Re: [PATCH net] net: vrf: remove redundant vrf neigh entry

[David Ahern]

On 3/22/19 3:10 PM, linmiaohe wrote:
> From: Miaohe Lin <linmiaohe@xxxxxxxxxx>
> 
> When vrf->rth is created, it wouldn't change in his lifetime.And in
> func vrf_finish_output, we always create a neigh with ip_hdr(skb)->daddr
> as key because rth->rt_gateway is equal to 0. But I think we only need
> one vrf neigh entry because we strip the ethernet header and reset the
> dst_entry in vrf_process_v4_outbound.
> So I set rth->rt_gateway to loopback addr(It's ok to set any other
> valid ip address, just choose one.). And we would only create one vrf
> neigh entry. This helps saving some memory and improving the hitting
> rate of neigh lookup.
> If there is something I misunderstand, it's very nice of you to
> let me know. Thanks a lot.
> 
> Signed-off-by: linmiaohe <linmiaohe@xxxxxxxxxx>
> ---
>  drivers/net/vrf.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/drivers/net/vrf.c b/drivers/net/vrf.c
> index 7c1430ed0244..2b0227fb8f53 100644
> --- a/drivers/net/vrf.c
> +++ b/drivers/net/vrf.c
> @@ -738,6 +738,7 @@ static int vrf_rtable_create(struct net_device *dev)
>  		return -ENOMEM;
> 
>  	rth->dst.output	= vrf_output;
> +	rth->rt_gateway = htonl(INADDR_LOOPBACK);
> 
>  	rcu_assign_pointer(vrf->rth, rth);
> 

Did you investigate how neighbor entries are getting created? The vrf
device has IFF_NOARP set, so neigh entries should not be created.