On Fri, May 17, 2013 at 10:26:14AM -0700, Eric Dumazet wrote:
> On Fri, 2013-05-17 at 15:50 +0900, Simon Horman wrote:
>
> > @@ -509,6 +511,8 @@ struct sk_buff {
> > __u32 reserved_tailroom;
> > };
> >
> > + __be16 inner_protocol;
> > + /* 16/48 bit hole */
> > sk_buff_data_t inner_transport_header;
> > sk_buff_data_t inner_network_header;
> > sk_buff_data_t inner_mac_header;
>
> We are reaching the point where sk_buff is so big that per packet cost
> is killing us, and guys want linux kernel bypass.
>
> sizeof(sk_buff) = 0xf8 -> 0x100
>
> sizeof(skbuff_fclone_cache) = 0x200 -> 0x240
>
> So TCP stack performance is going to be hurt by this change, as the
> atomic_t containing fclone_ref will be in a separate cache line.
>
> __copy_skb_header() needs to be smarter and perform a bulk copy using
> long words
>
> Maybe we could use 16 bits instead of 32 for the inner_*_headers ?
Thanks. I have made a patch to implement that and it seems to
both reduce the size of struct sk_buff and leave a hole for inner_protocol.
I will also update this patch to guard inner_protocol with
#ifdef CONFIG_NET_MPLS_GSO in struct sk_buff as it is otherwise unused.
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
