On Thu, 9 Mar 2017 14:09:18 -0700
Subash Abhinov Kasiviswanathan <subas...@codeaurora.org> wrote:
> Certain system process significant unconnected UDP workload.
> It would be preferrable to disable UDP early demux for those systems
> and enable it for TCP only.
>
> Signed-off-by: Subash Abhinov Kasiviswanathan <subas...@codeaurora.org>
> Suggested-by: Eric Dumazet <eduma...@google.com>
> ---
This makes sense.
> diff --git a/net/ipv4/ip_input.c b/net/ipv4/ip_input.c
> index d6feabb..187feae 100644
> --- a/net/ipv4/ip_input.c
> +++ b/net/ipv4/ip_input.c
> @@ -329,7 +329,7 @@ static int ip_rcv_finish(struct net *net, struct sock
> *sk, struct sk_buff *skb)
> int protocol = iph->protocol;
>
> ipprot = rcu_dereference(inet_protos[protocol]);
> - if (ipprot && ipprot->early_demux) {
> + if (ipprot && ipprot->early_demux &&
> *ipprot->early_demux_enabled) {
> ipprot->early_demux(skb);
> /* must reload iph, skb->head might have changed */
> iph = ip_hdr(skb);
Another possible option would be change the function pointer for early_demux
instead of having
an additional conditional test (and cache line read). The downside of doing it
that way
is the code to turn the sysctl on/off gets more complicated than simple
standard proc_int_vec.
