TCP socket saves the minimum required header length in tcp_header_len
of struct tcp_sock, and later the value is used in __tcp_fast_path_on()
to generate a part of TCP header in tcp_sock(sk)->pred_flags.

In tcp_rcv_established(), if the incoming packet has the same pattern
with pred_flags, we enter the fast path and skip full option parsing.

The MD5 option is parsed in tcp_v[46]_rcv(), so we need not parse it
again later in tcp_rcv_established() unless other options exist.  We
add TCPOLEN_MD5SIG_ALIGNED to tcp_header_len in two paths to avoid the
slow path.

For passive open connections with MD5, we add TCPOLEN_MD5SIG_ALIGNED
to tcp_header_len in tcp_create_openreq_child() after 3WHS.

On the other hand, we do it in tcp_connect_init() for active open
connections.  However, the value is overwritten while processing
SYN+ACK or crossed SYN in tcp_rcv_synsent_state_process().

These two cases will have the wrong value in pred_flags and never go
into the fast path.

We could update tcp_header_len in tcp_rcv_synsent_state_process(), but
a test with slightly modified netperf which uses MD5 for each flow shows
that the slow path is actually a bit faster than the fast path.

  On c5.4xlarge EC2 instance (16 vCPU, 32 GiB mem)

  $ for i in {1..10}; do
  ./super_netperf $(nproc) -H localhost -l 10 -- -m 256 -M 256;
  done

  Avg of 10
  * 36e68ead  : 10.376 Gbps
  * all fast path : 10.374 Gbps (patch v2, See Link)
  * all slow path : 10.394 Gbps

The header prediction is not worth adding complexity for MD5, so let's
disable it for MD5.

Link: https://lore.kernel.org/netdev/20230803042214.38309-1-kuniyu@amazon.com/Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20230803224552.69398-2-kuniyu@amazon.comSigned-off-by: Jakub Kicinski <kuba@kernel.org>