Re: [PATCH net-next v1] ipvs: add consistent source hashing scheduling

[Vincent Bernat <vincent@bernat.im>]

 ❦  2 avril 2018 10:33 -0700, Eric Dumazet <eric.dumazet@gmail.com> :

>> +static inline u32
>> +ip_vs_csh_permutation(struct ip_vs_dest *d, int j)
>> +{
>> +	u32 offset, skip;
>> +	__be32 addr_fold = d->addr.ip;
>> +
>> +#ifdef CONFIG_IP_VS_IPV6
>> +	if (d->af == AF_INET6)
>> +		addr_fold = d->addr.ip6[0]^d->addr.ip6[1]^
>> +			d->addr.ip6[2]^d->addr.ip6[3];
>> +#endif
>> +	addr_fold = ntohl(addr_fold) + ntohs(d->port);
>> +	offset = hash_32(addr_fold, 32) % IP_VS_CSH_TAB_SIZE;
>> +	skip = (hash_32(addr_fold + 1, 32) % (IP_VS_CSH_TAB_SIZE - 1)) + 1;
>> +	return (offset + j * skip) % IP_VS_CSH_TAB_SIZE;
>> +}
>> +
>
> This does not look very strong to me, particularly the IPv6 folding
>
> I would rather use __ipv6_addr_jhash() instead of ipv6_addr_hash(),
> even if it is hard coded ;)

I can switch to ipv6_addr_hash(). However, switching to
__ipv6_addr_jhash seems useless as I would need to hardcode the initial
value: people use source hashing to get the same result from one host to
another. Am I missing something?
-- 
Each module should do one thing well.
            - The Elements of Programming Style (Kernighan & Plauger)