1 files changed, 92 insertions, 11 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
index 7cb711fc1ee2..4fffb2539331 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
@@ -32,6 +32,11 @@
 #include "amdgpu.h"
 #include "amdgpu_trace.h"
 
+struct amdgpu_sync_entry {
+	struct hlist_node	node;
+	struct fence		*fence;
+};
+
 /**
  * amdgpu_sync_create - zero init sync object
  *
@@ -49,9 +54,33 @@ void amdgpu_sync_create(struct amdgpu_sync *sync)
 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
 		sync->sync_to[i] = NULL;
 
+	hash_init(sync->fences);
 	sync->last_vm_update = NULL;
 }
 
+static bool amdgpu_sync_same_dev(struct amdgpu_device *adev, struct fence *f)
+{
+	struct amdgpu_fence *a_fence = to_amdgpu_fence(f);
+	struct amd_sched_fence *s_fence = to_amd_sched_fence(f);
+
+	if (a_fence)
+		return a_fence->ring->adev == adev;
+	if (s_fence)
+		return (struct amdgpu_device *)s_fence->scheduler->priv == adev;
+	return false;
+}
+
+static bool amdgpu_sync_test_owner(struct fence *f, void *owner)
+{
+	struct amdgpu_fence *a_fence = to_amdgpu_fence(f);
+	struct amd_sched_fence *s_fence = to_amd_sched_fence(f);
+	if (s_fence)
+		return s_fence->owner == owner;
+	if (a_fence)
+		return a_fence->owner == owner;
+	return false;
+}
+
 /**
  * amdgpu_sync_fence - remember to sync to this fence
  *
@@ -62,28 +91,54 @@ void amdgpu_sync_create(struct amdgpu_sync *sync)
 int amdgpu_sync_fence(struct amdgpu_device *adev, struct amdgpu_sync *sync,
 		      struct fence *f)
 {
+	struct amdgpu_sync_entry *e;
 	struct amdgpu_fence *fence;
 	struct amdgpu_fence *other;
+	struct fence *tmp, *later;
 
 	if (!f)
 		return 0;
 
+	if (amdgpu_sync_same_dev(adev, f) &&
+	    amdgpu_sync_test_owner(f, AMDGPU_FENCE_OWNER_VM)) {
+		if (sync->last_vm_update) {
+			tmp = sync->last_vm_update;
+			BUG_ON(f->context != tmp->context);
+			later = (f->seqno - tmp->seqno <= INT_MAX) ? f : tmp;
+			sync->last_vm_update = fence_get(later);
+			fence_put(tmp);
+		} else
+			sync->last_vm_update = fence_get(f);
+	}
+
 	fence = to_amdgpu_fence(f);
-	if (!fence || fence->ring->adev != adev)
-		return fence_wait(f, true);
+	if (!fence || fence->ring->adev != adev) {
+		hash_for_each_possible(sync->fences, e, node, f->context) {
+			struct fence *new;
+			if (unlikely(e->fence->context != f->context))
+				continue;
+			new = fence_get(fence_later(e->fence, f));
+			if (new) {
+				fence_put(e->fence);
+				e->fence = new;
+			}
+			return 0;
+		}
+
+		e = kmalloc(sizeof(struct amdgpu_sync_entry), GFP_KERNEL);
+		if (!e)
+			return -ENOMEM;
+
+		hash_add(sync->fences, &e->node, f->context);
+		e->fence = fence_get(f);
+		return 0;
+	}
 
 	other = sync->sync_to[fence->ring->idx];
 	sync->sync_to[fence->ring->idx] = amdgpu_fence_ref(
 		amdgpu_fence_later(fence, other));
 	amdgpu_fence_unref(&other);
 
-	if (fence->owner == AMDGPU_FENCE_OWNER_VM) {
-		other = sync->last_vm_update;
-		sync->last_vm_update = amdgpu_fence_ref(
-			amdgpu_fence_later(fence, other));
-		amdgpu_fence_unref(&other);
-	}
-
 	return 0;
 }
 
@@ -147,6 +202,24 @@ int amdgpu_sync_resv(struct amdgpu_device *adev,
 	return r;
 }
 
+int amdgpu_sync_wait(struct amdgpu_sync *sync)
+{
+	struct amdgpu_sync_entry *e;
+	struct hlist_node *tmp;
+	int i, r;
+
+	hash_for_each_safe(sync->fences, i, tmp, e, node) {
+		r = fence_wait(e->fence, false);
+		if (r)
+			return r;
+
+		hash_del(&e->node);
+		fence_put(e->fence);
+		kfree(e);
+	}
+	return 0;
+}
+
 /**
  * amdgpu_sync_rings - sync ring to all registered fences
  *
@@ -234,15 +307,23 @@ int amdgpu_sync_rings(struct amdgpu_sync *sync,
  */
 void amdgpu_sync_free(struct amdgpu_device *adev,
 		      struct amdgpu_sync *sync,
-		      struct amdgpu_fence *fence)
+		      struct fence *fence)
 {
+	struct amdgpu_sync_entry *e;
+	struct hlist_node *tmp;
 	unsigned i;
 
+	hash_for_each_safe(sync->fences, i, tmp, e, node) {
+		hash_del(&e->node);
+		fence_put(e->fence);
+		kfree(e);
+	}
+
 	for (i = 0; i < AMDGPU_NUM_SYNCS; ++i)
 		amdgpu_semaphore_free(adev, &sync->semaphores[i], fence);
 
 	for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
 		amdgpu_fence_unref(&sync->sync_to[i]);
 
-	amdgpu_fence_unref(&sync->last_vm_update);
+	fence_put(sync->last_vm_update);
 }