Just checked with kernel 6.19. This one hase the upstream patched mentioned included and works. So it looks like this patch, at least in the way it's backported on debian, is not working like intended..
b.t.w: the re6 tunnel configured runs (on one side) over a vlan on ethernet and I don't have crashes. Since the regression is with and without vlan tough it seems unlikly that vlan is actually the issue. Looking at what skb_vlan_inet_prepare does it should be equal to pskb_inet_may_pull(skb) tough, or at least I see no difference. Still, recompiling with !pskb_inet_may_pull(skb) instead of skb_vlan_inet_prepare(skb, true) immediatly fixes the issue for me. I'm far from following the changes in the network stack, but could it be that skb_reset_network_header behaves different on 6.19 and 6.12 so that skb_vlan_inet_prepare on 6.12 sees some junk when checking if it's a vlan header and messes with sizes in a way that it could fail here?
