Re: [RFC PATCH 4/5] mlx4: add support for fast rx drop bpf program

[Brenden Blanco]

On Sat, Apr 02, 2016 at 11:15:38PM -0700, Brenden Blanco wrote:
> On Fri, Apr 01, 2016 at 07:08:31PM -0700, Eric Dumazet wrote:
[...]
> > 2) priv->stats.rx_dropped is shared by all the RX queues -> false
> > sharing.
> > 
> >    This is probably the right time to add a rx_dropped field in struct
> > mlx4_en_rx_ring since you guys want to drop 14 Mpps, and 50 Mpps on
> > higher speed links.
> > 
> This sounds reasonable! Will look into it for the next spin.
I looked into this, and it seems to me that both the rx and tx dropped
stats are buggy. With commit a3333b35da1634f49aca541f2574a084221e2616,
specifically with the line
  stats->rx_dropped = be32_to_cpu(mlx4_en_stats->RDROP);
that occurs during the periodic ethtool task, whatever ++ was happening
in the rx/tx code is overwritten with the HW value. Since the SW stats
are incremented mostly in edge (oom) cases, nobody probably noticed. To
me it doesn't seem right to mix hard and soft counters, especially at
the risk of making a bad situation worse, so I'm planning to omit the
new bpf dropped++ stat and we can discuss ways to fix this other bug
separately.