In IPv4 code, we pull the ARP header into the linear section of the skb:

  [...]
  case bpf_htons(ETH_P_ARP):
    if (is_defined(ENABLE_ARP_PASSTHROUGH) ||
        is_defined(ENABLE_ARP_RESPONDER) ||
        CONFIG(enable_l2_announcements)) {
        if (!revalidate_data_arp_pull(ctx, &data, &data_end, &arp)) {
	  ret = DROP_INVALID;
	  goto drop_err_ingress;
        }

        [...]

        if (CONFIG(enable_l2_announcements)) {
          ret = handle_l2_announcement(ctx, NULL);
       [...]

This is however not the case in IPv6, and so a NIC driver which does not
pull in anything the handle_l2_announcement() could fail given only the
Ethernet + IPv6 header is in the linear section. Then we return with
CTX_ACT_OK and the packet goes up the stack. Given this is slow-path, pull
in skb->len. For XDP its a no-op.

Related: #43774
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Co-developed-by: Marc Suñé <marc.sune@isovalent.com>
Signed-off-by: Marc Suñé <marc.sune@isovalent.com>