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: 605ad7f184
("tcp: refine TSO autosizing")
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit is contained in:
parent
e9647d1e74
commit
6c09fa09d4
|
@ -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 */
|
||||
|
|
Loading…
Reference in a new issue