drm/amdgpu: use only one reservation object for each VM v2

Reduces the locking and fencing overhead.

v2: add comment why we need the duplicates list in the GEM op.

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c

@@ -455,11 +455,12 @@ static void amdgpu_gem_va_update_vm(struct amdgpu_device *adev,
 	struct ttm_validate_buffer tv, *entry;
 	struct amdgpu_bo_list_entry *vm_bos;
 	struct ww_acquire_ctx ticket;
-	struct list_head list;
+	struct list_head list, duplicates;
 	unsigned domain;
 	int r;
 
 	INIT_LIST_HEAD(&list);
+	INIT_LIST_HEAD(&duplicates);
 
 	tv.bo = &bo_va->bo->tbo;
 	tv.shared = true;
@@ -469,7 +470,8 @@ static void amdgpu_gem_va_update_vm(struct amdgpu_device *adev,
 	if (!vm_bos)
 		return;
 
-	r = ttm_eu_reserve_buffers(&ticket, &list, true, NULL);
+	/* Provide duplicates to avoid -EALREADY */
+	r = ttm_eu_reserve_buffers(&ticket, &list, true, &duplicates);
 	if (r)
 		goto error_free;
 

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c

@@ -685,31 +685,6 @@ static int amdgpu_vm_update_ptes(struct amdgpu_device *adev,
 	return 0;
 }
 
-/**
- * amdgpu_vm_fence_pts - fence page tables after an update
- *
- * @vm: requested vm
- * @start: start of GPU address range
- * @end: end of GPU address range
- * @fence: fence to use
- *
- * Fence the page tables in the range @start - @end (cayman+).
- *
- * Global and local mutex must be locked!
- */
-static void amdgpu_vm_fence_pts(struct amdgpu_vm *vm,
-				uint64_t start, uint64_t end,
-				struct fence *fence)
-{
-	unsigned i;
-
-	start >>= amdgpu_vm_block_size;
-	end >>= amdgpu_vm_block_size;
-
-	for (i = start; i <= end; ++i)
-		amdgpu_bo_fence(vm->page_tables[i].bo, fence, true);
-}
-
 /**
  * amdgpu_vm_bo_update_mapping - update a mapping in the vm page table
  *
@@ -813,8 +788,7 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
 	if (r)
 		goto error_free;
 
-	amdgpu_vm_fence_pts(vm, mapping->it.start,
-			    mapping->it.last + 1, f);
+	amdgpu_bo_fence(vm->page_directory, f, true);
 	if (fence) {
 		fence_put(*fence);
 		*fence = fence_get(f);
@@ -1089,6 +1063,7 @@ int amdgpu_vm_bo_map(struct amdgpu_device *adev,
 
 	/* walk over the address space and allocate the page tables */
 	for (pt_idx = saddr; pt_idx <= eaddr; ++pt_idx) {
+		struct reservation_object *resv = vm->page_directory->tbo.resv;
 		struct amdgpu_bo *pt;
 
 		if (vm->page_tables[pt_idx].bo)
@@ -1097,11 +1072,13 @@ int amdgpu_vm_bo_map(struct amdgpu_device *adev,
 		/* drop mutex to allocate and clear page table */
 		mutex_unlock(&vm->mutex);
 
+		ww_mutex_lock(&resv->lock, NULL);
 		r = amdgpu_bo_create(adev, AMDGPU_VM_PTE_COUNT * 8,
 				     AMDGPU_GPU_PAGE_SIZE, true,
 				     AMDGPU_GEM_DOMAIN_VRAM,
 				     AMDGPU_GEM_CREATE_NO_CPU_ACCESS,
-				     NULL, NULL, &pt);
+				     NULL, resv, &pt);
+		ww_mutex_unlock(&resv->lock);
 		if (r)
 			goto error_free;