drm/amdgpu: use bulk moves for efficient VM LRU handlingvm_bulk_move

WIP

Signed-off-by: Christian König <christian.koenig@amd.com>

2 files changed, 18 insertions, 8 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index b3639533de4a..20d6aef1fbbb 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -197,9 +197,9 @@ static int amdgpu_vm_move_bos(int id, void *ptr, void *data)
 	if (bo->tbo.resv != vm->root.base.bo->tbo.resv)
 		return 0;
 
-	ttm_bo_move_to_lru_tail(&bo->tbo, NULL);
+	ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move);
 	if (bo->shadow)
-		ttm_bo_move_to_lru_tail(&bo->shadow->tbo, NULL);
+		ttm_bo_move_to_lru_tail(&bo->shadow->tbo, &vm->lru_bulk_move);
 
 	return 0;
 }
@@ -214,6 +214,7 @@ static int amdgpu_vm_move_bos(int id, void *ptr, void *data)
  * Make sure that all per VM PDs/PTs are grouped together on their LRU.
  */
 static void amdgpu_vm_move_levels(struct amdgpu_device *adev,
+				  struct amdgpu_vm *vm,
 				  struct amdgpu_vm_pt *parent,
 				  unsigned level)
 {
@@ -222,9 +223,10 @@ static void amdgpu_vm_move_levels(struct amdgpu_device *adev,
 
 	if (bo) {
 		if (bo->parent)
-			ttm_bo_move_to_lru_tail(&bo->tbo, NULL);
+			ttm_bo_move_to_lru_tail(&bo->tbo, &vm->lru_bulk_move);
 		if (bo->shadow)
-			ttm_bo_move_to_lru_tail(&bo->shadow->tbo, NULL);
+			ttm_bo_move_to_lru_tail(&bo->shadow->tbo,
+						&vm->lru_bulk_move);
 	}
 
 	if (!parent->entries)
@@ -232,7 +234,7 @@ static void amdgpu_vm_move_levels(struct amdgpu_device *adev,
 
 	num_entries = amdgpu_vm_num_entries(adev, level);
 	for (i = 0; i < num_entries; i++)
-		amdgpu_vm_move_levels(adev, &parent->entries[i],
+		amdgpu_vm_move_levels(adev, vm, &parent->entries[i],
 				      level + 1);
 }
 
@@ -290,8 +292,14 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct drm_file *filp,
 	}
 	spin_unlock(&vm->status_lock);
 
-	if (!validated)
+	if (!validated) {
+		spin_lock(&glob->lru_lock);
+		ttm_bo_bulk_move_lru_tail(&vm->lru_bulk_move);
+		spin_unlock(&glob->lru_lock);
 		return 0;
+	}
+
+	memset(&vm->lru_bulk_move, 0, sizeof(vm->lru_bulk_move));
 
 	spin_lock(&glob->lru_lock);
 
@@ -299,7 +307,7 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct drm_file *filp,
 	idr_for_each(&filp->object_idr, amdgpu_vm_move_bos, vm);
 	spin_unlock(&filp->table_lock);
 
-	amdgpu_vm_move_levels(adev, &vm->root,
+	amdgpu_vm_move_levels(adev, vm, &vm->root,
 			      adev->vm_manager.root_level);
 
 	spin_unlock(&glob->lru_lock);
@@ -512,7 +520,7 @@ static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev,
 			entry->base.bo = pt;
 			list_add_tail(&entry->base.bo_list, &pt->va);
 			spin_lock(&vm->status_lock);
-			list_add(&entry->base.vm_status, &vm->relocated);
+			list_add(&entry->base.vm_status, &vm->evicted);
 			spin_unlock(&vm->status_lock);
 		}
 
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
index a43a493f5c08..4c4346b80b8e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
@@ -206,6 +206,8 @@ struct amdgpu_vm {
 
 	/* Limit non-retry fault storms */
 	unsigned int		fault_credit;
+
+	struct ttm_lru_bulk_move	lru_bulk_move;
 };
 
 struct amdgpu_vm_manager {