Re: [PATCH net-next-2.6] net: optimize INET input path further

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

Le mercredi 01 décembre 2010 à 12:37 -0500, Brian Haley a écrit :
> On 12/01/2010 12:04 AM, Eric Dumazet wrote:
> >  struct sock_common {
> > -	/*
> > -	 * first fields are not copied in sock_copy()
> > +	/* skc_daddr and skc_rcv_saddr must be grouped :
> > +	 * cf INET_MATCH() and INET_TW_MATCH()
> >  	 */
> > -	union {
> > -		struct hlist_node	skc_node;
> > -		struct hlist_nulls_node skc_nulls_node;
> > -	};
> > -	atomic_t		skc_refcnt;
> > -	int			skc_tx_queue_mapping;
> > +	__be32			skc_daddr;
> > +	__be32			skc_rcv_saddr;
> >  
> >  	union  {
> >  		unsigned int	skc_hash;
> 
> Putting IPv4 addresses in sock_common doesn't make it so common anymore :)
> 
> Is it possible to make this a union so other address families like IPv6
> can benefit from this as well, or will that blow the whole cache line
> effect you were trying to achieve?

This might be OK, depending on cache line size and/or arch.

On x86_32 for example, that might even be a good thing, because refcnt
might still be in the first 64bytes of socket.

By the way, ipv6 sock includes inet, so includes ipv4 addresses too, I
only moved them in the 'whole structure'