+            drm_syncobj_find_signal_pt_for_wait_pt(syncobj,
+                                   wait_pt);
+        }
+    }
+

      spin_unlock(&syncobj->lock);
+    wake_up_all(&syncobj->syncobj_timeline.wq);

(quoted diff) +9 -0
+
+    return 0;
+fail:
+    for (i = 0; i < num_fences; i++)
+        dma_fence_put(fences[i]);
+    kfree(fences);
+out:
+    kfree(signal_pt);
+    return ret;

  }
  -void drm_syncobj_remove_callback(struct drm_syncobj *syncobj, -                 struct drm_syncobj_cb *cb)
+static void drm_syncobj_timeline_garbage_collection(struct drm_syncobj *syncobj)
  {
+    struct drm_syncobj_timeline *timeline = &syncobj->syncobj_timeline;
+    struct rb_node *node;
+    struct drm_syncobj_wait_pt *wait_pt;
+    struct drm_syncobj_signal_pt *signal_pt, *tmp;
+
      spin_lock(&syncobj->lock);
-    list_del_init(&cb->node);
+    node = rb_first(&timeline->wait_pt_tree);
+    while (node) {
+        wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node); +        if (!dma_fence_is_signaled(&wait_pt->base.base)) { +            node = rb_next(node);
+            continue;
+        }
+        rb_erase(&wait_pt->node,
+             &timeline->wait_pt_tree);
+        RB_CLEAR_NODE(&wait_pt->node);
+        /* kfree(wait_pt) is excuted by fence put */
+        dma_fence_put(&wait_pt->base.base);
+        node = rb_first(&timeline->wait_pt_tree);
+    }
+    list_for_each_entry_safe(signal_pt, tmp,
+                 &timeline->signal_pt_list, list) {
+        if (dma_fence_is_signaled(&signal_pt->base->base)) { +            timeline->timeline = signal_pt->value;
+            list_del(&signal_pt->list);
+ dma_fence_put(&signal_pt->base->base);
+            kfree(signal_pt);
+        } else {
+            /*signal_pt is in order in list, from small to big, so +             * the later must not be signal either */
+            break;
+        }
+    }
+
      spin_unlock(&syncobj->lock);
  }
-
  /**
   * drm_syncobj_replace_fence - replace fence in a sync object.
   * @syncobj: Sync object to replace fence in

(quoted diff) +21 -20
@@ -176,28 +368,29 @@ void drm_syncobj_replace_fence(struct 
 drm_syncobj *syncobj,
                      u64 point,
                      struct dma_fence *fence)
   {
-    struct dma_fence *old_fence;
-    struct drm_syncobj_cb *cur, *tmp;
-
-    if (fence)
-        dma_fence_get(fence);
-
-    spin_lock(&syncobj->lock);
-
-    old_fence = rcu_dereference_protected(syncobj->fence,
- lockdep_is_held(&syncobj->lock));
-    rcu_assign_pointer(syncobj->fence, fence);
-
-    if (fence != old_fence) {
-        list_for_each_entry_safe(cur, tmp, &syncobj->cb_list, node) {
-            list_del_init(&cur->node);
-            cur->func(syncobj, cur);
+    drm_syncobj_timeline_garbage_collection(syncobj);
+    if (syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+        if (fence)
+            drm_syncobj_timeline_create_signal_pt(syncobj, fence,
+                                  point);
+        return;
+    } else if (syncobj->type == DRM_SYNCOBJ_TYPE_NORMAL) {
+        u64 pt_value;
+
+        if (!fence) {
+            drm_syncobj_timeline_fini(syncobj,
+ &syncobj->syncobj_timeline);
+            drm_syncobj_timeline_init(syncobj,
+ &syncobj->syncobj_timeline);
+            return;
           }
+        pt_value = syncobj->syncobj_timeline.signal_point +
+            DRM_SYNCOBJ_NORMAL_POINT;
+        drm_syncobj_timeline_create_signal_pt(syncobj, fence, 
 pt_value);
+        return;
+    } else {
+        DRM_ERROR("the syncobj type isn't support\n");
       }
-
-    spin_unlock(&syncobj->lock);
-
-    dma_fence_put(old_fence);
   }

  EXPORT_SYMBOL(drm_syncobj_replace_fence);
  @@ -220,6 +413,120 @@ static int drm_syncobj_assign_null_handle(struct drm_syncobj *syncobj) return 0;
  }
  +static struct drm_syncobj_wait_pt * +drm_syncobj_timeline_lookup_wait_pt_fence(struct drm_syncobj
*syncobj,
+                      u64 point,
+                      struct dma_fence **fence)
+{
+    struct rb_node *node =
syncobj->syncobj_timeline.wait_pt_tree.rb_node;
+    struct drm_syncobj_wait_pt *wait_pt = NULL;

(quoted diff) +24 -0
+
+    spin_lock(&syncobj->lock);
+    while(node) {
+        int result;
+
+        wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node);
+        result = point - wait_pt->value;
+        if (result < 0) {
+            node = node->rb_left;
+        } else if (result > 0) {
+            node = node->rb_right;
+        } else {
+            if (fence)
+                *fence = dma_fence_get(&wait_pt->base.base);
+            break;
+        }
+    }
+    spin_unlock(&syncobj->lock);
+
+    return wait_pt;
+}
+
+static struct drm_syncobj_wait_pt *
+drm_syncobj_timeline_create_wait_pt_and_return_fence(struct 

drm_syncobj *syncobj,
+                             u64 point,
+                             struct dma_fence **fence) +{
+    struct drm_syncobj_wait_pt *wait_pt = kzalloc(sizeof(*wait_pt), +                              GFP_KERNEL);
+    struct drm_syncobj_stub_fence *base;
+    struct rb_node **new = &(syncobj->syncobj_timeline.wait_pt_tree.rb_node), *parent = NULL;

(quoted diff) +13 -0
+
+    if (!wait_pt)
+        return NULL;
+    base = &wait_pt->base;
+    spin_lock_init(&base->lock);
+    dma_fence_init(&base->base,
+               &drm_syncobj_stub_fence_ops,
+               &base->lock,
+ syncobj->syncobj_timeline.timeline_context, point);
+    wait_pt->value = point;
+    wait_pt->signal_pt_fence = NULL;
+
+    /* wait pt must be in an order, so that we can easily lookup 

and signal
+     * it */
+    spin_lock(&syncobj->lock);
+    while(*new) {
+        struct drm_syncobj_wait_pt *this =
+            rb_entry(*new, struct drm_syncobj_wait_pt, node); +        int result = wait_pt->value - this->value;
+
+        parent = *new;
+        if (result < 0)
+            new = &((*new)->rb_left);
+        else if (result > 0)
+            new = &((*new)->rb_right);
+        else
+            goto exist;
+    }
+    if (fence)
+        *fence = dma_fence_get(&wait_pt->base.base);
+    rb_link_node(&wait_pt->node, parent, new);
+    rb_insert_color(&wait_pt->node, &syncobj->syncobj_timeline.wait_pt_tree);
+    spin_unlock(&syncobj->lock);
+    return wait_pt;
+exist:
+    spin_unlock(&syncobj->lock);
+    dma_fence_put(&wait_pt->base.base);
+    wait_pt = drm_syncobj_timeline_lookup_wait_pt_fence(syncobj, point,
+                                fence);
+    return wait_pt;

(quoted diff) +4 -0
+}
+
+static struct dma_fence *
+drm_syncobj_timeline_point_get(struct drm_syncobj *syncobj, u64 

point, u64 flags)

(quoted diff) +6 -0
+{
+    struct drm_syncobj_wait_pt *wait_pt;
+    struct dma_fence *fence = NULL;
+
+    /* check if the wait pt exists */
+    wait_pt = drm_syncobj_timeline_lookup_wait_pt_fence(syncobj, 

point,
+                                &fence);
+    if (!fence) {
+        /* This is a new wait pt, so create it */
+        wait_pt = drm_syncobj_timeline_create_wait_pt_and_return_fence(syncobj, point, +                                  &fence);
+        if (!fence)
+            return NULL;
+    }
+    if (wait_pt) {
+        int ret = 0;
+
+        if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) { +            ret = wait_event_interruptible_timeout(syncobj->syncobj_timeline.wq, +                                   wait_pt->value <= syncobj->syncobj_timeline.signal_point,
+ msecs_to_jiffies(10000)); /* wait 10s */
+            if (ret <= 0)
+                return NULL;
+        }
+        spin_lock(&syncobj->lock);
+        drm_syncobj_find_signal_pt_for_wait_pt(syncobj, wait_pt); +        spin_unlock(&syncobj->lock);
+        return fence;
+    }
+    return NULL;

(quoted diff) +2 -0
+}
+

  /**
   * drm_syncobj_find_fence - lookup and reference the fence in a sync object
   * @file_private: drm file private pointer

(quoted diff) +31 -4
@@ -234,20 +541,46 @@ static int 
 drm_syncobj_assign_null_handle(struct drm_syncobj *syncobj)
    * contains a reference to the fence, which must be released by calling
    * dma_fence_put().
    */
-int drm_syncobj_find_fence(struct drm_file *file_private,
-               u32 handle, u64 point,
-               struct dma_fence **fence)
+int drm_syncobj_search_fence(struct drm_syncobj *syncobj, u64 point,
+                 u64 flags, struct dma_fence **fence)
   {
-    struct drm_syncobj *syncobj = drm_syncobj_find(file_private, 
 handle);
       int ret = 0;
         if (!syncobj)
           return -ENOENT;
   -    *fence = drm_syncobj_fence_get(syncobj);
+    drm_syncobj_timeline_garbage_collection(syncobj);
+    if (syncobj->type == DRM_SYNCOBJ_TYPE_NORMAL) {
+        /*NORMAL syncobj always wait on last pt */
+        u64 tail_pt_value = syncobj->syncobj_timeline.signal_point;
+
+        if (tail_pt_value == 0)
+            tail_pt_value += DRM_SYNCOBJ_NORMAL_POINT;
+        /* NORMAL syncobj doesn't care point value */
+        WARN_ON(point != 0);
+        *fence = drm_syncobj_timeline_point_get(syncobj, 
 tail_pt_value,
+                            flags);
+    } else if (syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+        *fence = drm_syncobj_timeline_point_get(syncobj, point,
+                            flags);
+    } else {
+        DRM_ERROR("Don't support this type syncobj\n");
+        *fence = NULL;
+    }
       if (!*fence) {
           ret = -EINVAL;
       }
+    return ret;
+}
+EXPORT_SYMBOL(drm_syncobj_search_fence);
+int drm_syncobj_find_fence(struct drm_file *file_private,
+               u32 handle, u64 point,
+               struct dma_fence **fence) {
+    struct drm_syncobj *syncobj = drm_syncobj_find(file_private, 
 handle);
+
+    int ret = drm_syncobj_search_fence(syncobj, point,
+                    DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT,
+                    fence);

      drm_syncobj_put(syncobj);
      return ret;
  }

(quoted diff) +6 -14
@@ -264,7 +597,7 @@ void drm_syncobj_free(struct kref *kref)
       struct drm_syncobj *syncobj = container_of(kref,
                              struct drm_syncobj,
                              refcount);
-    drm_syncobj_replace_fence(syncobj, 0, NULL);
+    drm_syncobj_timeline_fini(syncobj, &syncobj->syncobj_timeline);
       kfree(syncobj);
   }
   EXPORT_SYMBOL(drm_syncobj_free);
@@ -292,8 +625,12 @@ int drm_syncobj_create(struct drm_syncobj 
 **out_syncobj, uint32_t flags,
           return -ENOMEM;
         kref_init(&syncobj->refcount);
-    INIT_LIST_HEAD(&syncobj->cb_list);
       spin_lock_init(&syncobj->lock);
+    if (flags & DRM_SYNCOBJ_CREATE_TYPE_TIMELINE)
+        syncobj->type = DRM_SYNCOBJ_TYPE_TIMELINE;
+    else
+        syncobj->type = DRM_SYNCOBJ_TYPE_NORMAL;
+    drm_syncobj_timeline_init(syncobj, &syncobj->syncobj_timeline);
         if (flags & DRM_SYNCOBJ_CREATE_SIGNALED) {
           ret = drm_syncobj_assign_null_handle(syncobj);
@@ -651,7 +988,6 @@ struct syncobj_wait_entry {
       struct task_struct *task;
       struct dma_fence *fence;
       struct dma_fence_cb fence_cb;
-    struct drm_syncobj_cb syncobj_cb;
   };
     static void syncobj_wait_fence_func(struct dma_fence *fence,
@@ -663,18 +999,6 @@ static void syncobj_wait_fence_func(struct 
 dma_fence *fence,
       wake_up_process(wait->task);
   }
   -static void syncobj_wait_syncobj_func(struct drm_syncobj *syncobj,
-                      struct drm_syncobj_cb *cb)
-{
-    struct syncobj_wait_entry *wait =
-        container_of(cb, struct syncobj_wait_entry, syncobj_cb);
-
-    /* This happens inside the syncobj lock */
-    wait->fence = 
 dma_fence_get(rcu_dereference_protected(syncobj->fence,
- lockdep_is_held(&syncobj->lock)));
-    wake_up_process(wait->task);
-}
-
   static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,

                            uint32_t count,
                            uint32_t flags,

(quoted diff) +4 -15
@@ -698,14 +1022,11 @@ static signed long 
 drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
       signaled_count = 0;
       for (i = 0; i < count; ++i) {
           entries[i].task = current;
-        entries[i].fence = drm_syncobj_fence_get(syncobjs[i]);
+        ret = drm_syncobj_search_fence(syncobjs[i], 0, 0,
+                           &entries[i].fence);
           if (!entries[i].fence) {
-            if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) {
-                continue;
-            } else {
-                ret = -EINVAL;
-                goto cleanup_entries;
-            }
+            ret = -EINVAL;
+            goto cleanup_entries;
           }
             if (dma_fence_is_signaled(entries[i].fence)) {
@@ -733,15 +1054,6 @@ static signed long 
 drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
        * fallthough and try a 0 timeout wait!
        */
   -    if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) {
-        for (i = 0; i < count; ++i) {
- drm_syncobj_fence_get_or_add_callback(syncobjs[i],
-                                  &entries[i].fence,
- &entries[i].syncobj_cb,
- syncobj_wait_syncobj_func);
-        }
-    }
-
       do {
           set_current_state(TASK_INTERRUPTIBLE);

  @@ -789,13 +1101,10 @@ static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs, cleanup_entries:
      for (i = 0; i < count; ++i) {
-        if (entries[i].syncobj_cb.func)
-            drm_syncobj_remove_callback(syncobjs[i],
-                            &entries[i].syncobj_cb);
+        dma_fence_put(entries[i].fence);
          if (entries[i].fence_cb.func)
              dma_fence_remove_callback(entries[i].fence,
                            &entries[i].fence_cb);
-        dma_fence_put(entries[i].fence);
      }
      kfree(entries);
  @@ -970,12 +1279,21 @@ drm_syncobj_reset_ioctl(struct drm_device *dev, void *data,
      if (ret < 0)
          return ret;
  -    for (i = 0; i < args->count_handles; i++)
-        drm_syncobj_replace_fence(syncobjs[i], 0, NULL);

(quoted diff) +12 -1
-
+    for (i = 0; i < args->count_handles; i++) {
+        if (syncobjs[i]->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+            DRM_ERROR("timeline syncobj cannot reset!\n");
+            ret = -EINVAL;
+            goto out;
+        }
+        drm_syncobj_timeline_fini(syncobjs[i],
+ &syncobjs[i]->syncobj_timeline);
+        drm_syncobj_timeline_init(syncobjs[i],
+ &syncobjs[i]->syncobj_timeline);
+    }
+out:

      drm_syncobj_array_free(syncobjs, args->count_handles); -    return 0;
+    return ret;
  }
    int

drivers/gpu/drm/i915/i915_gem_execbuffer.c +3 -0
index 0a8d2d64f380..579e91a5858b 100644
@@ -2137,7 +2137,9 @@ await_fence_array(struct i915_execbuffer *eb,
           if (!(flags & I915_EXEC_FENCE_WAIT))
               continue;
   -        fence = drm_syncobj_fence_get(syncobj);
+        drm_syncobj_search_fence(syncobj, 0,
+ DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT,
+                     &fence);
           if (!fence)
               return -EINVAL;

  diff --git a/include/drm/drm_syncobj.h b/include/drm/drm_syncobj.h index 425432b85a87..657c97dc25ec 100644

include/drm/drm_syncobj.h +20 -7
@@ -30,6 +30,25 @@
     struct drm_syncobj_cb;
   +enum drm_syncobj_type {
+    DRM_SYNCOBJ_TYPE_NORMAL,
+    DRM_SYNCOBJ_TYPE_TIMELINE
+};
+
+struct drm_syncobj_timeline {
+    wait_queue_head_t    wq;
+    u64 timeline_context;
+    /**
+     * @timeline: syncobj timeline
+     */
+    u64 timeline;
+    u64 signal_point;
+
+
+    struct rb_root wait_pt_tree;
+    struct list_head signal_pt_list;
+};
+
   /**
    * struct drm_syncobj - sync object.
    *
@@ -41,19 +60,16 @@ struct drm_syncobj {
        */
       struct kref refcount;
       /**
-     * @fence:
-     * NULL or a pointer to the fence bound to this object.
-     *
-     * This field should not be used directly. Use 
 drm_syncobj_fence_get()
-     * and drm_syncobj_replace_fence() instead.
+     * @type: indicate syncobj type
        */
-    struct dma_fence __rcu *fence;
+    enum drm_syncobj_type type;
       /**
-     * @cb_list: List of callbacks to call when the &fence gets 

replaced.
+     * @syncobj_timeline: timeline
       */
-    struct list_head cb_list;
+    struct drm_syncobj_timeline syncobj_timeline; +
      /**
-     * @lock: Protects &cb_list and write-locks &fence. +     * @lock: Protects timeline list and write-locks &fence. */
      spinlock_t lock;
      /**

include/drm/drm_syncobj.h +0 -9
@@ -62,25 +78,6 @@ struct drm_syncobj {
       struct file *file;
   };
   -typedef void (*drm_syncobj_func_t)(struct drm_syncobj *syncobj,
-                   struct drm_syncobj_cb *cb);
-
-/**
- * struct drm_syncobj_cb - callback for drm_syncobj_add_callback
- * @node: used by drm_syncob_add_callback to append this struct to
- *      &drm_syncobj.cb_list
- * @func: drm_syncobj_func_t to call
- *
- * This struct will be initialized by drm_syncobj_add_callback, 

additional
- * data can be passed along by embedding drm_syncobj_cb in another struct.
- * The callback will get called the next time drm_syncobj_replace_fence is
- * called.
- */

include/drm/drm_syncobj.h +0 -5
-struct drm_syncobj_cb {
-    struct list_head node;
-    drm_syncobj_func_t func;
-};
-

  void drm_syncobj_free(struct kref *kref);
    /**

include/drm/drm_syncobj.h +0 -0
@@ -106,29 +103,6 @@ drm_syncobj_put(struct drm_syncobj *obj)
       kref_put(&obj->refcount, drm_syncobj_free);
   }

  -/**
- * drm_syncobj_fence_get - get a reference to a fence in a sync object
- * @syncobj: sync object.
- *
- * This acquires additional reference to &drm_syncobj.fence contained in @obj,
- * if not NULL. It is illegal to call this without already holding a reference.
- * No locks required.
- *
- * Returns:
- * Either the fence of @obj or NULL if there's none.
- */

include/drm/drm_syncobj.h +0 -12
-static inline struct dma_fence *
-drm_syncobj_fence_get(struct drm_syncobj *syncobj)
-{
-    struct dma_fence *fence;
-
-    rcu_read_lock();
-    fence = dma_fence_get_rcu_safe(&syncobj->fence);
-    rcu_read_unlock();
-
-    return fence;
-}
-

  struct drm_syncobj *drm_syncobj_find(struct drm_file *file_private, u32 handle);
  void drm_syncobj_replace_fence(struct drm_syncobj *syncobj, u64 point,

include/drm/drm_syncobj.h +2 -0
@@ -142,5 +116,7 @@ int drm_syncobj_create(struct drm_syncobj 
 **out_syncobj, uint32_t flags,
   int drm_syncobj_get_handle(struct drm_file *file_private,
                  struct drm_syncobj *syncobj, u32 *handle);
   int drm_syncobj_get_fd(struct drm_syncobj *syncobj, int *p_fd);
+int drm_syncobj_search_fence(struct drm_syncobj *syncobj, u64 point,
+                 u64 flags, struct dma_fence **fence);
     #endif
include/uapi/drm/drm.h +0 -1
index 300f336633f2..cebdb2541eb7 100644
--- a/include/uapi/drm/drm.h
Quoted 9 lines

include/uapi/drm/drm.h struct drm_prime_handle { · line 722

Quoted 8 lines

+++ b/include/uapi/drm/drm.h

include/uapi/drm/drm.h +1 -0
@@ -717,6 +717,7 @@ struct drm_prime_handle {
   struct drm_syncobj_create {
       __u32 handle;
   #define DRM_SYNCOBJ_CREATE_SIGNALED (1 << 0)
+#define DRM_SYNCOBJ_CREATE_TYPE_TIMELINE (1 << 1)
       __u32 flags;
   };
Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 7117b33b-c24f-6cb5-5372-143385f7a7b5-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org

#14 lore raw
Quoted 16 lines · Christian König

Am 12.09.2018 um 12:20 schrieb zhoucm1:

Quoted 3 lines

[SNIP]

Quoted 1 line
Quoted 1 line

Drop the term semaphore here, better use syncobj.

This is from VK_KHR_timeline_semaphore extension describe, not my invention, I just quote it. In kernel side, we call syncobj, in UMD, they still call semaphore.

Yeah, but we don't care about close source UMD names in the kernel and' the open source UMD calls it syncobj as well.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 6 lines

[SNIP]
+static void drm_syncobj_find_signal_pt_for_wait_pt(struct drm_syncobj *syncobj,
+                           struct drm_syncobj_wait_pt *wait_pt) +{

That whole approach still looks horrible complicated to me.

Quoted 8 lines
Quoted 6 lines

It's already very close to what you said before.

Quoted 3 lines
Quoted 3 lines

Especially the separation of signal and wait pt is completely unnecessary as far as I can see. When a wait pt is requested we just need to search for the signal point which it will trigger.

Yeah, I tried this, but when I implement cpu wait ioctl on specific point, we need a advanced wait pt fence, otherwise, we could still need old syncobj cb.

Why? I mean you just need to call drm_syncobj_find_fence() and when that one returns NULL you use wait_event_*() to wait for a signal point >= your wait point to appear and try again.

e.g. when there are 3 syncobjs(A,B,C) to wait, all syncobjABC have no fence yet, as you said, during drm_syncobj_find_fence(A) is working on wait_event, syncobjB and syncobjC could already be signaled, then we don't know which one is first signaled, which is need when wait ioctl returns.

Back to my implementation, it already fixes all your concerns before, and can be able to easily used in wait_ioctl. When you feel that is complicated, I guess that is because we merged all logic to that and much clean up in one patch. In fact, it already is very simple, timeline_init/fini, create signal/wait_pt, find signal_pt for wait_pt, garbage collection, just them.

Thanks,
David Zhou

Quoted 720 lines

Regards,
Christian.

Quoted 715 lines

Thanks,
David Zhou

Quoted 702 lines
Quoted 692 lines

Regards,
Christian.

Quoted 688 lines

+    struct drm_syncobj_timeline *timeline =
&syncobj->syncobj_timeline;
+    struct drm_syncobj_signal_pt *signal_pt;
+    int ret;

(quoted diff) +14 -0
+
+    if (wait_pt->signal_pt_fence) {
+        return;
+    } else if ((syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) &&
+           (wait_pt->value <= timeline->timeline)) {
+        dma_fence_signal(&wait_pt->base.base);
+        rb_erase(&wait_pt->node,
+             &timeline->wait_pt_tree);
+        RB_CLEAR_NODE(&wait_pt->node);
+        dma_fence_put(&wait_pt->base.base);
+        return;
+    }
+
+    list_for_each_entry(signal_pt, &timeline->signal_pt_list, 

list) {
+        if (wait_pt->value < signal_pt->value)
+            continue;
+        if ((syncobj->type == DRM_SYNCOBJ_TYPE_NORMAL) && +            (wait_pt->value != signal_pt->value))
+            continue;
+        wait_pt->signal_pt_fence = dma_fence_get(&signal_pt->base->base);
+        ret = dma_fence_add_callback(wait_pt->signal_pt_fence, +                         &wait_pt->wait_cb,
+                         wait_pt_func);
+        if (ret == -ENOENT) {
+            dma_fence_signal(&wait_pt->base.base);
+            dma_fence_put(wait_pt->signal_pt_fence);
+            wait_pt->signal_pt_fence = NULL;
+            rb_erase(&wait_pt->node,
+                 &timeline->wait_pt_tree);
+            RB_CLEAR_NODE(&wait_pt->node);
+            dma_fence_put(&wait_pt->base.base);
+        } else if (ret < 0) {
+            dma_fence_put(wait_pt->signal_pt_fence);
+            DRM_ERROR("add callback error!");
+        } else {
+            /* after adding callback, remove from rb tree */ +            rb_erase(&wait_pt->node,
+                 &timeline->wait_pt_tree);
+            RB_CLEAR_NODE(&wait_pt->node);
+        }
+        return;
+    }
+    /* signaled pt was released */
+    if (!wait_pt->signal_pt_fence && (wait_pt->value <= +                      timeline->signal_point)) {
+        dma_fence_signal(&wait_pt->base.base);
+        rb_erase(&wait_pt->node,
+             &timeline->wait_pt_tree);
+        RB_CLEAR_NODE(&wait_pt->node);
+        dma_fence_put(&wait_pt->base.base);
+    }
  }
  -void drm_syncobj_add_callback(struct drm_syncobj *syncobj, -                  struct drm_syncobj_cb *cb,
-                  drm_syncobj_func_t func)
+static int drm_syncobj_timeline_create_signal_pt(struct drm_syncobj *syncobj,
+                         struct dma_fence *fence,
+                         u64 point)
  {
+    struct drm_syncobj_signal_pt *signal_pt =
+        kzalloc(sizeof(struct drm_syncobj_signal_pt), GFP_KERNEL); +    struct drm_syncobj_signal_pt *tail_pt;
+    struct dma_fence **fences;
+    struct rb_node *node;
+    struct drm_syncobj_wait_pt *tail_wait_pt = NULL;
+    int num_fences = 0;
+    int ret = 0, i;

(quoted diff) +18 -0
+
+    if (!signal_pt)
+        return -ENOMEM;
+    if (syncobj->syncobj_timeline.signal_point >= point) {
+        DRM_WARN("A later signal is ready!");
+        goto out;
+    }
+    if (!fence)
+        goto out;
+
+    fences = kmalloc_array(sizeof(void *), 2, GFP_KERNEL);
+    if (!fences)
+        goto out;
+    fences[num_fences++] = dma_fence_get(fence);
+    /* timeline syncobj must take this dependency */
+    if (syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+        spin_lock(&syncobj->lock);
+        if 

(!list_empty(&syncobj->syncobj_timeline.signal_pt_list)) { +            tail_pt = list_last_entry(&syncobj->syncobj_timeline.signal_pt_list, +                          struct drm_syncobj_signal_pt, list); +            fences[num_fences++] = dma_fence_get(&tail_pt->base->base);
+        }
+        spin_unlock(&syncobj->lock);
+    }
+    signal_pt->base = dma_fence_array_create(num_fences, fences,
+ syncobj->syncobj_timeline.timeline_context,
+                         point, false);
+    if (!signal_pt->base)
+        goto fail;
+
+    signal_pt->value = point; spin_lock(&syncobj->lock);
-    drm_syncobj_add_callback_locked(syncobj, cb, func); +    syncobj->syncobj_timeline.signal_point = point;
+    INIT_LIST_HEAD(&signal_pt->list);
+    list_add_tail(&signal_pt->list, &syncobj->syncobj_timeline.signal_pt_list);
+
+    /* check if there is advanced wait */
+    node = rb_last(&syncobj->syncobj_timeline.wait_pt_tree); +    if (node)
+        tail_wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node);
+    if (tail_wait_pt && !tail_wait_pt->signal_pt_fence) { +        for (node = rb_first(&syncobj->syncobj_timeline.wait_pt_tree);
+             node != NULL; node = rb_next(node)) {
+            struct drm_syncobj_wait_pt *wait_pt =
+                rb_entry(node, struct drm_syncobj_wait_pt, +                     node);

(quoted diff) +6 -0
+
+ drm_syncobj_find_signal_pt_for_wait_pt(syncobj,
+                                   wait_pt);
+        }
+    }
+

      spin_unlock(&syncobj->lock);
+    wake_up_all(&syncobj->syncobj_timeline.wq);

(quoted diff) +9 -0
+
+    return 0;
+fail:
+    for (i = 0; i < num_fences; i++)
+        dma_fence_put(fences[i]);
+    kfree(fences);
+out:
+    kfree(signal_pt);
+    return ret;

  }
  -void drm_syncobj_remove_callback(struct drm_syncobj *syncobj, -                 struct drm_syncobj_cb *cb)
+static void drm_syncobj_timeline_garbage_collection(struct drm_syncobj *syncobj)
  {
+    struct drm_syncobj_timeline *timeline = &syncobj->syncobj_timeline;
+    struct rb_node *node;
+    struct drm_syncobj_wait_pt *wait_pt;
+    struct drm_syncobj_signal_pt *signal_pt, *tmp;
+
      spin_lock(&syncobj->lock);
-    list_del_init(&cb->node);
+    node = rb_first(&timeline->wait_pt_tree);
+    while (node) {
+        wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node); +        if (!dma_fence_is_signaled(&wait_pt->base.base)) { +            node = rb_next(node);
+            continue;
+        }
+        rb_erase(&wait_pt->node,
+             &timeline->wait_pt_tree);
+        RB_CLEAR_NODE(&wait_pt->node);
+        /* kfree(wait_pt) is excuted by fence put */
+        dma_fence_put(&wait_pt->base.base);
+        node = rb_first(&timeline->wait_pt_tree);
+    }
+    list_for_each_entry_safe(signal_pt, tmp,
+                 &timeline->signal_pt_list, list) {
+        if (dma_fence_is_signaled(&signal_pt->base->base)) { +            timeline->timeline = signal_pt->value;
+            list_del(&signal_pt->list);
+ dma_fence_put(&signal_pt->base->base);
+            kfree(signal_pt);
+        } else {
+            /*signal_pt is in order in list, from small to big, so +             * the later must not be signal either */
+            break;
+        }
+    }
+
      spin_unlock(&syncobj->lock);
  }
-
  /**
   * drm_syncobj_replace_fence - replace fence in a sync object.
   * @syncobj: Sync object to replace fence in

(quoted diff) +21 -20
@@ -176,28 +368,29 @@ void drm_syncobj_replace_fence(struct 
 drm_syncobj *syncobj,
                      u64 point,
                      struct dma_fence *fence)
   {
-    struct dma_fence *old_fence;
-    struct drm_syncobj_cb *cur, *tmp;
-
-    if (fence)
-        dma_fence_get(fence);
-
-    spin_lock(&syncobj->lock);
-
-    old_fence = rcu_dereference_protected(syncobj->fence,
- lockdep_is_held(&syncobj->lock));
-    rcu_assign_pointer(syncobj->fence, fence);
-
-    if (fence != old_fence) {
-        list_for_each_entry_safe(cur, tmp, &syncobj->cb_list, 
 node) {
-            list_del_init(&cur->node);
-            cur->func(syncobj, cur);
+    drm_syncobj_timeline_garbage_collection(syncobj);
+    if (syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+        if (fence)
+ drm_syncobj_timeline_create_signal_pt(syncobj, fence,
+                                  point);
+        return;
+    } else if (syncobj->type == DRM_SYNCOBJ_TYPE_NORMAL) {
+        u64 pt_value;
+
+        if (!fence) {
+            drm_syncobj_timeline_fini(syncobj,
+ &syncobj->syncobj_timeline);
+            drm_syncobj_timeline_init(syncobj,
+ &syncobj->syncobj_timeline);
+            return;
           }
+        pt_value = syncobj->syncobj_timeline.signal_point +
+            DRM_SYNCOBJ_NORMAL_POINT;
+        drm_syncobj_timeline_create_signal_pt(syncobj, fence, 
 pt_value);
+        return;
+    } else {
+        DRM_ERROR("the syncobj type isn't support\n");
       }
-
-    spin_unlock(&syncobj->lock);
-
-    dma_fence_put(old_fence);

  }
  EXPORT_SYMBOL(drm_syncobj_replace_fence);
  @@ -220,6 +413,120 @@ static int drm_syncobj_assign_null_handle(struct drm_syncobj *syncobj) return 0;
  }
  +static struct drm_syncobj_wait_pt * +drm_syncobj_timeline_lookup_wait_pt_fence(struct drm_syncobj
*syncobj,
+                      u64 point,
+                      struct dma_fence **fence)
+{
+    struct rb_node *node =
syncobj->syncobj_timeline.wait_pt_tree.rb_node;
+    struct drm_syncobj_wait_pt *wait_pt = NULL;

(quoted diff) +24 -0
+
+    spin_lock(&syncobj->lock);
+    while(node) {
+        int result;
+
+        wait_pt = rb_entry(node, struct drm_syncobj_wait_pt, node);
+        result = point - wait_pt->value;
+        if (result < 0) {
+            node = node->rb_left;
+        } else if (result > 0) {
+            node = node->rb_right;
+        } else {
+            if (fence)
+                *fence = dma_fence_get(&wait_pt->base.base);
+            break;
+        }
+    }
+    spin_unlock(&syncobj->lock);
+
+    return wait_pt;
+}
+
+static struct drm_syncobj_wait_pt *
+drm_syncobj_timeline_create_wait_pt_and_return_fence(struct 

drm_syncobj *syncobj,
+                             u64 point,
+                             struct dma_fence **fence) +{
+    struct drm_syncobj_wait_pt *wait_pt = kzalloc(sizeof(*wait_pt), +                              GFP_KERNEL);
+    struct drm_syncobj_stub_fence *base;
+    struct rb_node **new = &(syncobj->syncobj_timeline.wait_pt_tree.rb_node), *parent = NULL;

(quoted diff) +13 -0
+
+    if (!wait_pt)
+        return NULL;
+    base = &wait_pt->base;
+    spin_lock_init(&base->lock);
+    dma_fence_init(&base->base,
+               &drm_syncobj_stub_fence_ops,
+               &base->lock,
+ syncobj->syncobj_timeline.timeline_context, point);
+    wait_pt->value = point;
+    wait_pt->signal_pt_fence = NULL;
+
+    /* wait pt must be in an order, so that we can easily lookup 

and signal
+     * it */
+    spin_lock(&syncobj->lock);
+    while(*new) {
+        struct drm_syncobj_wait_pt *this =
+            rb_entry(*new, struct drm_syncobj_wait_pt, node); +        int result = wait_pt->value - this->value;
+
+        parent = *new;
+        if (result < 0)
+            new = &((*new)->rb_left);
+        else if (result > 0)
+            new = &((*new)->rb_right);
+        else
+            goto exist;
+    }
+    if (fence)
+        *fence = dma_fence_get(&wait_pt->base.base);
+    rb_link_node(&wait_pt->node, parent, new);
+    rb_insert_color(&wait_pt->node, &syncobj->syncobj_timeline.wait_pt_tree);
+    spin_unlock(&syncobj->lock);
+    return wait_pt;
+exist:
+    spin_unlock(&syncobj->lock);
+    dma_fence_put(&wait_pt->base.base);
+    wait_pt = drm_syncobj_timeline_lookup_wait_pt_fence(syncobj, point,
+                                fence);
+    return wait_pt;

(quoted diff) +4 -0
+}
+
+static struct dma_fence *
+drm_syncobj_timeline_point_get(struct drm_syncobj *syncobj, u64 

point, u64 flags)

(quoted diff) +6 -0
+{
+    struct drm_syncobj_wait_pt *wait_pt;
+    struct dma_fence *fence = NULL;
+
+    /* check if the wait pt exists */
+    wait_pt = drm_syncobj_timeline_lookup_wait_pt_fence(syncobj, 

point,
+                                &fence);
+    if (!fence) {
+        /* This is a new wait pt, so create it */
+        wait_pt = drm_syncobj_timeline_create_wait_pt_and_return_fence(syncobj, point, +                                  &fence);
+        if (!fence)
+            return NULL;
+    }
+    if (wait_pt) {
+        int ret = 0;
+
+        if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) { +            ret = wait_event_interruptible_timeout(syncobj->syncobj_timeline.wq, +                                   wait_pt->value <= syncobj->syncobj_timeline.signal_point,
+ msecs_to_jiffies(10000)); /* wait 10s */
+            if (ret <= 0)
+                return NULL;
+        }
+        spin_lock(&syncobj->lock);
+        drm_syncobj_find_signal_pt_for_wait_pt(syncobj, wait_pt); +        spin_unlock(&syncobj->lock);
+        return fence;
+    }
+    return NULL;

(quoted diff) +2 -0
+}
+

  /**
   * drm_syncobj_find_fence - lookup and reference the fence in a sync object
   * @file_private: drm file private pointer

(quoted diff) +31 -4
@@ -234,20 +541,46 @@ static int 
 drm_syncobj_assign_null_handle(struct drm_syncobj *syncobj)
    * contains a reference to the fence, which must be released by calling
    * dma_fence_put().
    */
-int drm_syncobj_find_fence(struct drm_file *file_private,
-               u32 handle, u64 point,
-               struct dma_fence **fence)
+int drm_syncobj_search_fence(struct drm_syncobj *syncobj, u64 point,
+                 u64 flags, struct dma_fence **fence)
   {
-    struct drm_syncobj *syncobj = drm_syncobj_find(file_private, 
 handle);
       int ret = 0;
         if (!syncobj)
           return -ENOENT;
   -    *fence = drm_syncobj_fence_get(syncobj);
+    drm_syncobj_timeline_garbage_collection(syncobj);
+    if (syncobj->type == DRM_SYNCOBJ_TYPE_NORMAL) {
+        /*NORMAL syncobj always wait on last pt */
+        u64 tail_pt_value = syncobj->syncobj_timeline.signal_point;
+
+        if (tail_pt_value == 0)
+            tail_pt_value += DRM_SYNCOBJ_NORMAL_POINT;
+        /* NORMAL syncobj doesn't care point value */
+        WARN_ON(point != 0);
+        *fence = drm_syncobj_timeline_point_get(syncobj, 
 tail_pt_value,
+                            flags);
+    } else if (syncobj->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+        *fence = drm_syncobj_timeline_point_get(syncobj, point,
+                            flags);
+    } else {
+        DRM_ERROR("Don't support this type syncobj\n");
+        *fence = NULL;
+    }
       if (!*fence) {
           ret = -EINVAL;
       }
+    return ret;
+}
+EXPORT_SYMBOL(drm_syncobj_search_fence);
+int drm_syncobj_find_fence(struct drm_file *file_private,
+               u32 handle, u64 point,
+               struct dma_fence **fence) {
+    struct drm_syncobj *syncobj = drm_syncobj_find(file_private, 
 handle);
+
+    int ret = drm_syncobj_search_fence(syncobj, point,
+ DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT,
+                    fence);

      drm_syncobj_put(syncobj);
      return ret;
  }

(quoted diff) +6 -14
@@ -264,7 +597,7 @@ void drm_syncobj_free(struct kref *kref)
       struct drm_syncobj *syncobj = container_of(kref,
                              struct drm_syncobj,
                              refcount);
-    drm_syncobj_replace_fence(syncobj, 0, NULL);
+    drm_syncobj_timeline_fini(syncobj, &syncobj->syncobj_timeline);
       kfree(syncobj);
   }
   EXPORT_SYMBOL(drm_syncobj_free);
@@ -292,8 +625,12 @@ int drm_syncobj_create(struct drm_syncobj 
 **out_syncobj, uint32_t flags,
           return -ENOMEM;
         kref_init(&syncobj->refcount);
-    INIT_LIST_HEAD(&syncobj->cb_list);
       spin_lock_init(&syncobj->lock);
+    if (flags & DRM_SYNCOBJ_CREATE_TYPE_TIMELINE)
+        syncobj->type = DRM_SYNCOBJ_TYPE_TIMELINE;
+    else
+        syncobj->type = DRM_SYNCOBJ_TYPE_NORMAL;
+    drm_syncobj_timeline_init(syncobj, &syncobj->syncobj_timeline);
         if (flags & DRM_SYNCOBJ_CREATE_SIGNALED) {
           ret = drm_syncobj_assign_null_handle(syncobj);
@@ -651,7 +988,6 @@ struct syncobj_wait_entry {
       struct task_struct *task;
       struct dma_fence *fence;
       struct dma_fence_cb fence_cb;
-    struct drm_syncobj_cb syncobj_cb;
   };
     static void syncobj_wait_fence_func(struct dma_fence *fence,
@@ -663,18 +999,6 @@ static void syncobj_wait_fence_func(struct 
 dma_fence *fence,
       wake_up_process(wait->task);
   }
   -static void syncobj_wait_syncobj_func(struct drm_syncobj *syncobj,
-                      struct drm_syncobj_cb *cb)
-{
-    struct syncobj_wait_entry *wait =
-        container_of(cb, struct syncobj_wait_entry, syncobj_cb);
-
-    /* This happens inside the syncobj lock */
-    wait->fence = 
 dma_fence_get(rcu_dereference_protected(syncobj->fence,
- lockdep_is_held(&syncobj->lock)));
-    wake_up_process(wait->task);
-}
-
   static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,

                            uint32_t count,
                            uint32_t flags,

(quoted diff) +4 -15
@@ -698,14 +1022,11 @@ static signed long 
 drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
       signaled_count = 0;
       for (i = 0; i < count; ++i) {
           entries[i].task = current;
-        entries[i].fence = drm_syncobj_fence_get(syncobjs[i]);
+        ret = drm_syncobj_search_fence(syncobjs[i], 0, 0,
+                           &entries[i].fence);
           if (!entries[i].fence) {
-            if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) {
-                continue;
-            } else {
-                ret = -EINVAL;
-                goto cleanup_entries;
-            }
+            ret = -EINVAL;
+            goto cleanup_entries;
           }
             if (dma_fence_is_signaled(entries[i].fence)) {
@@ -733,15 +1054,6 @@ static signed long 
 drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs,
        * fallthough and try a 0 timeout wait!
        */
   -    if (flags & DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT) {
-        for (i = 0; i < count; ++i) {
- drm_syncobj_fence_get_or_add_callback(syncobjs[i],
-                                  &entries[i].fence,
- &entries[i].syncobj_cb,
- syncobj_wait_syncobj_func);
-        }
-    }
-
       do {
           set_current_state(TASK_INTERRUPTIBLE);

  @@ -789,13 +1101,10 @@ static signed long drm_syncobj_array_wait_timeout(struct drm_syncobj **syncobjs, cleanup_entries:
      for (i = 0; i < count; ++i) {
-        if (entries[i].syncobj_cb.func)
-            drm_syncobj_remove_callback(syncobjs[i],
-                            &entries[i].syncobj_cb);
+        dma_fence_put(entries[i].fence);
          if (entries[i].fence_cb.func)
              dma_fence_remove_callback(entries[i].fence,
                            &entries[i].fence_cb);
-        dma_fence_put(entries[i].fence);
      }
      kfree(entries);
  @@ -970,12 +1279,21 @@ drm_syncobj_reset_ioctl(struct drm_device *dev, void *data,
      if (ret < 0)
          return ret;
  -    for (i = 0; i < args->count_handles; i++)
-        drm_syncobj_replace_fence(syncobjs[i], 0, NULL);

(quoted diff) +12 -1
-
+    for (i = 0; i < args->count_handles; i++) {
+        if (syncobjs[i]->type == DRM_SYNCOBJ_TYPE_TIMELINE) {
+            DRM_ERROR("timeline syncobj cannot reset!\n");
+            ret = -EINVAL;
+            goto out;
+        }
+        drm_syncobj_timeline_fini(syncobjs[i],
+ &syncobjs[i]->syncobj_timeline);
+        drm_syncobj_timeline_init(syncobjs[i],
+ &syncobjs[i]->syncobj_timeline);
+    }
+out:

      drm_syncobj_array_free(syncobjs, args->count_handles); -    return 0;
+    return ret;
  }
    int

drivers/gpu/drm/i915/i915_gem_execbuffer.c +3 -0
index 0a8d2d64f380..579e91a5858b 100644
@@ -2137,7 +2137,9 @@ await_fence_array(struct i915_execbuffer *eb,
           if (!(flags & I915_EXEC_FENCE_WAIT))
               continue;
   -        fence = drm_syncobj_fence_get(syncobj);
+        drm_syncobj_search_fence(syncobj, 0,
+ DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT,
+                     &fence);
           if (!fence)
               return -EINVAL;

  diff --git a/include/drm/drm_syncobj.h b/include/drm/drm_syncobj.h index 425432b85a87..657c97dc25ec 100644

include/drm/drm_syncobj.h +20 -7
@@ -30,6 +30,25 @@
     struct drm_syncobj_cb;
   +enum drm_syncobj_type {
+    DRM_SYNCOBJ_TYPE_NORMAL,
+    DRM_SYNCOBJ_TYPE_TIMELINE
+};
+
+struct drm_syncobj_timeline {
+    wait_queue_head_t    wq;
+    u64 timeline_context;
+    /**
+     * @timeline: syncobj timeline
+     */
+    u64 timeline;
+    u64 signal_point;
+
+
+    struct rb_root wait_pt_tree;
+    struct list_head signal_pt_list;
+};
+
   /**
    * struct drm_syncobj - sync object.
    *
@@ -41,19 +60,16 @@ struct drm_syncobj {
        */
       struct kref refcount;
       /**
-     * @fence:
-     * NULL or a pointer to the fence bound to this object.
-     *
-     * This field should not be used directly. Use 
 drm_syncobj_fence_get()
-     * and drm_syncobj_replace_fence() instead.
+     * @type: indicate syncobj type
        */
-    struct dma_fence __rcu *fence;
+    enum drm_syncobj_type type;
       /**
-     * @cb_list: List of callbacks to call when the &fence gets 

replaced.
+     * @syncobj_timeline: timeline
       */
-    struct list_head cb_list;
+    struct drm_syncobj_timeline syncobj_timeline; +
      /**
-     * @lock: Protects &cb_list and write-locks &fence. +     * @lock: Protects timeline list and write-locks &fence. */
      spinlock_t lock;
      /**

include/drm/drm_syncobj.h +0 -9
@@ -62,25 +78,6 @@ struct drm_syncobj {
       struct file *file;
   };
   -typedef void (*drm_syncobj_func_t)(struct drm_syncobj *syncobj,
-                   struct drm_syncobj_cb *cb);
-
-/**
- * struct drm_syncobj_cb - callback for drm_syncobj_add_callback
- * @node: used by drm_syncob_add_callback to append this struct to
- *      &drm_syncobj.cb_list
- * @func: drm_syncobj_func_t to call
- *
- * This struct will be initialized by drm_syncobj_add_callback, 

additional
- * data can be passed along by embedding drm_syncobj_cb in another struct.
- * The callback will get called the next time drm_syncobj_replace_fence is
- * called.
- */

include/drm/drm_syncobj.h +0 -5
-struct drm_syncobj_cb {
-    struct list_head node;
-    drm_syncobj_func_t func;
-};
-

  void drm_syncobj_free(struct kref *kref);
    /**

include/drm/drm_syncobj.h +0 -0
@@ -106,29 +103,6 @@ drm_syncobj_put(struct drm_syncobj *obj)
       kref_put(&obj->refcount, drm_syncobj_free);
   }

  -/**
- * drm_syncobj_fence_get - get a reference to a fence in a sync object
- * @syncobj: sync object.
- *
- * This acquires additional reference to &drm_syncobj.fence contained in @obj,
- * if not NULL. It is illegal to call this without already holding a reference.
- * No locks required.
- *
- * Returns:
- * Either the fence of @obj or NULL if there's none.
- */

include/drm/drm_syncobj.h +0 -12
-static inline struct dma_fence *
-drm_syncobj_fence_get(struct drm_syncobj *syncobj)
-{
-    struct dma_fence *fence;
-
-    rcu_read_lock();
-    fence = dma_fence_get_rcu_safe(&syncobj->fence);
-    rcu_read_unlock();
-
-    return fence;
-}
-

  struct drm_syncobj *drm_syncobj_find(struct drm_file *file_private, u32 handle);
  void drm_syncobj_replace_fence(struct drm_syncobj *syncobj, u64 point,

include/drm/drm_syncobj.h +2 -0
@@ -142,5 +116,7 @@ int drm_syncobj_create(struct drm_syncobj 
 **out_syncobj, uint32_t flags,
   int drm_syncobj_get_handle(struct drm_file *file_private,
                  struct drm_syncobj *syncobj, u32 *handle);
   int drm_syncobj_get_fd(struct drm_syncobj *syncobj, int *p_fd);
+int drm_syncobj_search_fence(struct drm_syncobj *syncobj, u64 point,
+                 u64 flags, struct dma_fence **fence);
     #endif
include/uapi/drm/drm.h +0 -1
index 300f336633f2..cebdb2541eb7 100644
--- a/include/uapi/drm/drm.h
Quoted 9 lines

include/uapi/drm/drm.h struct drm_prime_handle { · line 722

Quoted 8 lines

+++ b/include/uapi/drm/drm.h

include/uapi/drm/drm.h +1 -0
@@ -717,6 +717,7 @@ struct drm_prime_handle {
   struct drm_syncobj_create {
       __u32 handle;
   #define DRM_SYNCOBJ_CREATE_SIGNALED (1 << 0)
+#define DRM_SYNCOBJ_CREATE_TYPE_TIMELINE (1 << 1)
       __u32 flags;
   };
Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#15 lore raw

Am 13.09.2018 um 04:15 schrieb zhoucm1:

Quoted 12 lines · Christian König
Quoted 11 lines
Quoted 11 lines
Quoted 8 lines
Quoted 8 lines
Quoted 6 lines

[SNIP]
+static void drm_syncobj_find_signal_pt_for_wait_pt(struct drm_syncobj *syncobj,
+                           struct drm_syncobj_wait_pt *wait_pt) +{

That whole approach still looks horrible complicated to me.

It's already very close to what you said before.

Quoted 1 line
Quoted 1 line
Quoted 6 lines
Quoted 5 lines
Quoted 3 lines
Quoted 2 lines
Quoted 2 lines

Especially the separation of signal and wait pt is completely unnecessary as far as I can see. When a wait pt is requested we just need to search for the signal point which it will trigger.

Yeah, I tried this, but when I implement cpu wait ioctl on specific point, we need a advanced wait pt fence, otherwise, we could still need old syncobj cb.

Why? I mean you just need to call drm_syncobj_find_fence() and when that one returns NULL you use wait_event_*() to wait for a signal point >= your wait point to appear and try again.

e.g. when there are 3 syncobjs(A,B,C) to wait, all syncobjABC have no fence yet, as you said, during drm_syncobj_find_fence(A) is working on wait_event, syncobjB and syncobjC could already be signaled, then we don't know which one is first signaled, which is need when wait ioctl returns.

I don't really see a problem with that. When you wait for the first one you need to wait for A,B,C at the same time anyway.

So what you do is to register a fence callback on the fences you already have and for the syncobj which doesn't yet have a fence you make sure that they wake up your thread when they get one.

So essentially exactly what drm_syncobj_fence_get_or_add_callback() already does today.

Regards,
Christian.

Quoted 8 lines

Back to my implementation, it already fixes all your concerns before, and can be able to easily used in wait_ioctl. When you feel that is complicated, I guess that is because we merged all logic to that and much clean up in one patch. In fact, it already is very simple, timeline_init/fini, create signal/wait_pt, find signal_pt for wait_pt, garbage collection, just them.

Thanks,
David Zhou

Quoted 3 lines

Regards,
Christian.

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

A message that is not in the archive (sent off-list or not yet received): 6879599a-0270-3eae-f8a2-1efc154159bb-5C7GfCeVMHo@public.gmane.org

#17 lore raw
Quoted 45 lines

-----Original Message-----
From: Koenig, Christian
Sent: Thursday, September 13, 2018 2:56 PM
To: Zhou, David(ChunMing) <[email protected]>; Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 04:15 schrieb zhoucm1:

Quoted 20 lines · Christian König
Quoted 19 lines
Quoted 18 lines
Quoted 8 lines
Quoted 8 lines
Quoted 6 lines

[SNIP]
+static void drm_syncobj_find_signal_pt_for_wait_pt(struct drm_syncobj *syncobj,
+                           struct drm_syncobj_wait_pt *wait_pt) +{

That whole approach still looks horrible complicated to me.

It's already very close to what you said before.

Quoted 5 lines
Quoted 5 lines

Especially the separation of signal and wait pt is completely unnecessary as far as I can see.
When a wait pt is requested we just need to search for the signal point which it will trigger.

Yeah, I tried this, but when I implement cpu wait ioctl on specific point, we need a advanced wait pt fence, otherwise, we could still need old syncobj cb.

Quoted 8 lines
Quoted 3 lines

Why? I mean you just need to call drm_syncobj_find_fence() and when that one returns NULL you use wait_event_*() to wait for a signal point >= your wait point to appear and try again.

e.g. when there are 3 syncobjs(A,B,C) to wait, all syncobjABC have no fence yet, as you said, during drm_syncobj_find_fence(A) is working on wait_event, syncobjB and syncobjC could already be signaled, then we don't know which one is first signaled, which is need when wait ioctl returns.

I don't really see a problem with that. When you wait for the first one you need to wait for A,B,C at the same time anyway.

Quoted 6 lines

So what you do is to register a fence callback on the fences you already have and for the syncobj which doesn't yet have a fence you make sure that they wake up your thread when they get one.

So essentially exactly what drm_syncobj_fence_get_or_add_callback() already does today.

So do you mean we need still use old syncobj CB for that? Advanced wait pt is bad?

Thanks,
David Zhou

Quoted 17 lines

Regards,
Christian.

Quoted 13 lines

Back to my implementation, it already fixes all your concerns before, and can be able to easily used in wait_ioctl. When you feel that is complicated, I guess that is because we merged all logic to that and much clean up in one patch. In fact, it already is very simple, timeline_init/fini, create signal/wait_pt, find signal_pt for wait_pt, garbage collection, just them.

Thanks,
David Zhou

Quoted 3 lines

Regards,
Christian.

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): BY1PR12MB0502C7E29899EBF4B0379B0BB41A0-PicGAnIBOobrCwm+z9iKNgdYzm3356FpvxpqHgZTriW3zl9H0oFU5g@public.gmane.org

#19 lore raw

Am 13.09.2018 um 09:43 schrieb Zhou, David(ChunMing):

Quoted 39 lines
Quoted 39 lines

-----Original Message-----
From: Koenig, Christian
Sent: Thursday, September 13, 2018 2:56 PM
To: Zhou, David(ChunMing) <[email protected]>; Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 04:15 schrieb zhoucm1:

Quoted 16 lines · Christian König
Quoted 15 lines
Quoted 15 lines
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines

[SNIP]
+static void drm_syncobj_find_signal_pt_for_wait_pt(struct drm_syncobj *syncobj,
+                           struct drm_syncobj_wait_pt *wait_pt) +{

That whole approach still looks horrible complicated to me.

It's already very close to what you said before.

Quoted 4 lines
Quoted 4 lines

Especially the separation of signal and wait pt is completely unnecessary as far as I can see.
When a wait pt is requested we just need to search for the signal point which it will trigger.

Yeah, I tried this, but when I implement cpu wait ioctl on specific point, we need a advanced wait pt fence, otherwise, we could still need old syncobj cb.

Quoted 8 lines
Quoted 3 lines

Why? I mean you just need to call drm_syncobj_find_fence() and when that one returns NULL you use wait_event_*() to wait for a signal point >= your wait point to appear and try again.

e.g. when there are 3 syncobjs(A,B,C) to wait, all syncobjABC have no fence yet, as you said, during drm_syncobj_find_fence(A) is working on wait_event, syncobjB and syncobjC could already be signaled, then we don't know which one is first signaled, which is need when wait ioctl returns.

I don't really see a problem with that. When you wait for the first one you need to wait for A,B,C at the same time anyway.

Quoted 7 lines
Quoted 6 lines

So what you do is to register a fence callback on the fences you already have and for the syncobj which doesn't yet have a fence you make sure that they wake up your thread when they get one.

So essentially exactly what drm_syncobj_fence_get_or_add_callback() already does today.

So do you mean we need still use old syncobj CB for that?

Yes, as far as I can see it should work.

Quoted 1 line

Advanced wait pt is bad?

Well it isn't bad, I just don't see any advantage in it. The existing mechanism should already be able to handle that.

Christian.

Quoted 13 lines

Thanks,
David Zhou

Quoted 10 lines

Regards,
Christian.

Quoted 7 lines

Back to my implementation, it already fixes all your concerns before, and can be able to easily used in wait_ioctl. When you feel that is complicated, I guess that is because we merged all logic to that and much clean up in one patch. In fact, it already is very simple, timeline_init/fini, create signal/wait_pt, find signal_pt for wait_pt, garbage collection, just them.

Quoted 8 lines
Quoted 4 lines
Quoted 4 lines

Thanks,
David Zhou

Quoted 2 lines

Regards,
Christian.

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): ead45230-65dd-ecd8-5153-8312dd2a9480-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org

#21 lore raw
Quoted 58 lines

-----Original Message-----
From: Christian König <[email protected]>
Sent: Thursday, September 13, 2018 4:50 PM
To: Zhou, David(ChunMing) <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 09:43 schrieb Zhou, David(ChunMing):

Quoted 21 lines
Quoted 21 lines

-----Original Message-----
From: Koenig, Christian
Sent: Thursday, September 13, 2018 2:56 PM
To: Zhou, David(ChunMing) <[email protected]>; Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 04:15 schrieb zhoucm1:

Quoted 9 lines · Christian König
Quoted 8 lines
Quoted 8 lines
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines

[SNIP]
+static void drm_syncobj_find_signal_pt_for_wait_pt(struct drm_syncobj *syncobj,
+                           struct drm_syncobj_wait_pt +*wait_pt) {

That whole approach still looks horrible complicated to me.

It's already very close to what you said before.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines
Quoted 4 lines
Quoted 4 lines

Especially the separation of signal and wait pt is completely unnecessary as far as I can see.
When a wait pt is requested we just need to search for the signal point which it will trigger.

Yeah, I tried this, but when I implement cpu wait ioctl on specific point, we need a advanced wait pt fence, otherwise, we could still need old syncobj cb.

Why? I mean you just need to call drm_syncobj_find_fence() and

when

Quoted 16 lines
Quoted 16 lines
Quoted 7 lines
Quoted 2 lines

that one returns NULL you use wait_event_*() to wait for a signal point >= your wait point to appear and try again.

e.g. when there are 3 syncobjs(A,B,C) to wait, all syncobjABC have no fence yet, as you said, during drm_syncobj_find_fence(A) is working on wait_event, syncobjB and syncobjC could already be signaled, then we don't know which one is first signaled, which is need when wait ioctl returns.

I don't really see a problem with that. When you wait for the first one you need to wait for A,B,C at the same time anyway.

So what you do is to register a fence callback on the fences you already have and for the syncobj which doesn't yet have a fence you make sure that they wake up your thread when they get one.

So essentially exactly what drm_syncobj_fence_get_or_add_callback() already does today.

Quoted 7 lines
Quoted 1 line

So do you mean we need still use old syncobj CB for that?

Yes, as far as I can see it should work.

Quoted 1 line

Advanced wait pt is bad?

Well it isn't bad, I just don't see any advantage in it.

The advantage is to replace old syncobj cb.

Quoted 2 lines

The existing mechanism
should already be able to handle that.

I thought more a bit, we don't that mechanism at all, if use advanced wait pt, we can easily use fence array to achieve it for wait ioctl, we should use kernel existing feature as much as possible, not invent another, shouldn't we? I remember you said it before.

Thanks,
David Zhou

Quoted 24 lines

Christian.

Quoted 21 lines

Thanks,
David Zhou

Quoted 14 lines

Regards,
Christian.

Quoted 11 lines

Back to my implementation, it already fixes all your concerns before, and can be able to easily used in wait_ioctl. When you feel that is complicated, I guess that is because we merged all logic to that and much clean up in one patch. In fact, it already is very simple, timeline_init/fini, create signal/wait_pt, find signal_pt for wait_pt, garbage collection, just them.

Thanks,
David Zhou

Quoted 2 lines

Regards,
Christian.

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): BY1PR12MB0502DD202E93FFAD340FE931B41A0-PicGAnIBOobrCwm+z9iKNgdYzm3356FpvxpqHgZTriW3zl9H0oFU5g@public.gmane.org

#23 lore raw

Am 13.09.2018 um 11:11 schrieb Zhou, David(ChunMing):

Quoted 59 lines
Quoted 59 lines

-----Original Message-----
From: Christian König <[email protected]>
Sent: Thursday, September 13, 2018 4:50 PM
To: Zhou, David(ChunMing) <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 09:43 schrieb Zhou, David(ChunMing):

Quoted 20 lines
Quoted 20 lines

-----Original Message-----
From: Koenig, Christian
Sent: Thursday, September 13, 2018 2:56 PM
To: Zhou, David(ChunMing) <[email protected]>; Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 04:15 schrieb zhoucm1:

Quoted 9 lines · Christian König
Quoted 8 lines
Quoted 8 lines
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines

[SNIP]
+static void drm_syncobj_find_signal_pt_for_wait_pt(struct drm_syncobj *syncobj,
+                           struct drm_syncobj_wait_pt +*wait_pt) {

That whole approach still looks horrible complicated to me.

It's already very close to what you said before.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines
Quoted 4 lines
Quoted 4 lines

Especially the separation of signal and wait pt is completely unnecessary as far as I can see.
When a wait pt is requested we just need to search for the signal point which it will trigger.

Yeah, I tried this, but when I implement cpu wait ioctl on specific point, we need a advanced wait pt fence, otherwise, we could still need old syncobj cb.

Why? I mean you just need to call drm_syncobj_find_fence() and

when

Quoted 10 lines
Quoted 10 lines
Quoted 7 lines
Quoted 2 lines

that one returns NULL you use wait_event_*() to wait for a signal point >= your wait point to appear and try again.

e.g. when there are 3 syncobjs(A,B,C) to wait, all syncobjABC have no fence yet, as you said, during drm_syncobj_find_fence(A) is working on wait_event, syncobjB and syncobjC could already be signaled, then we don't know which one is first signaled, which is need when wait ioctl returns.

I don't really see a problem with that. When you wait for the first one you need to wait for A,B,C at the same time anyway.

Quoted 7 lines
Quoted 6 lines

So what you do is to register a fence callback on the fences you already have and for the syncobj which doesn't yet have a fence you make sure that they wake up your thread when they get one.

So essentially exactly what drm_syncobj_fence_get_or_add_callback() already does today.

So do you mean we need still use old syncobj CB for that?

Yes, as far as I can see it should work.

Quoted 8 lines
Quoted 2 lines
Quoted 1 line

Advanced wait pt is bad?

Well it isn't bad, I just don't see any advantage in it.

The advantage is to replace old syncobj cb.

Quoted 2 lines

The existing mechanism
should already be able to handle that.

I thought more a bit, we don't that mechanism at all, if use advanced wait pt, we can easily use fence array to achieve it for wait ioctl, we should use kernel existing feature as much as possible, not invent another, shouldn't we? I remember you said it before.

Yeah, but the syncobj cb is an existing feature. And I absolutely don't see a need to modify that and replace it with something far more complex.

Regards,
Christian.

Quoted 25 lines

Thanks,
David Zhou

Quoted 22 lines

Christian.

Quoted 20 lines

Thanks,
David Zhou

Quoted 14 lines

Regards,
Christian.

Quoted 11 lines

Back to my implementation, it already fixes all your concerns before, and can be able to easily used in wait_ioctl. When you feel that is complicated, I guess that is because we merged all logic to that and much clean up in one patch. In fact, it already is very simple, timeline_init/fini, create signal/wait_pt, find signal_pt for wait_pt, garbage collection, just them.

Thanks,
David Zhou

Quoted 2 lines

Regards,
Christian.

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 54f6a4e6-36e1-fb6d-967a-81a105ea271d-5C7GfCeVMHo@public.gmane.org

#25 lore raw
Quoted 76 lines

-----Original Message-----
From: Koenig, Christian
Sent: Thursday, September 13, 2018 5:20 PM
To: Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 11:11 schrieb Zhou, David(ChunMing):

Quoted 64 lines
Quoted 54 lines

-----Original Message-----
From: Christian König <[email protected]>
Sent: Thursday, September 13, 2018 4:50 PM
To: Zhou, David(ChunMing) <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 09:43 schrieb Zhou, David(ChunMing):

Quoted 20 lines
Quoted 20 lines

-----Original Message-----
From: Koenig, Christian
Sent: Thursday, September 13, 2018 2:56 PM
To: Zhou, David(ChunMing) <[email protected]>; Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 04:15 schrieb zhoucm1:

Quoted 9 lines · Christian König
Quoted 8 lines
Quoted 8 lines
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines

[SNIP]
+static void drm_syncobj_find_signal_pt_for_wait_pt(struct drm_syncobj *syncobj,
+                           struct drm_syncobj_wait_pt +*wait_pt) {

That whole approach still looks horrible complicated to me.

It's already very close to what you said before.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines
Quoted 4 lines
Quoted 4 lines

Especially the separation of signal and wait pt is completely unnecessary as far as I can see.
When a wait pt is requested we just need to search for the signal point which it will trigger.

Yeah, I tried this, but when I implement cpu wait ioctl on specific point, we need a advanced wait pt fence, otherwise, we could still need old syncobj cb.

Why? I mean you just need to call drm_syncobj_find_fence() and

when

Quoted 14 lines
Quoted 14 lines
Quoted 7 lines
Quoted 2 lines

that one returns NULL you use wait_event_*() to wait for a signal point >= your wait point to appear and try again.

e.g. when there are 3 syncobjs(A,B,C) to wait, all syncobjABC have no fence yet, as you said, during drm_syncobj_find_fence(A) is working on wait_event, syncobjB and syncobjC could already be signaled, then we don't know which one is first signaled, which is need when wait ioctl returns.

I don't really see a problem with that. When you wait for the first one you need to wait for A,B,C at the same time anyway.

So what you do is to register a fence callback on the fences you already have and for the syncobj which doesn't yet have a fence you make sure that they wake up your thread when they get one.

Quoted 7 lines
Quoted 3 lines
Quoted 2 lines

So essentially exactly what drm_syncobj_fence_get_or_add_callback() already does today.

So do you mean we need still use old syncobj CB for that?

Yes, as far as I can see it should work.

Quoted 1 line

Advanced wait pt is bad?

Well it isn't bad, I just don't see any advantage in it.

The advantage is to replace old syncobj cb.

Quoted 8 lines
Quoted 3 lines
Quoted 2 lines

The existing mechanism
should already be able to handle that.

I thought more a bit, we don't that mechanism at all, if use advanced wait

pt, we can easily use fence array to achieve it for wait ioctl, we should use kernel existing feature as much as possible, not invent another, shouldn't we? I remember you said it before.

Yeah, but the syncobj cb is an existing feature.

This is obviously a workaround when doing for wait ioctl, Do you see it used in other place?

Quoted 2 lines

And I absolutely don't see a
need to modify that and replace it with something far more complex.

The wait ioctl is simplified much more by fence array, not complex, and we just need to allocate a wait pt. If keeping old syncobj cb workaround, all wait pt logic still is there, just save allocation and wait pt handling, in fact, which part isn't complex at all. But compare with ugly syncobj cb, which is simpler.

Thanks,
David Zhou

Quoted 29 lines

Regards,
Christian.

Quoted 25 lines

Thanks,
David Zhou

Quoted 22 lines

Christian.

Quoted 20 lines

Thanks,
David Zhou

Quoted 14 lines

Regards,
Christian.

Quoted 11 lines

Back to my implementation, it already fixes all your concerns before, and can be able to easily used in wait_ioctl. When you feel that is complicated, I guess that is because we merged all logic to that and much clean up in one patch. In fact, it already is very simple, timeline_init/fini, create signal/wait_pt, find signal_pt for wait_pt, garbage collection, just them.

Thanks,
David Zhou

Quoted 2 lines

Regards,
Christian.

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): BY1PR12MB05020A2E66D19438D4069CFCB41A0-PicGAnIBOobrCwm+z9iKNgdYzm3356FpvxpqHgZTriW3zl9H0oFU5g@public.gmane.org

#27 lore raw

Am 13.09.2018 um 11:35 schrieb Zhou, David(ChunMing):

Quoted 80 lines
Quoted 80 lines

-----Original Message-----
From: Koenig, Christian
Sent: Thursday, September 13, 2018 5:20 PM
To: Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 11:11 schrieb Zhou, David(ChunMing):

Quoted 58 lines
Quoted 58 lines

-----Original Message-----
From: Christian König <[email protected]>
Sent: Thursday, September 13, 2018 4:50 PM
To: Zhou, David(ChunMing) <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 09:43 schrieb Zhou, David(ChunMing):

Quoted 20 lines
Quoted 20 lines

-----Original Message-----
From: Koenig, Christian
Sent: Thursday, September 13, 2018 2:56 PM
To: Zhou, David(ChunMing) <[email protected]>; Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 04:15 schrieb zhoucm1:

Quoted 9 lines · Christian König
Quoted 8 lines
Quoted 8 lines
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines

[SNIP]
+static void drm_syncobj_find_signal_pt_for_wait_pt(struct drm_syncobj *syncobj,
+                           struct drm_syncobj_wait_pt +*wait_pt) {

That whole approach still looks horrible complicated to me.

It's already very close to what you said before.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines
Quoted 4 lines
Quoted 4 lines

Especially the separation of signal and wait pt is completely unnecessary as far as I can see.
When a wait pt is requested we just need to search for the signal point which it will trigger.

Yeah, I tried this, but when I implement cpu wait ioctl on specific point, we need a advanced wait pt fence, otherwise, we could still need old syncobj cb.

Why? I mean you just need to call drm_syncobj_find_fence() and

when

Quoted 10 lines
Quoted 10 lines
Quoted 7 lines
Quoted 2 lines

that one returns NULL you use wait_event_*() to wait for a signal point >= your wait point to appear and try again.

e.g. when there are 3 syncobjs(A,B,C) to wait, all syncobjABC have no fence yet, as you said, during drm_syncobj_find_fence(A) is working on wait_event, syncobjB and syncobjC could already be signaled, then we don't know which one is first signaled, which is need when wait ioctl returns.

I don't really see a problem with that. When you wait for the first one you need to wait for A,B,C at the same time anyway.

Quoted 7 lines
Quoted 6 lines

So what you do is to register a fence callback on the fences you already have and for the syncobj which doesn't yet have a fence you make sure that they wake up your thread when they get one.

So essentially exactly what drm_syncobj_fence_get_or_add_callback() already does today.

So do you mean we need still use old syncobj CB for that?

Yes, as far as I can see it should work.

Quoted 7 lines
Quoted 2 lines
Quoted 1 line

Advanced wait pt is bad?

Well it isn't bad, I just don't see any advantage in it.

The advantage is to replace old syncobj cb.

Quoted 2 lines

The existing mechanism
should already be able to handle that.

I thought more a bit, we don't that mechanism at all, if use advanced wait

pt, we can easily use fence array to achieve it for wait ioctl, we should use kernel existing feature as much as possible, not invent another, shouldn't we? I remember you said it before.

Quoted 6 lines
Quoted 1 line

Yeah, but the syncobj cb is an existing feature.

This is obviously a workaround when doing for wait ioctl, Do you see it used in other place?

Quoted 2 lines

And I absolutely don't see a
need to modify that and replace it with something far more complex.

The wait ioctl is simplified much more by fence array, not complex, and we just need to allocate a wait pt. If keeping old syncobj cb workaround, all wait pt logic still is there, just save allocation and wait pt handling, in fact, which part isn't complex at all. But compare with ugly syncobj cb, which is simpler.

I strongly disagree on that. You just need to extend the syncobj cb with the sequence number and you are done.

We could clean that up in the long term by adding some wait_multi event macro, but for now just adding the sequence number should do the trick.

Regards,
Christian.

Quoted 30 lines

Thanks,
David Zhou

Quoted 27 lines

Regards,
Christian.

Quoted 24 lines

Thanks,
David Zhou

Quoted 22 lines

Christian.

Quoted 20 lines

Thanks,
David Zhou

Quoted 14 lines

Regards,
Christian.

Quoted 11 lines

Back to my implementation, it already fixes all your concerns before, and can be able to easily used in wait_ioctl. When you feel that is complicated, I guess that is because we merged all logic to that and much clean up in one patch. In fact, it already is very simple, timeline_init/fini, create signal/wait_pt, find signal_pt for wait_pt, garbage collection, just them.

Thanks,
David Zhou

Quoted 2 lines

Regards,
Christian.

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): c077c499-dcad-2545-d918-75f13a1d3a14-5C7GfCeVMHo@public.gmane.org

#29 lore raw
Quoted 84 lines · Christian König

Am 13.09.2018 um 11:35 schrieb Zhou, David(ChunMing):

Quoted 83 lines
Quoted 73 lines

-----Original Message-----
From: Koenig, Christian
Sent: Thursday, September 13, 2018 5:20 PM
To: Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 11:11 schrieb Zhou, David(ChunMing):

Quoted 62 lines
Quoted 53 lines

-----Original Message-----
From: Christian König <[email protected]>
Sent: Thursday, September 13, 2018 4:50 PM
To: Zhou, David(ChunMing) <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 09:43 schrieb Zhou, David(ChunMing):

Quoted 20 lines
Quoted 20 lines

-----Original Message-----
From: Koenig, Christian
Sent: Thursday, September 13, 2018 2:56 PM
To: Zhou, David(ChunMing) <[email protected]>; Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 04:15 schrieb zhoucm1:

Quoted 9 lines · Christian König
Quoted 8 lines
Quoted 8 lines
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines

[SNIP]
+static void drm_syncobj_find_signal_pt_for_wait_pt(struct
drm_syncobj *syncobj,
+                           struct drm_syncobj_wait_pt
+*wait_pt) {

That whole approach still looks horrible complicated to me.

It's already very close to what you said before.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines
Quoted 4 lines
Quoted 4 lines

Especially the separation of signal and wait pt is completely unnecessary as far as I can see.
When a wait pt is requested we just need to search for the signal point which it will trigger.

Yeah, I tried this, but when I implement cpu wait ioctl on specific point, we need a advanced wait pt fence, otherwise, we could still need old syncobj cb.

Why? I mean you just need to call drm_syncobj_find_fence() and

when

Quoted 14 lines
Quoted 14 lines
Quoted 7 lines
Quoted 2 lines

that one returns NULL you use wait_event_*() to wait for a signal point >= your wait point to appear and try again.

e.g. when there are 3 syncobjs(A,B,C) to wait, all syncobjABC have no fence yet, as you said, during drm_syncobj_find_fence(A) is working on wait_event, syncobjB and syncobjC could already be signaled, then we don't know which one is first signaled, which is need when wait ioctl returns.

I don't really see a problem with that. When you wait for the first one you need to wait for A,B,C at the same time anyway.

So what you do is to register a fence callback on the fences you already have and for the syncobj which doesn't yet have a fence you make sure that they wake up your thread when they get one.

Quoted 7 lines
Quoted 3 lines
Quoted 2 lines

So essentially exactly what drm_syncobj_fence_get_or_add_callback() already does today.

So do you mean we need still use old syncobj CB for that?

Yes, as far as I can see it should work.

Quoted 1 line

    Advanced wait pt is bad?

Well it isn't bad, I just don't see any advantage in it.

The advantage is to replace old syncobj cb.

Quoted 8 lines
Quoted 3 lines
Quoted 2 lines

The existing mechanism
should already be able to handle that.

I thought more a bit, we don't that mechanism at all, if use advanced wait

pt, we can easily use fence array to achieve it for wait ioctl, we should use kernel existing feature as much as possible, not invent another, shouldn't we? I remember  you said  it before.

Yeah, but the syncobj cb is an existing feature.

This is obviously a workaround when doing for wait ioctl, Do you see it used in other place?

Quoted 7 lines
Quoted 3 lines
Quoted 2 lines

And I absolutely don't see a
need to modify that and replace it with something far more complex.

The wait ioctl is simplified much more by fence array, not complex, and we just need  to allocate a wait pt.  If keeping old syncobj cb workaround, all wait pt logic still is there, just save allocation and wait pt handling, in fact, which part isn't complex at all. But compare with ugly syncobj cb, which is simpler.

I strongly disagree on that. You just need to extend the syncobj cb with the sequence number and you are done.

We could clean that up in the long term by adding some wait_multi event macro, but for now just adding the sequence number should do the trick.

Quote from Daniel Vetter comment when v1, "

Specifically for this stuff here having unified future fence semantics will allow drivers to do clever stuff with them.

"
I think the advanced wait pt is a similar concept as 'future fence' what Daniel Vetter said before, which obviously a right direction.

Anyway, I will change the patch as you like if no other comment, so that the patch can pass soon.

Thanks,
David Zhou

Quoted 35 lines

Regards,
Christian.

Quoted 30 lines

Thanks,
David Zhou

Quoted 27 lines

Regards,
Christian.

Quoted 24 lines

Thanks,
David Zhou

Quoted 22 lines

Christian.

Quoted 20 lines

Thanks,
David Zhou

Quoted 14 lines

Regards,
Christian.

Quoted 11 lines

Back to my implementation, it already fixes all your concerns before, and can be able to easily used in wait_ioctl. When you feel that is complicated, I guess that is because we merged all logic to that and much clean up in one patch. In fact, it already is very simple, timeline_init/fini, create signal/wait_pt, find signal_pt for wait_pt, garbage collection, just them.

Thanks,
David Zhou

Quoted 2 lines

Regards,
Christian.

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): a47d925d-c37f-d60c-db8d-dfb4ea570dca-5C7GfCeVMHo@public.gmane.org

#31 lore raw
Quoted 97 lines · On 2018年09月14日 11:14, zhoucm1 wrote:
Quoted 84 lines · Christian König

Am 13.09.2018 um 11:35 schrieb Zhou, David(ChunMing):

Quoted 83 lines
Quoted 73 lines

-----Original Message-----
From: Koenig, Christian
Sent: Thursday, September 13, 2018 5:20 PM
To: Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 11:11 schrieb Zhou, David(ChunMing):

Quoted 62 lines
Quoted 53 lines

-----Original Message-----
From: Christian König <[email protected]>
Sent: Thursday, September 13, 2018 4:50 PM
To: Zhou, David(ChunMing) <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 09:43 schrieb Zhou, David(ChunMing):

Quoted 20 lines
Quoted 20 lines

-----Original Message-----
From: Koenig, Christian
Sent: Thursday, September 13, 2018 2:56 PM
To: Zhou, David(ChunMing) <[email protected]>; Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 04:15 schrieb zhoucm1:

Quoted 9 lines · Christian König
Quoted 8 lines
Quoted 8 lines
Quoted 6 lines
Quoted 6 lines

[SNIP]
+static void drm_syncobj_find_signal_pt_for_wait_pt(struct
drm_syncobj *syncobj,
+                           struct drm_syncobj_wait_pt
+*wait_pt) {
That whole approach still looks horrible complicated to me.

It's already very close to what you said before.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines
Quoted 4 lines
Quoted 4 lines

Especially the separation of signal and wait pt is completely
unnecessary as far as I can see.
When a wait pt is requested we just need to search for the
signal point which it will trigger.

Yeah, I tried this, but when I implement cpu wait ioctl on specific point, we need a advanced wait pt fence, otherwise, we could still need old syncobj cb.

Why? I mean you just need to call drm_syncobj_find_fence() and

when

Quoted 14 lines
Quoted 14 lines
Quoted 7 lines
Quoted 2 lines

that one returns NULL you use wait_event_*() to wait for a signal point >= your wait point to appear and try again.

e.g. when there are 3 syncobjs(A,B,C) to wait, all syncobjABC have no fence yet, as you said, during drm_syncobj_find_fence(A) is working on wait_event, syncobjB and syncobjC could already be signaled, then we don't know which one is first signaled, which is need when wait ioctl returns.

I don't really see a problem with that. When you wait for the first one you need to wait for A,B,C at the same time anyway.

So what you do is to register a fence callback on the fences you already have and for the syncobj which doesn't yet have a fence you make sure that they wake up your thread when they get one.

Quoted 7 lines
Quoted 3 lines
Quoted 2 lines

So essentially exactly what drm_syncobj_fence_get_or_add_callback() already does today.

So do you mean we need still use old syncobj CB for that?

Yes, as far as I can see it should work.

Quoted 1 line

    Advanced wait pt is bad?

Well it isn't bad, I just don't see any advantage in it.

The advantage is to replace old syncobj cb.

Quoted 8 lines
Quoted 3 lines
Quoted 2 lines

The existing mechanism
should already be able to handle that.

I thought more a bit, we don't that mechanism at all, if use advanced wait

pt, we can easily use fence array to achieve it for wait ioctl, we should use kernel existing feature as much as possible, not invent another, shouldn't we? I remember  you said  it before.

Yeah, but the syncobj cb is an existing feature.

This is obviously a workaround when doing for wait ioctl, Do you see it used in other place?

Quoted 7 lines
Quoted 3 lines
Quoted 2 lines

And I absolutely don't see a
need to modify that and replace it with something far more complex.

The wait ioctl is simplified much more by fence array, not complex, and we just need  to allocate a wait pt.  If keeping old syncobj cb workaround, all wait pt logic still is there, just save allocation and wait pt handling, in fact, which part isn't complex at all. But compare with ugly syncobj cb, which is simpler.

I strongly disagree on that. You just need to extend the syncobj cb with the sequence number and you are done.

We could clean that up in the long term by adding some wait_multi event macro, but for now just adding the sequence number should do the trick.

Quote from Daniel Vetter comment when v1, "

Quoted 8 lines

Specifically for this stuff here having unified future fence semantics will allow drivers to do clever stuff with them.

"
I think the advanced wait pt is a similar concept as 'future fence' what Daniel Vetter said before, which obviously a right direction.

Anyway, I will change the patch as you like if no other comment, so that the patch can pass soon.

When I try to remove wait pt future fence, I encounter another problem, drm_syncobj_find_fence cannot get a fence if signal pt already is collected as garbage, then CS will report error, any idea for that? I still think the future fence is right thing, Could you give futher thought on it again? Otherwise, we could need various workarounds.

Thanks,
David Zhou

Quoted 39 lines

Thanks,
David Zhou

Quoted 35 lines

Regards,
Christian.

Quoted 30 lines

Thanks,
David Zhou

Quoted 27 lines

Regards,
Christian.

Quoted 24 lines

Thanks,
David Zhou

Quoted 22 lines

Christian.

Quoted 20 lines

Thanks,
David Zhou

Quoted 14 lines

Regards,
Christian.

Quoted 11 lines

Back to my implementation, it already fixes all your concerns before, and can be able to easily used in wait_ioctl. When you feel that is complicated, I guess that is because we merged all logic to that and much clean up in one patch. In fact, it already is very simple, timeline_init/fini, create signal/wait_pt, find signal_pt for wait_pt, garbage collection, just them.

Thanks,
David Zhou

Quoted 2 lines

Regards,
Christian.

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#32 lore raw

Am 14.09.2018 um 05:59 schrieb zhoucm1:

Quoted 103 lines · On 2018年09月14日 11:14, zhoucm1 wrote:
Quoted 100 lines · Christian König
Quoted 84 lines

Am 13.09.2018 um 11:35 schrieb Zhou, David(ChunMing):

Quoted 83 lines
Quoted 73 lines

-----Original Message-----
From: Koenig, Christian
Sent: Thursday, September 13, 2018 5:20 PM
To: Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 11:11 schrieb Zhou, David(ChunMing):

Quoted 62 lines
Quoted 53 lines

-----Original Message-----
From: Christian König <[email protected]>
Sent: Thursday, September 13, 2018 4:50 PM
To: Zhou, David(ChunMing) <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 09:43 schrieb Zhou, David(ChunMing):

Quoted 20 lines
Quoted 20 lines

-----Original Message-----
From: Koenig, Christian
Sent: Thursday, September 13, 2018 2:56 PM
To: Zhou, David(ChunMing) <[email protected]>; Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 04:15 schrieb zhoucm1:

Quoted 9 lines · Christian König
Quoted 8 lines
Quoted 8 lines
Quoted 6 lines

[SNIP]
+static void drm_syncobj_find_signal_pt_for_wait_pt(struct
drm_syncobj *syncobj,
+                           struct drm_syncobj_wait_pt
+*wait_pt) {
That whole approach still looks horrible complicated to me.

It's already very close to what you said before.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines
Quoted 4 lines

Especially the separation of signal and wait pt is completely
unnecessary as far as I can see.
When a wait pt is requested we just need to search for the
signal point which it will trigger.

Yeah, I tried this, but when I implement cpu wait ioctl on specific point, we need a advanced wait pt fence, otherwise, we could still need old syncobj cb.

Why? I mean you just need to call drm_syncobj_find_fence() and

when

Quoted 14 lines
Quoted 14 lines
Quoted 7 lines
Quoted 2 lines

that one returns NULL you use wait_event_*() to wait for a signal point >= your wait point to appear and try again.

e.g. when there are 3 syncobjs(A,B,C) to wait, all syncobjABC have no fence yet, as you said, during drm_syncobj_find_fence(A) is working on wait_event, syncobjB and syncobjC could already be signaled, then we don't know which one is first signaled, which is need when wait ioctl returns.

I don't really see a problem with that. When you wait for the first one you need to wait for A,B,C at the same time anyway.

So what you do is to register a fence callback on the fences you already have and for the syncobj which doesn't yet have a fence you make sure that they wake up your thread when they get one.

Quoted 7 lines
Quoted 3 lines
Quoted 2 lines

So essentially exactly what drm_syncobj_fence_get_or_add_callback() already does today.

So do you mean we need still use old syncobj CB for that?

Yes, as far as I can see it should work.

Quoted 1 line

    Advanced wait pt is bad?

Well it isn't bad, I just don't see any advantage in it.

The advantage is to replace old syncobj cb.

Quoted 8 lines
Quoted 3 lines
Quoted 2 lines

The existing mechanism
should already be able to handle that.

I thought more a bit, we don't that mechanism at all, if use advanced wait

pt, we can easily use fence array to achieve it for wait ioctl, we should use kernel existing feature as much as possible, not invent another, shouldn't we? I remember  you said  it before.

Yeah, but the syncobj cb is an existing feature.

This is obviously a workaround when doing for wait ioctl, Do you see it used in other place?

Quoted 7 lines
Quoted 3 lines
Quoted 2 lines

And I absolutely don't see a
need to modify that and replace it with something far more complex.

The wait ioctl is simplified much more by fence array, not complex, and we just need  to allocate a wait pt.  If keeping old syncobj cb workaround, all wait pt logic still is there, just save allocation and wait pt handling, in fact, which part isn't complex at all. But compare with ugly syncobj cb, which is simpler.

I strongly disagree on that. You just need to extend the syncobj cb with the sequence number and you are done.

We could clean that up in the long term by adding some wait_multi event macro, but for now just adding the sequence number should do the trick.

Quote from Daniel Vetter comment when v1, "

Specifically for this stuff here having unified future fence semantics will allow drivers to do clever stuff with them.

Quoted 6 lines
Quoted 5 lines

"
I think the advanced wait pt is a similar concept as 'future fence' what Daniel Vetter said before, which obviously a right direction.

Anyway, I will change the patch as you like if no other comment, so that the patch can pass soon.

When I try to remove wait pt future fence, I encounter another problem, drm_syncobj_find_fence cannot get a fence if signal pt already is collected as garbage, then CS will report error, any idea for that?

Well when the signal pt is already garbage collected you know that it is already signaled. So you can just return a dummy fence.

I actually thought that this was the intention of the dummy fence rename :)

Christian.

Quoted 44 lines

I still think the future fence is right thing, Could you give futher thought on it again? Otherwise, we could need various workarounds.

Thanks,
David Zhou

Quoted 39 lines

Thanks,
David Zhou

Quoted 35 lines

Regards,
Christian.

Quoted 30 lines

Thanks,
David Zhou

Quoted 27 lines

Regards,
Christian.

Quoted 24 lines

Thanks,
David Zhou

Quoted 22 lines

Christian.

Quoted 20 lines

Thanks,
David Zhou

Quoted 14 lines

Regards,
Christian.

Quoted 11 lines

Back to my implementation, it already fixes all your concerns before, and can be able to easily used in wait_ioctl. When you feel that is complicated, I guess that is because we merged all logic to that and much clean up in one patch. In fact, it already is very simple, timeline_init/fini, create signal/wait_pt, find signal_pt for wait_pt, garbage collection, just them.

Thanks,
David Zhou

Quoted 2 lines

Regards,
Christian.

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#33 lore raw
Quoted 148 lines

-----Original Message-----
From: Koenig, Christian
Sent: Friday, September 14, 2018 3:27 PM
To: Zhou, David(ChunMing) <[email protected]>; Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]; Daniel Vetter <[email protected]>
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 14.09.2018 um 05:59 schrieb zhoucm1:

Quoted 59 lines · On 2018年09月14日 11:14, zhoucm1 wrote:
Quoted 56 lines · Christian König
Quoted 53 lines

Am 13.09.2018 um 11:35 schrieb Zhou, David(ChunMing):

Quoted 52 lines
Quoted 52 lines

-----Original Message-----
From: Koenig, Christian
Sent: Thursday, September 13, 2018 5:20 PM
To: Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 11:11 schrieb Zhou, David(ChunMing):

Quoted 41 lines
Quoted 41 lines

-----Original Message-----
From: Christian König <[email protected]>
Sent: Thursday, September 13, 2018 4:50 PM
To: Zhou, David(ChunMing) <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 09:43 schrieb Zhou, David(ChunMing):

Quoted 29 lines
Quoted 29 lines

-----Original Message-----
From: Koenig, Christian
Sent: Thursday, September 13, 2018 2:56 PM
To: Zhou, David(ChunMing) <[email protected]>; Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 04:15 schrieb zhoucm1:

Quoted 17 lines · Christian König
Quoted 16 lines
Quoted 16 lines

[SNIP]
+static void
+drm_syncobj_find_signal_pt_for_wait_pt(struct
drm_syncobj *syncobj,
+                           struct drm_syncobj_wait_pt
+*wait_pt) {
That whole approach still looks horrible complicated to me.
It's already very close to what you said before.

Especially the separation of signal and wait pt is
completely unnecessary as far as I can see.
When a wait pt is requested we just need to search for
the signal point which it will trigger.
Yeah, I tried this, but when I implement cpu wait ioctl on
specific point, we need a advanced wait pt fence,
otherwise, we could still need old syncobj cb.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

Why? I mean you just need to call drm_syncobj_find_fence() and

when

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 2 lines

that one returns NULL you use wait_event_*() to wait for a signal point >= your wait point to appear and try again.

e.g. when there are 3 syncobjs(A,B,C) to wait, all syncobjABC have no fence yet, as you said, during drm_syncobj_find_fence(A) is working on wait_event,

syncobjB

Quoted 59 lines
Quoted 59 lines
Quoted 42 lines
Quoted 34 lines
Quoted 23 lines
Quoted 23 lines
Quoted 11 lines
Quoted 11 lines
Quoted 11 lines
Quoted 3 lines

and syncobjC could already be signaled, then we don't know which one is first signaled, which is need when wait ioctl returns.

I don't really see a problem with that. When you wait for the first one you need to wait for A,B,C at the same time anyway.

So what you do is to register a fence callback on the fences you already have and for the syncobj which doesn't yet have a fence you make sure that they wake up your thread when they get one.

Quoted 8 lines
Quoted 4 lines
Quoted 3 lines

So essentially exactly what
drm_syncobj_fence_get_or_add_callback()
already does today.

So do you mean we need still use old syncobj CB for that?

Yes, as far as I can see it should work.

Quoted 1 line

    Advanced wait pt is bad?

Well it isn't bad, I just don't see any advantage in it.

The advantage is to replace old syncobj cb.

Quoted 2 lines

The existing mechanism
should already be able to handle that.

Quoted 8 lines
Quoted 2 lines

I thought more a bit, we don't that mechanism at all, if use advanced wait

pt, we can easily use fence array to achieve it for wait ioctl, we should use kernel existing feature as much as possible, not invent another, shouldn't we?
I remember  you said  it before.

Yeah, but the syncobj cb is an existing feature.

This is obviously a workaround when doing for wait ioctl, Do you see it used in other place?

Quoted 7 lines
Quoted 2 lines

And I absolutely don't see a
need to modify that and replace it with something far more complex.

The wait ioctl is simplified much more by fence array, not complex, and we just need  to allocate a wait pt.  If keeping old syncobj cb workaround, all wait pt logic still is there, just save allocation and wait pt handling, in fact, which part isn't complex at all. But compare with ugly syncobj cb, which is simpler.

Quoted 6 lines

I strongly disagree on that. You just need to extend the syncobj cb with the sequence number and you are done.

We could clean that up in the long term by adding some wait_multi event macro, but for now just adding the sequence number should do the trick.

Quote from Daniel Vetter comment when v1, "

Specifically for this stuff here having unified future fence semantics will allow drivers to do clever stuff with them.

"
I think the advanced wait pt is a similar concept as 'future fence' what Daniel Vetter said before, which obviously a right direction.

Quoted 6 lines
Quoted 2 lines

Anyway, I will change the patch as you like if no other comment, so that the patch can pass soon.

When I try to remove wait pt future fence, I encounter another problem, drm_syncobj_find_fence cannot get a fence if signal pt already is collected as garbage, then CS will report error, any idea for that?

Quoted 4 lines

Well when the signal pt is already garbage collected you know that it is already signaled. So you can just return a dummy fence.

I actually thought that this was the intention of the dummy fence rename :)

In fact, this is one of future fence functionality, then why not we unify them? Daniel Vetter also commented that before v1. Future fence can unify both your dummy fence and old syncobj cb.
If you have no objection, I will prepare a patch to let us see how simple wait_ioctl using fence array.

Thanks,
David Zhou

Quoted 49 lines

Christian.

Quoted 46 lines

I still think the future fence is right thing, Could you give futher thought on it again? Otherwise, we could need various workarounds.

Thanks,
David Zhou

Quoted 40 lines

Thanks,
David Zhou

Quoted 36 lines

Regards,
Christian.

Quoted 31 lines

Thanks,
David Zhou

Quoted 28 lines

Regards,
Christian.

Quoted 25 lines

Thanks,
David Zhou

Quoted 23 lines

Christian.

Quoted 21 lines

Thanks,
David Zhou

Quoted 15 lines

Regards,
Christian.

Quoted 12 lines

Back to my implementation, it already fixes all your concerns before, and can be able to easily used in wait_ioctl. When you feel that is complicated, I guess that is because we merged all logic to that and much clean up in one patch. In fact, it already is very simple, timeline_init/fini, create signal/wait_pt, find signal_pt for wait_pt, garbage collection, just them.

Thanks,
David Zhou

Quoted 2 lines

Regards,
Christian.

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

A message that is not in the archive (sent off-list or not yet received): BY1PR12MB050245C96DE62338FB349AE5B4190-PicGAnIBOobrCwm+z9iKNgdYzm3356FpvxpqHgZTriW3zl9H0oFU5g@public.gmane.org

#35 lore raw

Am 14.09.2018 um 09:46 schrieb Zhou, David(ChunMing):

Quoted 141 lines
Quoted 141 lines

-----Original Message-----
From: Koenig, Christian
Sent: Friday, September 14, 2018 3:27 PM
To: Zhou, David(ChunMing) <[email protected]>; Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]; Daniel Vetter <[email protected]>
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 14.09.2018 um 05:59 schrieb zhoucm1:

Quoted 56 lines · On 2018年09月14日 11:14, zhoucm1 wrote:
Quoted 54 lines · Christian König
Quoted 52 lines

Am 13.09.2018 um 11:35 schrieb Zhou, David(ChunMing):

Quoted 51 lines
Quoted 51 lines

-----Original Message-----
From: Koenig, Christian
Sent: Thursday, September 13, 2018 5:20 PM
To: Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 11:11 schrieb Zhou, David(ChunMing):

Quoted 41 lines
Quoted 41 lines

-----Original Message-----
From: Christian König <[email protected]>
Sent: Thursday, September 13, 2018 4:50 PM
To: Zhou, David(ChunMing) <[email protected]>; Koenig, Christian <[email protected]>; [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 09:43 schrieb Zhou, David(ChunMing):

Quoted 29 lines
Quoted 29 lines

-----Original Message-----
From: Koenig, Christian
Sent: Thursday, September 13, 2018 2:56 PM
To: Zhou, David(ChunMing) <[email protected]>; Zhou, David(ChunMing) <[email protected]>; dri- [email protected]
Cc: Dave Airlie <[email protected]>; Rakos, Daniel <[email protected]>; [email protected]
Subject: Re: [PATCH 1/3] [RFC]drm: add syncobj timeline support v4

Am 13.09.2018 um 04:15 schrieb zhoucm1:

Quoted 17 lines · Christian König
Quoted 16 lines

[SNIP]
+static void
+drm_syncobj_find_signal_pt_for_wait_pt(struct
drm_syncobj *syncobj,
+                           struct drm_syncobj_wait_pt
+*wait_pt) {
That whole approach still looks horrible complicated to me.
It's already very close to what you said before.

Especially the separation of signal and wait pt is
completely unnecessary as far as I can see.
When a wait pt is requested we just need to search for
the signal point which it will trigger.
Yeah, I tried this, but when I implement cpu wait ioctl on
specific point, we need a advanced wait pt fence,
otherwise, we could still need old syncobj cb.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

Why? I mean you just need to call drm_syncobj_find_fence()
and

when

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 2 lines

that one returns NULL you use wait_event_*() to wait for a
signal point >= your wait point to appear and try again.

e.g. when there are 3 syncobjs(A,B,C) to wait, all syncobjABC have no fence yet, as you said, during drm_syncobj_find_fence(A) is working on wait_event,

syncobjB

Quoted 63 lines
Quoted 52 lines
Quoted 41 lines
Quoted 34 lines
Quoted 23 lines
Quoted 23 lines
Quoted 11 lines
Quoted 11 lines
Quoted 11 lines
Quoted 3 lines

and syncobjC could already be signaled, then we don't know which one is first signaled, which is need when wait ioctl returns.

I don't really see a problem with that. When you wait for the first one you need to wait for A,B,C at the same time anyway.

So what you do is to register a fence callback on the fences you already have and for the syncobj which doesn't yet have a fence you make sure that they wake up your thread when they get one.

Quoted 8 lines
Quoted 4 lines
Quoted 3 lines

So essentially exactly what
drm_syncobj_fence_get_or_add_callback()
already does today.

So do you mean we need still use old syncobj CB for that?

Yes, as far as I can see it should work.

Quoted 1 line

    Advanced wait pt is bad?

Well it isn't bad, I just don't see any advantage in it.

The advantage is to replace old syncobj cb.

Quoted 2 lines

The existing mechanism
should already be able to handle that.

Quoted 8 lines
Quoted 2 lines

I thought more a bit, we don't that mechanism at all, if use advanced wait

pt, we can easily use fence array to achieve it for wait ioctl, we should use kernel existing feature as much as possible, not invent another, shouldn't we?
I remember  you said  it before.

Yeah, but the syncobj cb is an existing feature.

This is obviously a workaround when doing for wait ioctl, Do you see it used in other place?

Quoted 7 lines
Quoted 2 lines

And I absolutely don't see a
need to modify that and replace it with something far more complex.

The wait ioctl is simplified much more by fence array, not complex, and we just need  to allocate a wait pt.  If keeping old syncobj cb workaround, all wait pt logic still is there, just save allocation and wait pt handling, in fact, which part isn't complex at all. But compare with ugly syncobj cb, which is simpler.

Quoted 6 lines

I strongly disagree on that. You just need to extend the syncobj cb with the sequence number and you are done.

We could clean that up in the long term by adding some wait_multi event macro, but for now just adding the sequence number should do the trick.

Quote from Daniel Vetter comment when v1, "

Specifically for this stuff here having unified future fence semantics will allow drivers to do clever stuff with them.

Quoted 7 lines

"
I think the advanced wait pt is a similar concept as 'future fence' what Daniel Vetter said before, which obviously a right direction.

Anyway, I will change the patch as you like if no other comment, so that the patch can pass soon.

When I try to remove wait pt future fence, I encounter another problem, drm_syncobj_find_fence cannot get a fence if signal pt already is collected as garbage, then CS will report error, any idea for that?

Quoted 7 lines
Quoted 4 lines

Well when the signal pt is already garbage collected you know that it is already signaled. So you can just return a dummy fence.

I actually thought that this was the intention of the dummy fence rename :)

In fact, this is one of future fence functionality, then why not we unify them? Daniel Vetter also commented that before v1. Future fence can unify both your dummy fence and old syncobj cb.
If you have no objection, I will prepare a patch to let us see how simple wait_ioctl using fence array.

I do have objections, please keep the existing wait as it is.

Regards,
Christian.

Quoted 46 lines

Thanks,
David Zhou

Quoted 42 lines

Christian.

Quoted 40 lines

I still think the future fence is right thing, Could you give futher thought on it again? Otherwise, we could need various workarounds.

Thanks,
David Zhou

Quoted 35 lines

Thanks,
David Zhou

Quoted 33 lines

Regards,
Christian.

Quoted 30 lines

Thanks,
David Zhou

Quoted 28 lines

Regards,
Christian.

Quoted 25 lines

Thanks,
David Zhou

Quoted 23 lines

Christian.

Quoted 21 lines

Thanks,
David Zhou

Quoted 15 lines

Regards,
Christian.

Quoted 12 lines

Back to my implementation, it already fixes all your concerns before, and can be able to easily used in wait_ioctl. When you feel that is complicated, I guess that is because we merged all logic to that and much clean up in one patch. In fact, it already is very simple, timeline_init/fini, create signal/wait_pt, find signal_pt for wait_pt, garbage collection, just them.

Thanks,
David Zhou

Quoted 2 lines

Regards,
Christian.

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx