tcp: don't clear lost_skb_hint when not necessary
authorIlpo Järvinen <ilpo.jarvinen@helsinki.fi>
Sun, 21 Sep 2008 04:25:52 +0000 (21:25 -0700)
committerDavid S. Miller <davem@davemloft.net>
Sun, 21 Sep 2008 04:25:52 +0000 (21:25 -0700)
Most importantly avoid doing it with cumulative ACK. However,
since we have lost_cnt_hint in the picture as well needing
adjustments, it's not as trivial as dealing with
retransmit_skb_hint (and cannot be done in the all place we
could trivially leave retransmit_skb_hint untouched).

With the previous patch, this should mostly remove O(n^2)
behavior while cumulative ACKs start flowing once rexmit
after a lossy round-trip made it through.

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

index 44a4fff..85627f8 100644 (file)
@@ -2844,6 +2844,7 @@ static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets)
        int flag = 0;
        u32 pkts_acked = 0;
        u32 reord = tp->packets_out;
+       u32 prior_sacked = tp->sacked_out;
        s32 seq_rtt = -1;
        s32 ca_seq_rtt = -1;
        ktime_t last_ackt = net_invalid_timestamp();
@@ -2925,9 +2926,11 @@ static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets)
 
                tcp_unlink_write_queue(skb, sk);
                sk_wmem_free_skb(sk, skb);
-               tcp_clear_retrans_hints_partial(tp);
+               tp->scoreboard_skb_hint = NULL;
                if (skb == tp->retransmit_skb_hint)
                        tp->retransmit_skb_hint = NULL;
+               if (skb == tp->lost_skb_hint)
+                       tp->lost_skb_hint = NULL;
        }
 
        if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
@@ -2946,6 +2949,15 @@ static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets)
                        /* Non-retransmitted hole got filled? That's reordering */
                        if (reord < prior_fackets)
                                tcp_update_reordering(sk, tp->fackets_out - reord, 0);
+
+                       /* No need to care for underflows here because
+                        * the lost_skb_hint gets NULLed if we're past it
+                        * (or something non-trivial happened)
+                        */
+                       if (tcp_is_fack(tp))
+                               tp->lost_cnt_hint -= pkts_acked;
+                       else
+                               tp->lost_cnt_hint -= prior_sacked - tp->sacked_out;
                }
 
                tp->fackets_out -= min(pkts_acked, tp->fackets_out);