From: Eric Dumazet
> Sent: 31 March 2017 14:25
> On Fri, 2017-03-31 at 11:47 +0200, Paolo Abeni wrote:
> > In the udp_sock struct, the 'forward_deficit' and 'pcflag' fields
> > share the same cacheline. While the first is dirtied by
> > udp_recvmsg, the latter is read, possibly several times, by the
> > bottom half processing to discriminate between udp and udplite
> > sockets.
> >
> > With this patch, sk->sk_protocol is used to check is the socket is
> > really an udplite one, avoiding some cache misses per
> > packet and improving the performance under udp_flood with
> > small packet up to 10%.
...
> I am pretty sure we agreed in the past that forward_deficit would need
> to be placed on a cache line of its own. Somehow we manage to not
> implement this properly.
>
> What about other fields like encap_rcv, encap_destroy, gro_receive,
> gro_complete. They really should have the same false sharing issue.
>
> Proper fix is :
...
> - /* This field is dirtied by udp_recvmsg() */
> - int forward_deficit;
> + /* This field is dirtied by udp_recvmsg().
> + * Make sure it wont share a cache line with prior fields.
> + */
> + int forward_deficit ____cacheline_aligned_in_smp;
Is that really sensible on systems with large cache lines?
David
