Re: [PATCH net-next] net: Optimize IPv6 path in ip_neigh_for_gw()

[Breno Leitao]

Hello David,

On Fri, Oct 04, 2024 at 11:01:29AM -0600, David Ahern wrote:
> On 10/4/24 10:27 AM, Breno Leitao wrote:
> > Branch annotation traces from approximately 200 IPv6-enabled hosts
> > revealed that the 'likely' branch in ip_neigh_for_gw() was consistently
> > mispredicted. Given the increasing prevalence of IPv6 in modern networks,
> > this commit adjusts the function to favor the IPv6 path.
> > 
> > Swap the order of the conditional statements and move the 'likely'
> > annotation to the IPv6 case. This change aims to improve performance in
> > IPv6-dominant environments by reducing branch mispredictions.
> > 
> > This optimization aligns with the trend of IPv6 becoming the default IP
> > version in many deployments, and should benefit modern network
> > configurations.
> > 
> > Signed-off-by: Breno Leitao <leitao@xxxxxxxxxx>
> > ---
> >  include/net/route.h | 6 +++---
> >  1 file changed, 3 insertions(+), 3 deletions(-)
> > 
> > diff --git a/include/net/route.h b/include/net/route.h
> > index 1789f1e6640b..b90b7b1effb8 100644
> > --- a/include/net/route.h
> > +++ b/include/net/route.h
> > @@ -389,11 +389,11 @@ static inline struct neighbour *ip_neigh_for_gw(struct rtable *rt,
> >  	struct net_device *dev = rt->dst.dev;
> >  	struct neighbour *neigh;
> >  
> > -	if (likely(rt->rt_gw_family == AF_INET)) {
> > -		neigh = ip_neigh_gw4(dev, rt->rt_gw4);
> > -	} else if (rt->rt_gw_family == AF_INET6) {
> > +	if (likely(rt->rt_gw_family == AF_INET6)) {
> >  		neigh = ip_neigh_gw6(dev, &rt->rt_gw6);
> >  		*is_v6gw = true;
> > +	} else if (rt->rt_gw_family == AF_INET) {
> > +		neigh = ip_neigh_gw4(dev, rt->rt_gw4);
> >  	} else {
> >  		neigh = ip_neigh_gw4(dev, ip_hdr(skb)->daddr);
> >  	}
> 
> This is an IPv4 function allowing support for IPv6 addresses as a
> nexthop. It is appropriate for IPv4 family checks to be first.

Right. In which case is this called on IPv6 only systems?

On my IPv6-only 200 systems, the annotated branch predictor is showing
it is mispredicted 100% of the time.

Thanks for the review
--breno