Re: netfilter conntrack tcp lock

Re: netfilter conntrack tcp lock

[Paul E. McKenney]

On Tue, Apr 21, 2009 at 09:04:12PM -0700, Stephen Hemminger wrote:
> Not sure what the performance impact would be but simply changing tcp_lock
> (in nf_conntrack_proto_tcp) to a spin_lock might get a performance boost.
> I thought I heard Paul say read_locks are way slower than spin_lock.

In -rt, read_locks are really exclusive locks.  :-/

And an uncontended read_lock is somewhat slower than a spin_lock,
but not overwhelmingly so.

> Alternatively, going to some form of bit based hash lock might work.

I have to ask the stupid question...  Can conntrack entries be hashed
or otherwise partitioned, with a lock then assigned to each partition?

						Thanx, Paul

Re: netfilter conntrack tcp lock

[Eric Dumazet]

Stephen Hemminger a écrit :
> Not sure what the performance impact would be but simply changing tcp_lock
> (in nf_conntrack_proto_tcp) to a spin_lock might get a performance boost.
> I thought I heard Paul say read_locks are way slower than spin_lock.
> 
> Alternatively, going to some form of bit based hash lock might work.
> 
> 

I thought Patrick was OK with adding a spinlock into each nf_conn struct ?
(We shrunk this struct with the rcu_head removal)

A straight forward locking schem. We have spinlocks everywhere in kernel afterall, and
conntracking deserves this.
 
Just resubmit one of your prior patch perhaps ? I can dig it in a couple of hours eventually.

Thanks

--
To unsubscribe from this list: send the line "unsubscribe netfilter-devel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: netfilter conntrack tcp lock

[Patrick McHardy]

Eric Dumazet wrote:
> Stephen Hemminger a écrit :
>> Not sure what the performance impact would be but simply changing tcp_lock
>> (in nf_conntrack_proto_tcp) to a spin_lock might get a performance boost.
>> I thought I heard Paul say read_locks are way slower than spin_lock.
>>
>> Alternatively, going to some form of bit based hash lock might work.
>>
> I thought Patrick was OK with adding a spinlock into each nf_conn struct ?
> (We shrunk this struct with the rcu_head removal)

Yes, my main concern was the further size increase.

> A straight forward locking schem. We have spinlocks everywhere in kernel afterall, and
> conntracking deserves this.
>  
> Just resubmit one of your prior patch perhaps ? I can dig it in a couple of hours eventually.

Yes, please.

--
To unsubscribe from this list: send the line "unsubscribe netfilter-devel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html