On 11/26/2018 03:38 AM, Florian Westphal wrote:
> The (out-of-tree) Multipath-TCP implementation needs a significant amount
> of extra space in the skb control buffer.
>
> Increasing skb->cb[] size in mainline is a non-starter for memory and
> and performance reasons (f.e. increase in cb size also moves several
> frequently-accessed fields to other cache lines).
>
One thing that could be done without too much impact is to provide a cbext[]
only for TCP packets in write/rtx queue, that is not in sk_buff but
on the struct sk_buff_fclones
This extra space would not be 0-initialized at alloc_skb()
and would not be copied at skb_clone()
I mentioned this idea a while back already.
diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
index
73902acf2b71c8800d81b744a936a7420f33b459..c4bfc2fd98eb9723c0219d5cd8bf5b28afaf5398
100644
--- a/include/linux/skbuff.h
+++ b/include/linux/skbuff.h
@@ -1018,6 +1018,8 @@ struct sk_buff_fclones {
struct sk_buff skb2;
refcount_t fclone_ref;
+
+ char cbext[128] __aligned(8);
};
/**
