Re: [PATCH v2] netdev:bfin_mac: reclaim and free tx skb as soon as possible after transfer

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

Le jeudi 03 juin 2010 à 16:57 +0800, Junchang Wang a écrit :
> Hi Eric,
> 
> On Thu, Jun 3, 2010 at 12:05 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> > Not related to your patch, but reviewing it I see this driver still do
> > the "dev->stats.tx_packets++; dev->stats.tx_bytes += (skb->len);"
> >
> > This is not necessary and expensive, since we update txq stats in core
> > network stack.
> >
> >        rc = ops->ndo_start_xmit(skb, dev);
> >        if (rc == NETDEV_TX_OK)
> >                txq_trans_update(txq);  << here >>
> >
> 
> Good suggestion for drivers. But I wonder whether there are stats for
> received packets in core network stack.
> 

No its not there.

> I.e., can I replace "dev->stats.rx_packets++" and "dev->stats.rx_bytes
> += (skb->len);" with something already maintained by core stack? I
> failed to find them.
> 

As I said, core network takes care of three counters only, because it
was 'free', as they share a cache line with a spinlock we must hold when
calling xmit function.

In receive path, we dont dirty a cache line in core network, so updating
counters would add a cost. (modern NICs handle stats in firmware)