I noticed, looking at tcpdumps, that timewait ACKs were getting sent
with an incorrect MD5 signature when signatures were enabled.
I broke this in 
49a72dfb8814c2d65bd9f8c9c6daf6395a1ec58d ("tcp: Fix
MD5 signatures for non-linear skbs"). I didn't take into account that
the skb passed to tcp_*_send_ack was the inbound packet, thus the
source and dest addresses need to be swapped when calculating the MD5
pseudoheader.
Signed-off-by: Adam Langley <agl@imperialviolet.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
                rep.th.doff = arg.iov[0].iov_len/4;
 
                tcp_v4_md5_hash_hdr((__u8 *) &rep.opt[offset],
-                                   key, ip_hdr(skb)->daddr,
-                                   ip_hdr(skb)->saddr, &rep.th);
+                                   key, ip_hdr(skb)->saddr,
+                                   ip_hdr(skb)->daddr, &rep.th);
        }
 #endif
        arg.csum = csum_tcpudp_nofold(ip_hdr(skb)->daddr,
 
                *topt++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
                                (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
                tcp_v6_md5_hash_hdr((__u8 *)topt, key,
-                                   &ipv6_hdr(skb)->daddr,
-                                   &ipv6_hdr(skb)->saddr, t1);
+                                   &ipv6_hdr(skb)->saddr,
+                                   &ipv6_hdr(skb)->daddr, t1);
        }
 #endif