tcp: align tcp_xmit_size_goal() on tcp_tso_autosize()

With some mss values, it is possible tcp_xmit_size_goal() puts
one segment more in TSO packet than tcp_tso_autosize().

We send then one TSO packet followed by one single MSS.

It is not a serious bug, but we can do slightly better, especially
for drivers using netif_set_gso_max_size() to lower gso_max_size.

Using same formula avoids these corner cases and makes
tcp_xmit_size_goal() a bit faster.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Fixes: 605ad7f184b6 ("tcp: refine TSO autosizing")
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 9d72a0f..995a225 100644 (file)
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -835,17 +835,13 @@ static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
                                       int large_allowed)
 {
        struct tcp_sock *tp = tcp_sk(sk);
-       u32 new_size_goal, size_goal, hlen;
+       u32 new_size_goal, size_goal;
 
        if (!large_allowed || !sk_can_gso(sk))
                return mss_now;
 
-       /* Maybe we should/could use sk->sk_prot->max_header here ? */
-       hlen = inet_csk(sk)->icsk_af_ops->net_header_len +
-              inet_csk(sk)->icsk_ext_hdr_len +
-              tp->tcp_header_len;
-
-       new_size_goal = sk->sk_gso_max_size - 1 - hlen;
+       /* Note : tcp_tso_autosize() will eventually split this later */
+       new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
        new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
 
        /* We try hard to avoid divides here */