Even using percpu stats, we still hit tunnel dst_entry refcount in
ip6_tnl_xmit2()

Since we are in RCU locked section, we can use skb_dst_set_noref() and
avoid these atomic operations, leaving dst shared on cpus.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>