gk: fix checksum issue

[mengxiang0811]

Some examples that implement the RFC1624 [Eqn. 4] use the network order:
https://elixir.bootlin.com/linux/latest/source/net/ipv4/netfilter/ipt_ECN.c#L38

This patch removes the byte-order convertion for checksum.

[AltraMayor]

Notice that the given example actually implements equation 3 of RFC1624. Moreover, one's complement sum and one's complement (i.e. C's operator ~) preserve endianness, that is, if all inputs are little-endian, the result is little-endian. The same is true for big-endian. This endianness preservation is even independent of the endianness of the host! That is why the given example works.

We have assumed that the two's complement subtraction, which is currently being used, is the same as "subtracting complements with borrow" under one's complement as required in RFC1624. While these two operations often come up with the same result, they are often not equal too. And this is the bug we are facing. In addition, the two's complement subtraction is not endianness preserving!

The solution is to fix the current code using the following functions:

/*
 * One's complement sum.
 *
 * Notice that if @a and @b are little-endian, the result is also
 * little-endian. The same is true for big-endian. In order words,
 * this function preserves endianness.
 *
 * The endianness preservation is independent of the endianness of the host.
 */
static inline uint16_t
onec_add(uint16_t a, uint16_t b)
{
	uint16_t res = a + b;
	return res + (res < b);
}

/*
 * The result has the same endianness of the inputs as long as
 * all inputs have the same endianness.
 * The endianness preservation is independent of the endianness of the host.
 */
static inline uint16_t
new_ip_csum(uint16_t old_csum, uint16_t old16, uint16_t new16)
{
	/* According to RFC1624 [Eqn. 3]. */
	return ~onec_add(onec_add(~old_csum, ~old16), new16);
}

Don't forget to update the description of the patch with a better text, for example, explain when the bug happens.

[mengxiang0811]

Thanks so much @AltraMayor ! This rebase is ready for another review.

[AltraMayor]

The code is ready for merge, but please drop the whole parenthesis in this section of the patch comment: the two's complement (i.e. C's operator ~). The operator ~ produces one's complements.

[mengxiang0811]

This rebase is ready for another review.