net: skb_vlan_untag(): don't reset transport offset if set by GRO layer

Similar to commit fda55eca5a33f
("net: introduce skb_transport_header_was_set()"), avoid resetting
transport offsets that were already set by GRO layer. This not only
mirrors the behavior of __netif_receive_skb_core(), but also makes
sense when it comes to UDP GSO fraglists forwarding: transport offset
of such skbs is set only once by GRO receive callback and remains
untouched and correct up to the xmitting driver in 1:1 case, but
becomes junk after untagging in ingress VLAN case and breaks UDP
GSO offload. This does not happen after this change, and all types
of forwarding of UDP GSO fraglists work as expected.

Since v1 [1]:
 - keep the code 1:1 with __netif_receive_skb_core() (Jakub).

[1] https://lore.kernel.org/netdev/zYurwsZRN7BkqSoikWQLVqHyxz18h4LhHU4NFa2Vw@cp4-web-038.plabs.ch

Signed-off-by: Alexander Lobakin <alobakin@pm.me>
Link: https://lore.kernel.org/r/7JgIkgEztzt0W6ZtC9V9Cnk5qfkrUFYcpN871syCi8@cp4-web-040.plabs.ch
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

diff --git a/net/core/skbuff.c b/net/core/skbuff.c
index c5e6c0b83a9266c4f5f1412df37fa744f13b423f..c9a5a3c262c8307678a889da04929e255c62b59e 100644 (file)
--- a/net/core/skbuff.c
+++ b/net/core/skbuff.c
@@ -5441,7 +5441,8 @@ struct sk_buff *skb_vlan_untag(struct sk_buff *skb)
                goto err_free;
 
        skb_reset_network_header(skb);
-       skb_reset_transport_header(skb);
+       if (!skb_transport_header_was_set(skb))
+               skb_reset_transport_header(skb);
        skb_reset_mac_len(skb);
 
        return skb;