The IP checksum algorithm yields identical results regardless of whether
arithmetic little-endian or big-endian, but in practice OVS only passes in
big-endian data, so it seems reasonable to annotate these functions that
way.
Occasionally the checksum test on "make check" would fail. This commit
fixes the problem, which was that the partial checksum may need more than
one reduction step to obtain a final checksum. Now running checksum tests
in a continuous loop yields no failures.
