| Message ID | 20120621235011.29846.29715.stgit@gitlad.jf.intel.com |
|---|---|
| State | Accepted, archived |
| Delegated to: | David Miller |
| Headers | show |
From: Alexander Duyck <alexander.h.duyck@intel.com> Date: Thu, 21 Jun 2012 16:58:31 -0700 > This change is meant to add a control for disabling early socket demux. > The main motivation behind this patch is to provide an option to disable > the feature as it adds an additional cost to routing that reduces overall > throughput by up to 5%. For example one of my systems went from 12.1Mpps > to 11.6 after the early socket demux was added. It looks like the reason > for the regression is that we are now having to perform two lookups, first > the one for an established socket, and then the one for the routing table. > > By adding this patch and toggling the value for ip_early_demux to 0 I am > able to get back to the 12.1Mpps I was previously seeing. > > Cc: David S. Miller <davem@davemloft.net> > Cc: Eric Dumazet <edumazet@google.com> > Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com> I applied this for now, making a minor change to move the local variables down into the new basic block you created. There has got to be a way to make this really cheap. At the very least we can have the GRO code store away the ports and therefore allow us to just do a direct call to try and demux the socket. Thus, we'd avoid all of pskb_may_pull() et al. packet validations, and packet header pointer calculations. Furthermore, we can reduce to overhead by making a special inet established hash demux that doesn't check for time-wait sockets, reducing the number of probes to 1 from 2. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, 2012-06-22 at 17:15 -0700, David Miller wrote: > I applied this for now, making a minor change to move the local > variables down into the new basic block you created. > Hmm, sorry to come late, but you left NET_IPV4_EARLY_DEMUX=126 sysctl, while this is deprecated way... > There has got to be a way to make this really cheap. At the very > least we can have the GRO code store away the ports and therefore > allow us to just do a direct call to try and demux the socket. Thus, > we'd avoid all of pskb_may_pull() et al. packet validations, and > packet header pointer calculations. > > Furthermore, we can reduce to overhead by making a special inet > established hash demux that doesn't check for time-wait sockets, > reducing the number of probes to 1 from 2. The timewait hash chain is on the same cache line than established one. And on a router, both chains are empty with a 99.999 % probability. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
From: Eric Dumazet <eric.dumazet@gmail.com> Date: Sat, 23 Jun 2012 07:45:26 +0200 > On Fri, 2012-06-22 at 17:15 -0700, David Miller wrote: > >> Furthermore, we can reduce to overhead by making a special inet >> established hash demux that doesn't check for time-wait sockets, >> reducing the number of probes to 1 from 2. > > The timewait hash chain is on the same cache line than established one. > And on a router, both chains are empty with a 99.999 % probability. I understand this. Probably a lot of the overhead has to do with the function calls and, as I mentioned, the transport layer probing and validation. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/include/linux/sysctl.h b/include/linux/sysctl.h index c34b4c8..20825e5 100644 --- a/include/linux/sysctl.h +++ b/include/linux/sysctl.h @@ -425,6 +425,7 @@ enum NET_TCP_ALLOWED_CONG_CONTROL=123, NET_TCP_MAX_SSTHRESH=124, NET_TCP_FRTO_RESPONSE=125, + NET_IPV4_EARLY_DEMUX=126, }; enum { diff --git a/include/net/ip.h b/include/net/ip.h index 83e0619..50841bd 100644 --- a/include/net/ip.h +++ b/include/net/ip.h @@ -210,6 +210,9 @@ extern int inet_peer_threshold; extern int inet_peer_minttl; extern int inet_peer_maxttl; +/* From ip_input.c */ +extern int sysctl_ip_early_demux; + /* From ip_output.c */ extern int sysctl_ip_dynaddr; diff --git a/kernel/sysctl_binary.c b/kernel/sysctl_binary.c index a650694..6a3cf82 100644 --- a/kernel/sysctl_binary.c +++ b/kernel/sysctl_binary.c @@ -415,6 +415,8 @@ static const struct bin_table bin_net_ipv4_table[] = { { CTL_INT, NET_IPV4_IPFRAG_SECRET_INTERVAL, "ipfrag_secret_interval" }, /* NET_IPV4_IPFRAG_MAX_DIST "ipfrag_max_dist" no longer used */ + { CTL_INT, NET_IPV4_EARLY_DEMUX, "ip_early_demux" }, + { CTL_INT, 2088 /* NET_IPQ_QMAX */, "ip_queue_maxlen" }, /* NET_TCP_DEFAULT_WIN_SCALE unused */ diff --git a/net/ipv4/ip_input.c b/net/ipv4/ip_input.c index 93b092c..07de38d 100644 --- a/net/ipv4/ip_input.c +++ b/net/ipv4/ip_input.c @@ -313,6 +313,8 @@ drop: return true; } +int sysctl_ip_early_demux __read_mostly = 1; + static int ip_rcv_finish(struct sk_buff *skb) { const struct iphdr *iph = ip_hdr(skb); @@ -325,14 +327,15 @@ static int ip_rcv_finish(struct sk_buff *skb) if (skb_dst(skb) == NULL) { const struct net_protocol *ipprot; int protocol = iph->protocol; - int err; - - rcu_read_lock(); - ipprot = rcu_dereference(inet_protos[protocol]); - err = -ENOENT; - if (ipprot && ipprot->early_demux) - err = ipprot->early_demux(skb); - rcu_read_unlock(); + int err = -ENOENT; + + if (sysctl_ip_early_demux) { + rcu_read_lock(); + ipprot = rcu_dereference(inet_protos[protocol]); + if (ipprot && ipprot->early_demux) + err = ipprot->early_demux(skb); + rcu_read_unlock(); + } if (err) { err = ip_route_input_noref(skb, iph->daddr, iph->saddr, diff --git a/net/ipv4/sysctl_net_ipv4.c b/net/ipv4/sysctl_net_ipv4.c index ef32956..12aa0c5 100644 --- a/net/ipv4/sysctl_net_ipv4.c +++ b/net/ipv4/sysctl_net_ipv4.c @@ -301,6 +301,13 @@ static struct ctl_table ipv4_table[] = { .proc_handler = proc_dointvec }, { + .procname = "ip_early_demux", + .data = &sysctl_ip_early_demux, + .maxlen = sizeof(int), + .mode = 0644, + .proc_handler = proc_dointvec + }, + { .procname = "ip_dynaddr", .data = &sysctl_ip_dynaddr, .maxlen = sizeof(int),
This change is meant to add a control for disabling early socket demux. The main motivation behind this patch is to provide an option to disable the feature as it adds an additional cost to routing that reduces overall throughput by up to 5%. For example one of my systems went from 12.1Mpps to 11.6 after the early socket demux was added. It looks like the reason for the regression is that we are now having to perform two lookups, first the one for an established socket, and then the one for the routing table. By adding this patch and toggling the value for ip_early_demux to 0 I am able to get back to the 12.1Mpps I was previously seeing. Cc: David S. Miller <davem@davemloft.net> Cc: Eric Dumazet <edumazet@google.com> Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com> --- I am open to any comments or suggestions on this patch. I had seen the earlier discussions and saw mention of adding a control for disabling the early demux feature so I figured I would just code it up real quick once I ran into the regression. I am assuming it is okay to disable the early demux code since I suspect there is other code in place that will still handle the demux for the TCP sockets later. Also I wasn't sure about the sysctl since I haven't set one up before. include/linux/sysctl.h | 1 + include/net/ip.h | 3 +++ kernel/sysctl_binary.c | 2 ++ net/ipv4/ip_input.c | 19 +++++++++++-------- net/ipv4/sysctl_net_ipv4.c | 7 +++++++ 5 files changed, 24 insertions(+), 8 deletions(-) -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html