Re: [PATCH v3 1/2] core: Factor out flow calculation from get_rps_cpu

[Changli Gao <xiaosuo@gmail.com>]

On Tue, Aug 3, 2010 at 1:57 PM, Krishna Kumar2 <krkumar2@in.ibm.com> wrote:
> Hi Changli,
>
> Good catch.
>
> Instead of adding support for ethernet header or pull/push,
> I could defer the skb_push(ETH_HLEN), something like:
>
> static int macvtap_forward(struct net_device *dev, struct sk_buff *skb)
> {
>        struct macvtap_queue *q = macvtap_get_queue(dev, skb);
>        if (!q)
>                goto drop;
>
>        if (skb_queue_len(&q->sk.sk_receive_queue) >= dev->tx_queue_len)
>                goto drop;
>
> +       skb_push(skb, ETH_HLEN);
>        ...
> }
>
> and remove the same in macvtap_receive. Will this be better?
>

I am confused by the call sites of macvlan_dev.receive and
macvlan_dev.forward. They both are possible to be called in both
RX(skb->data points to network header) and TX(skb->data points to
ethernet) paths. The current code in macvtap shows that
macvlan_dev.receive should be called in network layer, and
macvlan_dev.forward should be called in dev layer. Am I correct?


-- 
Regards,
Changli Gao(xiaosuo@gmail.com)