[ Just Cc-ing linux-sctp@ ]
On Mon, Dec 23, 2019 at 06:45:35PM +0100, Lorenzo Bianconi wrote:
> Hi all,
>
> I am currently working on the following scenario:
>
> sctp --> ipv4 --> gretap --> ipv6 --> eth
>
> If NETIF_F_SCTP_CRC is not supported by the network device (it is the
> case for gre), sctp will fallback computing the crc32 in sw with
> sctp_gso_make_checksum(), where SKB_GSO_CB(skb)->csum is set to ~0 by
> gso_reset_checksum(). After the gso segmentation, gre_gso_segment()
> will try to compute gre csum with gso_make_checksum() even if skb->ip_summed
> is set to CHECKSUM_NONE (and so using ~0 as partial).
> One possible (trivial and not tested) solution would be to recompute the
> gre checksum, doing in gre_gso_segment() something like:
>
> if (skb->ip_summed == CHECKSUM_NONE) {
> ...
> err = skb_checksum_help(skb);
> if (err < 0)
> return return ERR_PTR(err);
> } else {
> *pcsum = gso_make_checksum(skb, 0);
> }
>
> One possible improvement would be offload the GRE checksum computation if the
> hw exports this capability in netdev_features and fall back to the sw
> implementation if not.
> Am I missing something? Is there a better approach?
>
> Regards,
> Lorenzo
