[ovs-dev] [PATCH V3 18/19] netdev-offload-dpdk: Support offload of clone tnl_push/output actions

[Sathya Perla]

On Sun, Dec 8, 2019 at 6:53 PM Eli Britstein <elibr at mellanox.com> wrote:
>
...
> +static int
> +parse_clone_actions(struct netdev *netdev,
> +                    struct flow_actions *actions,
> +                    const struct nlattr *clone_actions,
> +                    const size_t clone_actions_len,
> +                    struct offload_info *info)
> +{
> +    const struct nlattr *ca;
> +    unsigned int cleft;
> +
> +    NL_ATTR_FOR_EACH_UNSAFE (ca, cleft, clone_actions, clone_actions_len) {
> +        int clone_type = nl_attr_type(ca);
> +
> +        if (clone_type == OVS_ACTION_ATTR_TUNNEL_PUSH) {
> +            const struct ovs_action_push_tnl *tnl_push = nl_attr_get(ca);
> +            struct rte_flow_action_raw_encap *raw_encap =
> +                xzalloc(sizeof *raw_encap);
> +
> +            raw_encap->data = (uint8_t *)tnl_push->header;
> +            raw_encap->preserve = NULL;
> +            raw_encap->size = tnl_push->header_len;
> +
> +            add_flow_action(actions, RTE_FLOW_ACTION_TYPE_RAW_ENCAP,
> +                            raw_encap);

Hi, converting OVS_TUNNLE_PUSH into RTE_RAW_ENCAP loses the 'tnl_type'
information provided by OVS. RAW_ENCAP provides the tunnel header just
as a blob of bits which may not be ideal for NIC HW to offload.

How about using tnl_push->tnl_type field to parse the header and
compose specific tunnel encap actions like RTE_VXLAN_ENCAP,
RTE_NVGRE_ENCAP etc.
This would be also be more inline with how it's done with TC-offload.

thanks!