it was actually double the full size in current kernels where the reduction is null, so overran the mmapped() buffer).