To avoid allocating ridiculous amounts of memory for corrupted files,
read the input in chunks limited to filesize or an arbitrary large
amount when that is not known (chosen to be 50M).