Re: neighbour table RCU

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

Stephen Hemminger a écrit :
> On Tue, 01 Sep 2009 08:50:17 +0200
> Eric Dumazet <eric.dumazet@gmail.com> wrote:
> 
>> Stephen Hemminger a écrit :
>>> Looking at the neighbour table, it should be possible to get
>>> rid of the two reader/writer locks.  The hash table lock is pretty
>>> amenable to RCU, but the dynamic resizing makes it non-trivial.
>>> Thinking of using a combination of RCU and sequence counts so that the
>>> reader would just rescan if resize was in progress.
>> I am not sure neigh_tbl_lock rwlock should be changed, I did not
>> see any contention on it.
>>
>>> The reader/writer lock on the neighbour entry is more of a problem.
>>> Probably would be simpler/faster to change it into a spinlock and
>>> be done with it.
>>>
>>> The reader/writer lock is also used for the proxy list hash table,
>>> but that can just be a simple spinlock.
>>>
>> This is probably is the only thing we want to do at this moment,
>> halving atomic ops on neigh_resolve_output()
>>
>> But why neigh_resolve_output() was called so much in the bench
>> is the question...
>>
> 
> Every packet has to have an ARP resolution.
> 

Sure, but I thought we had a cache ?

static inline int ip_finish_output2(struct sk_buff *skb)
{
...
	if (dst->hh)
		return neigh_hh_output(dst->hh, skb);
	else if (dst->neighbour)
		return dst->neighbour->output(skb);  << should fill cache first time >>
...
}

in my pktgen benches, I always hit same dst so should take the hh cache ?