tcp: Reduce chance of collisions in inet6_hashfn().

For both IPv4 and IPv6 incoming TCP connections are tracked in a hash
table with a hash over the source & destination addresses and ports.
However, the IPv6 hash is insufficient and can lead to a high rate of
collisions.

The IPv6 hash used an XOR to fit everything into the 96 bits for the
fast jenkins hash, meaning it is possible for an external entity to
ensure the hash collides, thus falling back to a linear search in the
bucket, which is slow.

We take the approach of hash the full length of IPv6 address in
__ipv6_addr_jhash() so that all users can benefit from a more secure
version.

While this may look like it adds overhead, the reality of modern CPUs
means that this is unmeasurable in real world scenarios.

In simulating with llvm-mca, the increase in cycles for the hashing
code was ~16 cycles on Skylake (from a base of ~155), and an extra ~9
on Nehalem (base of ~173).

In commit dd6d2910 ("netfilter: conntrack: switch to siphash")
netfilter switched from a jenkins hash to a siphash, but even the faster
hsiphash is a more significant overhead (~20-30%) in some preliminary
testing.  So, in this patch, we keep to the more conservative approach to
ensure we don't add much overhead per SYN.

In testing, this results in a consistently even spread across the
connection buckets.  In both testing and real-world scenarios, we have
not found any measurable performance impact.

Fixes: 08dcdbf6 ("ipv6: use a stronger hash for tcp")
Signed-off-by: Stewart Smith <trawets@amazon.com>
Signed-off-by: Samuel Mendoza-Jonas <samjonas@amazon.com>
Suggested-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20230721222410.17914-1-kuniyu@amazon.com


Signed-off-by: Jakub Kicinski <kuba@kernel.org>