Re: [PATCH] net: increase skb->users instead of skb_clone()

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

Le jeudi 16 décembre 2010 à 22:05 +0800, Junchang Wang a écrit :
> On Thu, Dec 16, 2010 at 1:57 PM, Changli Gao <xiaosuo@gmail.com> wrote:
> > In dev_queue_xmit_nit(), we have to clone skbs as we need to mangle skbs,
> > however, we don't need to clone skbs for all the packet_types.
> >
> > Except for the first packet_type, we increase skb->users instead of
> > skb_clone().
> 
> Hi Changli,
> Take af_packet for example, I can't see benefit from this patch.
> 
> > +static inline int deliver_skb(struct sk_buff *skb,
> > +                             struct packet_type *pt_prev,
> > +                             struct net_device *orig_dev)
> > +{
> > +       atomic_inc(&skb->users);
> > +       return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
> > +}
> The increment call will incur skb_shared() failure in packet_rcv.
> In reality, packet_rcv has to clone this packet by itself.
> 

Yes, and no.

Consider the case you have one receiver.

Packet given after Changli patch wont be shared, so packet_rcv wont
clone it : Thats a win. Only one skb_clone() done instead of two.

Consider case with 2 receivers :

First time we call packet_rcv, packet is shared (because we call
deliver_skb(), so packet_rcv clones it. Normal situation, we really need
to clone it.

Second time, we give a non shared packet : Thats a win over previous
situation.