Re: [REPOST patch net-next V6] net: introduce ethernet teaming device

[Jiri Pirko <jpirko@redhat.com>]

Thu, Nov 10, 2011 at 12:53:09AM CET, eric.dumazet@gmail.com wrote:
>Le jeudi 10 novembre 2011 à 00:12 +0100, Eric Dumazet a écrit :
>> Le mercredi 09 novembre 2011 à 23:13 +0100, Jiri Pirko a écrit :
>> > This patch introduces new network device called team. It supposes to be
>> > very fast, simple, userspace-driven alternative to existing bonding
>> > driver.
>> 
>> ...
>> 
>> > +/*
>> > + * note: already called with rcu_read_lock
>> > + */
>> > +static netdev_tx_t team_xmit(struct sk_buff *skb, struct net_device *dev)
>> > +{
>> > +	struct team *team = netdev_priv(dev);
>> > +	bool tx_success = false;
>> > +	unsigned int len = skb->len;
>> > +
>> > +	/*
>> > +	 * Ensure transmit function is called only in case there is at least
>> > +	 * one port present.
>> > +	 */
>> > +	if (likely(!list_empty(&team->port_list) && team->mode_ops.transmit))
>> > +		tx_success = team->mode_ops.transmit(team, skb);
>
>Not clear why its so complex here.
>
>When you manipulate team->port_list and make it empty, why dont you set
>team->mode_ops.transmit to a helper function, freeing the skb and
>returning false.
>
>Also, instead of setting .transmit to NULL, you also could set it to
>same helper function.
>
>This way you could just do in fast path :
>
>	tx_succcess = team->mode_ops.transmit(team, skb);
>
>Avoiding two tests

This approach I was thinking of as well. I was not sure that it would be
so nice to change this function from port_add/port_del but I'm going to
look at this.

>
>> > +	if (tx_success) {
>> > +		struct team_pcpu_stats *pcpu_stats;
>> > +
>> > +		pcpu_stats = this_cpu_ptr(team->pcpu_stats);
>> > +		u64_stats_update_begin(&pcpu_stats->syncp);
>> > +		pcpu_stats->tx_packets++;
>> > +		pcpu_stats->tx_bytes += len;
>> > +		u64_stats_update_end(&pcpu_stats->syncp);
>> > +	} else {
>> > +		this_cpu_inc(team->pcpu_stats->tx_dropped);
>> > +	}
>> > +
>> > +	return NETDEV_TX_OK;
>> > +}
>> > +
>> 
>
>
>