TCP NewReno and single retransmit

[Marcelo Ricardo Leitner <mleitner@redhat.com>]

Hi,

We have a report from a customer saying that on a very calm connection, like 
having only a single data packet within some minutes, if this packet gets to 
be re-transmitted, retrans_stamp is only cleared when the next acked packet is 
received. But this may make we abort the connection too soon if this next 
packet also gets lost, because the reference for the initial loss is still for 
a big while ago..

                    local-machine              remote-machine
                         |                           |
          send#1---->(*1)|--------> data#1 --------->|
                   |     |                           |
                  RTO    :                           :
                   |     |                           |
                  ---(*2)|----> data#1(retrans) ---->|
                   | (*3)|<---------- ACK <----------|
                   |     |                           |
                   |     :                           :
                   |     :                           :
                   |     :                           :
                 16 minutes (or more)                :
                   |     :                           :
                   |     :                           :
                   |     :                           :
                   |     |                           |
          send#2---->(*4)|--------> data#2 --------->|
                   |     |                           |
                  RTO    :                           :
                   |     |                           |
                  ---(*5)|----> data#2(retrans) ---->|
                   |     |                           |
                   |     |                           |
                 RTO*2   :                           :
                   |     |                           |
                   |     |                           |
       ETIMEDOUT<----(*6)|                           |
    (diagram is not mine)

ETIMEDOUT happens way too early, because that's based on (*2) stamp.

Question is, can't we really clear retrans_stamp on step (*3)? Like with:

@@ -2382,31 +2382,32 @@ static inline bool tcp_may_undo(const struct tcp_sock *tp)
  static bool tcp_try_undo_recovery(struct sock *sk)
  {
         struct tcp_sock *tp = tcp_sk(sk);

         if (tcp_may_undo(tp)) {
                 int mib_idx;

                 /* Happy end! We did not retransmit anything
                  * or our original transmission succeeded.
                  */
                 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? 
"loss" : "retrans");
                 tcp_undo_cwnd_reduction(sk, false);
                 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
                         mib_idx = LINUX_MIB_TCPLOSSUNDO;
                 else
                         mib_idx = LINUX_MIB_TCPFULLUNDO;

                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
         }
         if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
                 /* Hold old state until something *above* high_seq
                  * is ACKed. For Reno it is MUST to prevent false
                  * fast retransmits (RFC2582). SACK TCP is safe. */
                 tcp_moderate_cwnd(tp);
+               tp->retrans_stamp = 0;
                 return true;
         }
         tcp_set_ca_state(sk, TCP_CA_Open);
         return false;
  }

We would still hold state, at least part of it.. WDYT?

Thanks,
Marcelo