On Wed, 2014-08-27 at 14:31 +0200, Hannes Frederic Sowa wrote:

One thing one should keep in mind is, that there must be a skb available
to trigger the flush, maybe this will hurt us one day.

Thinking more about it should we go with a coccinelle script and
replace/extend ndo_start_xmit with an additional argument?

This will be a pain for backports and things like that.

skb->xmit_more is a bit annoying, because it consumes one bit in all
skbs, while it could be one byte per cpu as ndo_start_xmit() is called
while BH are disabled.

It also forces a cache line dirtying, that will hurt qdisc like HTB
where skbs can be cooked/enqueued by remote cpus.

Also, most drivers will never benefit from this new infra.

Mellanox mlx4 uses the Blue Frame trick to avoid the iowrite() cost.

I am afraid Jesper tests are not complete to truly have a picture of the
extra costs, because he made sure no false sharing was possible (A
single cpu does everything and keeps skb in its cache)