]> err.no Git - linux-2.6/blobdiff - net/ipv4/tcp_output.c
[POWERPC] CPM: Rename commproc to cpm1 and cpm2_common.c to cpm2.c
[linux-2.6] / net / ipv4 / tcp_output.c
index 9df5b2a4da6e37d9bcd53d6fd490e26eee601975..f4c1eef89af0b3ac1e63ed91ecf3725a761d6949 100644 (file)
@@ -652,6 +652,26 @@ static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned
        }
 }
 
+/* When a modification to fackets out becomes necessary, we need to check
+ * skb is counted to fackets_out or not. Another important thing is to
+ * tweak SACK fastpath hint too as it would overwrite all changes unless
+ * hint is also changed.
+ */
+static void tcp_adjust_fackets_out(struct tcp_sock *tp, struct sk_buff *skb,
+                                  int decr)
+{
+       if (!tp->sacked_out || tcp_is_reno(tp))
+               return;
+
+       if (!before(tp->highest_sack, TCP_SKB_CB(skb)->seq))
+               tp->fackets_out -= decr;
+
+       /* cnt_hint is "off-by-one" compared with fackets_out (see sacktag) */
+       if (tp->fastpath_skb_hint != NULL &&
+           after(TCP_SKB_CB(tp->fastpath_skb_hint)->seq, TCP_SKB_CB(skb)->seq))
+               tp->fastpath_cnt_hint -= decr;
+}
+
 /* Function to create two new TCP segments.  Shrinks the given segment
  * to the specified size and appends a new segment with the rest of the
  * packet to the list.  This won't be called frequently, I hope.
@@ -667,7 +687,7 @@ int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss
 
        BUG_ON(len > skb->len);
 
-       clear_all_retrans_hints(tp);
+       tcp_clear_retrans_hints_partial(tp);
        nsize = skb_headlen(skb) - len;
        if (nsize < 0)
                nsize = 0;
@@ -692,7 +712,8 @@ int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss
        TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
        TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
 
-       if (tp->sacked_out && (TCP_SKB_CB(skb)->seq == tp->highest_sack))
+       if (tcp_is_sack(tp) && tp->sacked_out &&
+           (TCP_SKB_CB(skb)->seq == tp->highest_sack))
                tp->highest_sack = TCP_SKB_CB(buff)->seq;
 
        /* PSH and FIN should only be set in the second packet. */
@@ -746,21 +767,12 @@ int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss
                if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
                        tp->lost_out -= diff;
 
-               if (diff > 0) {
-                       /* Adjust Reno SACK estimate. */
-                       if (tcp_is_reno(tp)) {
-                               tcp_dec_pcount_approx_int(&tp->sacked_out, diff);
-                               tcp_verify_left_out(tp);
-                       }
-
-                       tcp_dec_pcount_approx_int(&tp->fackets_out, diff);
-                       /* SACK fastpath might overwrite it unless dealt with */
-                       if (tp->fastpath_skb_hint != NULL &&
-                           after(TCP_SKB_CB(tp->fastpath_skb_hint)->seq,
-                                 TCP_SKB_CB(skb)->seq)) {
-                               tcp_dec_pcount_approx_int(&tp->fastpath_cnt_hint, diff);
-                       }
+               /* Adjust Reno SACK estimate. */
+               if (tcp_is_reno(tp) && diff > 0) {
+                       tcp_dec_pcount_approx_int(&tp->sacked_out, diff);
+                       tcp_verify_left_out(tp);
                }
+               tcp_adjust_fackets_out(tp, skb, diff);
        }
 
        /* Link BUFF into the send queue. */
@@ -1150,8 +1162,7 @@ int tcp_may_send_now(struct sock *sk)
        return (skb &&
                tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
                             (tcp_skb_is_last(sk, skb) ?
-                             TCP_NAGLE_PUSH :
-                             tp->nonagle)));
+                             tp->nonagle : TCP_NAGLE_PUSH)));
 }
 
 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
@@ -1283,6 +1294,7 @@ static int tcp_mtu_probe(struct sock *sk)
        struct sk_buff *skb, *nskb, *next;
        int len;
        int probe_size;
+       int size_needed;
        unsigned int pif;
        int copy;
        int mss_now;
@@ -1301,27 +1313,20 @@ static int tcp_mtu_probe(struct sock *sk)
        /* Very simple search strategy: just double the MSS. */
        mss_now = tcp_current_mss(sk, 0);
        probe_size = 2*tp->mss_cache;
+       size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
        if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
                /* TODO: set timer for probe_converge_event */
                return -1;
        }
 
        /* Have enough data in the send queue to probe? */
-       len = 0;
-       if ((skb = tcp_send_head(sk)) == NULL)
-               return -1;
-       while ((len += skb->len) < probe_size && !tcp_skb_is_last(sk, skb))
-               skb = tcp_write_queue_next(sk, skb);
-       if (len < probe_size)
+       if (tp->write_seq - tp->snd_nxt < size_needed)
                return -1;
 
-       /* Receive window check. */
-       if (after(TCP_SKB_CB(skb)->seq + probe_size, tp->snd_una + tp->snd_wnd)) {
-               if (tp->snd_wnd < probe_size)
-                       return -1;
-               else
-                       return 0;
-       }
+       if (tp->snd_wnd < size_needed)
+               return -1;
+       if (after(tp->snd_nxt + size_needed, tp->snd_una + tp->snd_wnd))
+               return 0;
 
        /* Do we need to wait to drain cwnd? */
        pif = tcp_packets_in_flight(tp);
@@ -1340,7 +1345,6 @@ static int tcp_mtu_probe(struct sock *sk)
 
        skb = tcp_send_head(sk);
        tcp_insert_write_queue_before(nskb, skb, sk);
-       tcp_advance_send_head(sk, skb);
 
        TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
        TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
@@ -1707,8 +1711,9 @@ static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int m
                BUG_ON(tcp_skb_pcount(skb) != 1 ||
                       tcp_skb_pcount(next_skb) != 1);
 
-               /* changing transmit queue under us so clear hints */
-               clear_all_retrans_hints(tp);
+               if (WARN_ON(tcp_is_sack(tp) && tp->sacked_out &&
+                   (TCP_SKB_CB(next_skb)->seq == tp->highest_sack)))
+                       return;
 
                /* Ok.  We will be able to collapse the packet. */
                tcp_unlink_write_queue(next_skb, sk);
@@ -1726,10 +1731,6 @@ static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int m
                /* Update sequence range on original skb. */
                TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
 
-               if (WARN_ON(tp->sacked_out &&
-                   (TCP_SKB_CB(next_skb)->seq == tp->highest_sack)))
-                       return;
-
                /* Merge over control information. */
                flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
                TCP_SKB_CB(skb)->flags = flags;
@@ -1746,11 +1747,17 @@ static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int m
                if (tcp_is_reno(tp) && tp->sacked_out)
                        tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
 
-               /* Not quite right: it can be > snd.fack, but
-                * it is better to underestimate fackets.
-                */
-               tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
+               tcp_adjust_fackets_out(tp, next_skb, tcp_skb_pcount(next_skb));
                tp->packets_out -= tcp_skb_pcount(next_skb);
+
+               /* changed transmit queue under us so clear hints */
+               tcp_clear_retrans_hints_partial(tp);
+               /* manually tune sacktag skb hint */
+               if (tp->fastpath_skb_hint == next_skb) {
+                       tp->fastpath_skb_hint = skb;
+                       tp->fastpath_cnt_hint -= tcp_skb_pcount(skb);
+               }
+
                sk_stream_free_skb(sk, next_skb);
        }
 }
@@ -1784,7 +1791,7 @@ void tcp_simple_retransmit(struct sock *sk)
                }
        }
 
-       clear_all_retrans_hints(tp);
+       tcp_clear_all_retrans_hints(tp);
 
        if (!lost)
                return;
@@ -1899,6 +1906,8 @@ int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
                                printk(KERN_DEBUG "retrans_out leaked.\n");
                }
 #endif
+               if (!tp->retrans_out)
+                       tp->lost_retrans_low = tp->snd_nxt;
                TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
                tp->retrans_out += tcp_skb_pcount(skb);