On Sun, Aug 18, 2019 at 07:00:59PM +0300, Paul Blakey wrote:
> What do you guys say about the following diff on top of the last one?
> Use static key, and also have OVS_DP_CMD_SET command probe/enable the feature.
>
> This will allow userspace to probe the feature, and selectivly enable it via
> the
> OVS_DP_CMD_SET command.
I'm not convinced yet that we need something like this. Been
wondering, skb_ext_find() below is not that expensive if not in use.
It's just a bit check and that's it, it returns NULL.
And drivers will only be setting this if they have tc-offloading
enabled (assuming they won't be seeing it for chain 0 all the time).
On which case, with tc offloading, we need this in order to work
properly.
Is the bit checking really that worrysome?
> --- a/net/openvswitch/flow.c
> +++ b/net/openvswitch/flow.c
> @@ -853,8 +853,10 @@ int ovs_flow_key_extract(const struct ip_tunnel_info
> *tun_info,
> key->mac_proto = res;
>
> #if IS_ENABLED(CONFIG_NET_TC_SKB_EXT)
> - tc_ext = skb_ext_find(skb, TC_SKB_EXT);
> - key->recirc_id = tc_ext ? tc_ext->chain : 0;
> + if (static_branch_unlikely(&tc_recirc_sharing_support)) {
> + tc_ext = skb_ext_find(skb, TC_SKB_EXT);
> + key->recirc_id = tc_ext ? tc_ext->chain : 0;
> + }
> #else
> key->recirc_id = 0;
> #endif
> --
> 1.8.3.1
>
