Re: [PATCH v3 net-next 1/4] net: implement support for low latency socket polling

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

On Tue, 2013-05-21 at 10:28 +0300, Eliezer Tamir wrote:
> On 20/05/2013 18:29, Eric Dumazet wrote:
> > On Mon, 2013-05-20 at 13:16 +0300, Eliezer Tamir wrote:
> ---
> >> +static inline void skb_mark_ll(struct sk_buff *skb, struct napi_struct *napi)
> >> +{
> >> +	skb->dev_ref = napi;
> >> +}
> >> +
> >> +static inline void sk_mark_ll(struct sock *sk, struct sk_buff *skb)
> >> +{
> >> +	sk->dev_ref = skb->dev_ref;
> >> +}
> >
> > I do not see why it's safe to keep a pointer to a napi object without
> > taking a reference, or something to prevent object being removed.
> >
> > Using a genid might be enough. (some counter incremented every time a
> > napi is dismantled)
> 
> I really like this approach and I tried it.
> The main problem I had is that you need to increase the size of the skb 
> to store the generation id unless you stuff it in the flags2 bitfield.
> There appear to be only 7 useful bit left there.
> Is it OK to use them all up?
> 
> 
> > Alternatively, use a napi_id instead of a pointer.
> 
> I'm not sure I understand what you propose.

Oh well.

To get a pointer to a struct net_device, we can use ifindex, and do a
rcu lookup into a hash table to get the net_device. We do not need
{pointer,ifindex} but {ifindex} is enough

My suggestion is to not have skb->skb_ref but skb->napi_index : Its safe
to copy its value from skb->napi_index to sk->napi_index without
refcounting.

All NAPI need to get a unique napi_index, and be inserted in a hash
table for immediate/fast lookup.