Re: [PATCH] mlx4_en: Setting RSS hash result to skb->rxhash field

[Eric Dumazet <eric.dumazet@gmail.com>]

Le mercredi 04 mai 2011 à 15:44 +0100, Ben Hutchings a écrit :
> On Wed, 2011-05-04 at 16:37 +0300, Yevgeny Petrilin wrote:
> > Signed-off-by: Yevgeny Petrilin <yevgenyp@mellanox.co.il>
> > ---
> >  drivers/net/mlx4/en_rx.c |    3 +++
> >  1 files changed, 3 insertions(+), 0 deletions(-)
> > 
> > diff --git a/drivers/net/mlx4/en_rx.c b/drivers/net/mlx4/en_rx.c
> > index 62dd21b..bb4d66a 100644
> > --- a/drivers/net/mlx4/en_rx.c
> > +++ b/drivers/net/mlx4/en_rx.c
> > @@ -610,6 +610,8 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud
> >  					gro_skb->data_len = length;
> >  					gro_skb->truesize += length;
> >  					gro_skb->ip_summed = CHECKSUM_UNNECESSARY;
> > +					gro_skb->rxhash = be32_to_cpu(cqe->immed_rss_invalid) << 24;
> > +					skb_record_rx_queue(gro_skb, cq->ring);
> 
> An 8-bit hash is almost useless.  It's entirely useless if you then
> shift it into the top bits of rxhash.
> 

Agreed. This is very bad.

Yevgeny probably did this shift because get_rps_cpu()
does :

tcpu = map->cpus[((u64) skb->rxhash * map->len) >> 32];

(If rxhash is not a pure random u32 distribution, then high order bits
are more important than low order bits)