Re: [PATCH][RFC] udp: cache sock to avoid searching it twice

[Paolo Abeni <pabeni@redhat.com>]

Hi,

Adding Willem, I think he can be interested.

On Fri, 2018-11-09 at 14:21 +0800, Li RongQing wrote:
> GRO for UDP needs to lookup socket twice, first is in gro receive,
> second is gro complete, so if store sock to skb to avoid looking up
> twice, this can give small performance boost
> 
> netperf -t UDP_RR -l 10
> 
> Before:
> 	Rate per sec: 28746.01
> After:
> 	Rate per sec: 29401.67
> 
> Signed-off-by: Li RongQing <lirongqing@baidu.com>
> ---
>  net/ipv4/udp_offload.c | 18 +++++++++++++++++-
>  1 file changed, 17 insertions(+), 1 deletion(-)
> 
> diff --git a/net/ipv4/udp_offload.c b/net/ipv4/udp_offload.c
> index 0646d61f4fa8..429570112a33 100644
> --- a/net/ipv4/udp_offload.c
> +++ b/net/ipv4/udp_offload.c
> @@ -408,6 +408,11 @@ struct sk_buff *udp_gro_receive(struct list_head *head, struct sk_buff *skb,
>  
>  	if (udp_sk(sk)->gro_enabled) {
>  		pp = call_gro_receive(udp_gro_receive_segment, head, skb);
> +
> +		if (!IS_ERR(pp) && NAPI_GRO_CB(pp)->count > 1) {
> +			sock_hold(sk);
> +			pp->sk = sk;
> +		}
>  		rcu_read_unlock();
>  		return pp;
>  	}

What if 'pp' is NULL?

Aside from that, this replace a lookup with 2 atomic ops, and only when
such lookup is amortized on multiple aggregated packets: I'm unsure if
it's worthy and I don't understand how that improves RR tests (where
the socket can't see multiple, consecutive skbs, AFAIK).

Cheers,

Paolo