[TCP] FRTO: Improve interoperability with other undo_marker users
authorIlpo Järvinen <ilpo.jarvinen@helsinki.fi>
Thu, 20 Sep 2007 18:34:38 +0000 (11:34 -0700)
committerDavid S. Miller <davem@sunset.davemloft.net>
Wed, 10 Oct 2007 23:52:11 +0000 (16:52 -0700)
Basically this change enables it, previously other undo_marker
users were left with nothing. Reverse undo_marker logic
completely to get it set right in CA_Loss. On the other hand,
when spurious RTO is detected, clear it. Clearing might be too
heavy for some scenarios but seems safe enough starting point
for now and shouldn't have much effect except in majority of
cases (if in any).

By adding a new FLAG_ we avoid looping through write_queue when
RTO occurs.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/tcp_input.c

index 8fe754be8076708aee276a19ee6d183fae50145f..0feb10935be16bb5f675a194b0e3436de93b5d40 100644 (file)
@@ -104,6 +104,7 @@ int sysctl_tcp_abc __read_mostly;
 #define FLAG_ONLY_ORIG_SACKED  0x200 /* SACKs only non-rexmit sent before RTO */
 #define FLAG_SND_UNA_ADVANCED  0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
 #define FLAG_DSACKING_ACK      0x800 /* SACK blocks contained DSACK info */
+#define FLAG_NONHEAD_RETRANS_ACKED     0x1000 /* Non-head rexmitted data was ACKed */
 
 #define FLAG_ACKED             (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
 #define FLAG_NOT_DUP           (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
@@ -1594,6 +1595,8 @@ void tcp_enter_frto(struct sock *sk)
        tp->undo_retrans = 0;
 
        skb = tcp_write_queue_head(sk);
+       if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
+               tp->undo_marker = 0;
        if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
                TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
                tp->retrans_out -= tcp_skb_pcount(skb);
@@ -1643,6 +1646,8 @@ static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
                        /* ...enter this if branch just for the first segment */
                        flag |= FLAG_DATA_ACKED;
                } else {
+                       if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
+                               tp->undo_marker = 0;
                        TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
                }
 
@@ -1658,7 +1663,6 @@ static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
        tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
        tp->snd_cwnd_cnt = 0;
        tp->snd_cwnd_stamp = tcp_time_stamp;
-       tp->undo_marker = 0;
        tp->frto_counter = 0;
 
        tp->reordering = min_t(unsigned int, tp->reordering,
@@ -2584,20 +2588,6 @@ static int tcp_clean_rtx_queue(struct sock *sk, s32 *seq_rtt_p)
                        end_seq = scb->end_seq;
                }
 
-               /* Initial outgoing SYN's get put onto the write_queue
-                * just like anything else we transmit.  It is not
-                * true data, and if we misinform our callers that
-                * this ACK acks real data, we will erroneously exit
-                * connection startup slow start one packet too
-                * quickly.  This is severely frowned upon behavior.
-                */
-               if (!(scb->flags & TCPCB_FLAG_SYN)) {
-                       flag |= FLAG_DATA_ACKED;
-               } else {
-                       flag |= FLAG_SYN_ACKED;
-                       tp->retrans_stamp = 0;
-               }
-
                /* MTU probing checks */
                if (fully_acked && icsk->icsk_mtup.probe_size &&
                    !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
@@ -2610,6 +2600,9 @@ static int tcp_clean_rtx_queue(struct sock *sk, s32 *seq_rtt_p)
                                        tp->retrans_out -= packets_acked;
                                flag |= FLAG_RETRANS_DATA_ACKED;
                                seq_rtt = -1;
+                               if ((flag & FLAG_DATA_ACKED) ||
+                                   (packets_acked > 1))
+                                       flag |= FLAG_NONHEAD_RETRANS_ACKED;
                        } else if (seq_rtt < 0) {
                                seq_rtt = now - scb->when;
                                if (fully_acked)
@@ -2631,6 +2624,20 @@ static int tcp_clean_rtx_queue(struct sock *sk, s32 *seq_rtt_p)
                }
                tp->packets_out -= packets_acked;
 
+               /* Initial outgoing SYN's get put onto the write_queue
+                * just like anything else we transmit.  It is not
+                * true data, and if we misinform our callers that
+                * this ACK acks real data, we will erroneously exit
+                * connection startup slow start one packet too
+                * quickly.  This is severely frowned upon behavior.
+                */
+               if (!(scb->flags & TCPCB_FLAG_SYN)) {
+                       flag |= FLAG_DATA_ACKED;
+               } else {
+                       flag |= FLAG_SYN_ACKED;
+                       tp->retrans_stamp = 0;
+               }
+
                if (!fully_acked)
                        break;
 
@@ -2852,6 +2859,10 @@ static int tcp_process_frto(struct sock *sk, int flag)
        if (flag&FLAG_DATA_ACKED)
                inet_csk(sk)->icsk_retransmits = 0;
 
+       if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
+           ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
+               tp->undo_marker = 0;
+
        if (!before(tp->snd_una, tp->frto_highmark)) {
                tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
                return 1;
@@ -2916,6 +2927,7 @@ static int tcp_process_frto(struct sock *sk, int flag)
                        break;
                }
                tp->frto_counter = 0;
+               tp->undo_marker = 0;
        }
        return 0;
 }