Re: Bug in computing data_len in tcp_sendmsg?

[Eric Dumazet <eric.dumazet@gmail.com>]

Le vendredi 02 décembre 2011 à 15:24 -0500, David Miller a écrit :
> From: Eric Dumazet <eric.dumazet@gmail.com>
> Date: Fri, 02 Dec 2011 21:22:49 +0100
> 
> > Le vendredi 02 décembre 2011 à 13:40 -0500, David Miller a écrit :
> > 
> >> Yes, for non-SG this always was technically possible.
> >> 
> > 
> > Yes this can, I reproduced it very easily.
> > 
> > I find this hard to believe...
> 
> Grrr :-)
> 
> Ok, I'll think about this some more.

Maybe a quick fix would be to trim not len bytes but (len & ~3) bytes ?

This avoid reallocations and complex code for a 'should never happen in
normal circumstances'...

Retransmits could transmits 3 bytes already ACKed, is it a big deal ?

patch on top of linux/net tree :

diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
index 63170e2..4e108c5 100644
--- a/net/ipv4/tcp_output.c
+++ b/net/ipv4/tcp_output.c
@@ -1126,7 +1126,7 @@ int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
 
 	/* If len == headlen, we avoid __skb_pull to preserve alignment. */
 	if (unlikely(len < skb_headlen(skb)))
-		__skb_pull(skb, len);
+		__skb_pull(skb, len & ~3); /* preserve alignement of tcp/ip headers */
 	else
 		__pskb_trim_head(skb, len - skb_headlen(skb));