On Fri, 2014-09-26 at 13:06 +0200, Hannes Frederic Sowa wrote:

I thought that prefetchw() would be the canonical way to solve write
stalls in CPUs and prepare the specific cache line to be written to.
Interesting, thanks Eric.

True, and I have a tcp_ack() (tcp_clean_rtx_queue() more exactly) patch
that indeed uses prefetchw() to bring tcp_shinfo() in cache before the 
tcp_unlink_write_queue(skb, sk);

Performance went from 570000 packets per second to 820000 packets per
second on a TSO=off GSO=off workload.

But this applies after the https://patchwork.ozlabs.org/patch/392877/
patch (tcp: change tcp_skb_pcount() location )