]> err.no Git - linux-2.6/commitdiff
[TCP]: Must count fack_count also when skipping
authorIlpo Järvinen <ilpo.jarvinen@helsinki.fi>
Mon, 3 Mar 2008 20:10:16 +0000 (12:10 -0800)
committerDavid S. Miller <davem@davemloft.net>
Mon, 3 Mar 2008 20:10:16 +0000 (12:10 -0800)
It makes fackets_out to grow too slowly compared with the
real write queue.

This shouldn't cause those BUG_TRAP(packets <= tp->packets_out)
to trigger but how knows how such inconsistent fackets_out
affects here and there around TCP when everything is nowadays
assuming accurate fackets_out. So lets see if this silences
them all.

Reported by Guillaume Chazarain <guichaz@gmail.com>.

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

index 19c449f62672d7cf41f03a65f6d6dbc0cc3b3c04..7facdb0f69608be661e409c2ead641f5d0fb1bc8 100644 (file)
@@ -1367,7 +1367,7 @@ static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
  * a normal way
  */
 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
-                                       u32 skip_to_seq)
+                                       u32 skip_to_seq, int *fack_count)
 {
        tcp_for_write_queue_from(skb, sk) {
                if (skb == tcp_send_head(sk))
@@ -1375,6 +1375,8 @@ static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
 
                if (!before(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
                        break;
+
+               *fack_count += tcp_skb_pcount(skb);
        }
        return skb;
 }
@@ -1390,7 +1392,7 @@ static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
                return skb;
 
        if (before(next_dup->start_seq, skip_to_seq)) {
-               skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq);
+               skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq, fack_count);
                tcp_sacktag_walk(skb, sk, NULL,
                                 next_dup->start_seq, next_dup->end_seq,
                                 1, fack_count, reord, flag);
@@ -1537,7 +1539,8 @@ tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
 
                        /* Head todo? */
                        if (before(start_seq, cache->start_seq)) {
-                               skb = tcp_sacktag_skip(skb, sk, start_seq);
+                               skb = tcp_sacktag_skip(skb, sk, start_seq,
+                                                      &fack_count);
                                skb = tcp_sacktag_walk(skb, sk, next_dup,
                                                       start_seq,
                                                       cache->start_seq,
@@ -1565,7 +1568,8 @@ tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
                                goto walk;
                        }
 
-                       skb = tcp_sacktag_skip(skb, sk, cache->end_seq);
+                       skb = tcp_sacktag_skip(skb, sk, cache->end_seq,
+                                              &fack_count);
                        /* Check overlap against next cached too (past this one already) */
                        cache++;
                        continue;
@@ -1577,7 +1581,7 @@ tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb,
                                break;
                        fack_count = tp->fackets_out;
                }
-               skb = tcp_sacktag_skip(skb, sk, start_seq);
+               skb = tcp_sacktag_skip(skb, sk, start_seq, &fack_count);
 
 walk:
                skb = tcp_sacktag_walk(skb, sk, next_dup, start_seq, end_seq,