From: Eric Dumazet <eric.duma...@gmail.com> Date: Mon, 28 Nov 2016 10:32:15 -0800
> On Mon, 2016-11-28 at 10:17 -0800, Eric Dumazet wrote: > >> I was referring to Stephen bug report. >> >> It appears that Eli changelog was not very precise, because his bug was >> because of XFRM being involved. >> >> xfrm_output() -> xfrm_output_gso() -> skb_gso_segment() >> >> So XFRM calls skb_gso_segment() before ip_output() or >> ip6_finish_output2() had a chance to change skb->protocol > > So maybe the real fix would be to set skb->protocol at the right place, > before xfrm can be called, instead of chasing all skb producers ;) And the key for this would be dst_output() invocations.