From mboxrd@z Thu Jan 1 00:00:00 1970 From: Zhouyi Zhou Subject: Re: [PATCH RFC] netfilter: cacheline aligning in struct netns_ct Date: Wed, 12 Mar 2014 20:20:21 +0800 Message-ID: References: <1394610994-22794-1-git-send-email-yizhouzhou@ict.ac.cn> <063D6719AE5E284EB5DD2968C1650D6D0F6DB542@AcuExch.aculab.com> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1 Cc: "davem@davemloft.net" , "pablo@netfilter.org" , "jbenc@redhat.com" , "brouer@redhat.com" , "netdev@vger.kernel.org" , "linux-kernel@vger.kernel.org" , Zhouyi Zhou To: David Laight Return-path: In-Reply-To: <063D6719AE5E284EB5DD2968C1650D6D0F6DB542@AcuExch.aculab.com> Sender: linux-kernel-owner@vger.kernel.org List-Id: netdev.vger.kernel.org Thanks Jesper and David for reviewing and replying. Great commitment of Jesper for moving above three hlist_nulls_heads to percpu structure. Sorry to miss the newest advances in nf-next. And arbitrarily cache-line aligning structure members is not a good idea for large cache lines. Still I have an idea that could we pack the seldom changed members (nf_ct_proto included) to a structure and we allocated it from a cache aligned slub when the structure "net" is initialized. Zhouyi On Wed, Mar 12, 2014 at 5:36 PM, David Laight wrote: > From: Zhouyi Zhou >> not frequently changing components should share same cachelines >> >> Signed-off-by: Zhouyi Zhou >> --- >> include/net/netns/conntrack.h | 12 ++++++++---- >> 1 file changed, 8 insertions(+), 4 deletions(-) >> >> diff --git a/include/net/netns/conntrack.h b/include/net/netns/conntrack.h >> index fbcc7fa..69d2d58 100644 >> --- a/include/net/netns/conntrack.h >> +++ b/include/net/netns/conntrack.h >> @@ -65,6 +65,12 @@ struct nf_ip_net { >> struct netns_ct { >> atomic_t count; >> unsigned int expect_count; >> + struct hlist_nulls_head unconfirmed; >> + struct hlist_nulls_head dying; >> + struct hlist_nulls_head tmpl; >> + >> + /* not frequently changing components should share same cachelines */ >> + struct nf_ip_net nf_ct_proto ____cacheline_aligned_in_smp; > > I'm not sure this has the effect you are trying to achieve. > Adding an __attribute__((aligned(n))) to a structure member not only > adds padding before the member, but also pads the member to a multiple > of the specified size. > > With large cache lines you probably want neither. > IIRC one of the ppc systems has something like 256 byte cache lines > (and if it doesn't now, it might have soon). > So arbitrarily cache-line aligning structure members may not be a good idea. > > In any case reducing the number of cache lines accessed, and the number > dirtied is probably more important than putting 'readonly' data in its > own cache lines. > > David > > >