vdpa/mlx5: Avoid destroying MR on empty iotlb

The current code treats an empty iotlb provdied in set_map() as a
special case and destroy the memory region object. This must not be done
since the virtqueue objects reference this MR. Doing so will cause the
driver unload to emit errors and log timeouts caused by the firmware
complaining on busy resources.

This patch treats an empty iotlb as any other change of mapping. In this
case, mlx5_vdpa_create_mr() will fail and the entire set_map() call to
fail.

This issue has not been encountered before but was seen to occur in a
non-official version of qemu. Since qemu is a userspace program, the
driver must protect against such case.

Fixes: 94abbccd ("vdpa/mlx5: Add shared memory registration code")
Signed-off-by: Eli Cohen <elic@nvidia.com>
Link: https://lore.kernel.org/r/20210811053713.66658-1-elic@nvidia.comSigned-off-by: Michael S. Tsirkin <mst@redhat.com>

drivers/vdpa/mlx5/core/mr.c

@@ -512,11 +512,6 @@ void mlx5_vdpa_destroy_mr(struct mlx5_vdpa_dev *mvdev)
 	mutex_unlock(&mr->mkey_mtx);
 }
 
-static bool map_empty(struct vhost_iotlb *iotlb)
-{
-	return !vhost_iotlb_itree_first(iotlb, 0, U64_MAX);
-}
-
 int mlx5_vdpa_handle_set_map(struct mlx5_vdpa_dev *mvdev, struct vhost_iotlb *iotlb,
 			     bool *change_map)
 {
@@ -524,10 +519,6 @@ int mlx5_vdpa_handle_set_map(struct mlx5_vdpa_dev *mvdev, struct vhost_iotlb *io
 	int err = 0;
 
 	*change_map = false;
-	if (map_empty(iotlb)) {
-		mlx5_vdpa_destroy_mr(mvdev);
-		return 0;
-	}
 	mutex_lock(&mr->mkey_mtx);
 	if (mr->initialized) {
 		mlx5_vdpa_info(mvdev, "memory map update\n");