Re: [PATCH v2 nf-next] netfilter: conntrack: remove the central spinlock

From: Joe Perches <joe@perches.com>
To: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Pablo Neira Ayuso <pablo@netfilter.org>,
	netfilter-devel@vger.kernel.org, netdev <netdev@vger.kernel.org>,
	Tom Herbert <therbert@google.com>,
	Patrick McHardy <kaber@trash.net>
Subject: Re: [PATCH v2 nf-next] netfilter: conntrack: remove the central spinlock
Date: Wed, 22 May 2013 12:57:04 -0700	[thread overview]
Message-ID: <1369252624.2143.103.camel@joe-AO722> (raw)
In-Reply-To: <1369250772.3301.349.camel@edumazet-glaptop>

On Wed, 2013-05-22 at 12:26 -0700, Eric Dumazet wrote:
> On Wed, 2013-05-22 at 11:20 -0700, Joe Perches wrote:
> > On Wed, 2013-05-22 at 10:47 -0700, Eric Dumazet wrote:
> > > nf_conntrack_lock is a monolithic lock and suffers from huge contention
> > > on current generation servers (8 or more core/threads).
> > []
> > > diff --git a/include/net/netfilter/nf_conntrack.h b/include/net/netfilter/nf_conntrack.h
> > []
> > > @@ -76,7 +76,8 @@ struct nf_conn {
> > >             plus 1 for any connection(s) we are `master' for */
> > >  	struct nf_conntrack ct_general;
> > >  
> > > -	spinlock_t lock;
> > > +	spinlock_t	lock;
> > > +	u16		cpu;
> > trivia:
> > What's the real value in not using int here?
> On some machines, sizeof(spinlock_t) is 2
> So this addition doesn't increase size of the structure, as I fill a
> hole.
> Thats the case on x86 when NR_CPUS < 256

Ahh, nice.

It might also be nice to mark it if ever more than a u16
brace/flock/coven/cluster worth of cpus become feasible
as it seems int is used almost everywhere else.