From: Eli Cooper <redacted>

3.12-stable review patch.  If anyone has any objections, please let me know.

===============

commit b4e479a96fc398ccf83bb1cffb4ffef8631beaf1 upstream.

When xfrm is applied to TSO/GSO packets, it follows this path:

    xfrm_output() -> xfrm_output_gso() -> skb_gso_segment()

where skb_gso_segment() relies on skb->protocol to function properly.

This patch sets skb->protocol to ETH_P_IPV6 before dst_output() is called,
fixing a bug where GSO packets sent through an ipip6 tunnel are dropped
when xfrm is involved.

Signed-off-by: Eli Cooper <redacted>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Jiri Slaby <redacted>
---
 net/ipv6/output_core.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/net/ipv6/output_core.c b/net/ipv6/output_core.c
index 4bd870af05d6..ab40997a1c2a 100644
--- a/net/ipv6/output_core.c
+++ b/net/ipv6/output_core.c

@@ -117,6 +117,8 @@ int __ip6_local_out(struct sk_buff *skb)
 		len = 0;
 	ipv6_hdr(skb)->payload_len = htons(len);
 
+	skb->protocol = htons(ETH_P_IPV6);
+
 	return nf_hook(NFPROTO_IPV6, NF_INET_LOCAL_OUT, skb, NULL,
 		       skb_dst(skb)->dev, dst_output);
 }

-- 
2.11.0