diff mbox

bridge: per-cpu packet statistics

Message ID 20100301161658.5a61143b@nehalam
State Superseded, archived
Delegated to: David Miller
Headers show

Commit Message

stephen hemminger March 2, 2010, 12:16 a.m. UTC 
The shared packet statistics are a potential source of slow down
on bridged traffic. Convert to per-cpu array, but only keep those
statistics which change per-packet.

Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>

---

 net/bridge/br_device.c  |   43 ++++++++++++++++++++++++++++++++++++++-----
 net/bridge/br_if.c      |    6 ++++++
 net/bridge/br_input.c   |    5 +++--
 net/bridge/br_private.h |    8 ++++++++
 4 files changed, 55 insertions(+), 7 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

Comments

Eric Dumazet March 2, 2010, 6:01 a.m. UTC | #1 
Le lundi 01 mars 2010 à 16:16 -0800, Stephen Hemminger a écrit :
> The shared packet statistics are a potential source of slow down
> on bridged traffic. Convert to per-cpu array, but only keep those
> statistics which change per-packet.
> 
> Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
> 
> ---
> 
>  net/bridge/br_device.c  |   43 ++++++++++++++++++++++++++++++++++++++-----
>  net/bridge/br_if.c      |    6 ++++++
>  net/bridge/br_input.c   |    5 +++--
>  net/bridge/br_private.h |    8 ++++++++
>  4 files changed, 55 insertions(+), 7 deletions(-)
> 
> --- a/net/bridge/br_device.c	2010-03-01 08:22:23.476657998 -0800
> +++ b/net/bridge/br_device.c	2010-03-01 15:31:36.737227465 -0800
> @@ -26,11 +26,12 @@ netdev_tx_t br_dev_xmit(struct sk_buff *
>  	const unsigned char *dest = skb->data;
>  	struct net_bridge_fdb_entry *dst;
>  	struct net_bridge_mdb_entry *mdst;
> +	struct br_cpu_netstats *brstats = this_cpu_ptr(br->stats);
>  
> -	BR_INPUT_SKB_CB(skb)->brdev = dev;
> +	brstats->tx_packets++;
> +	brstats->tx_bytes += skb->len;


On TX path, this is not really necessary, since we already dirtied
txq->lock before calling br_dev_xmit(), we can use txq->tx_packets and
txq->tx_bytes for free ?

>  
> -	dev->stats.tx_packets++;
> -	dev->stats.tx_bytes += skb->len;
> +	BR_INPUT_SKB_CB(skb)->brdev = dev;
>  
>  	skb_reset_mac_header(skb);
>  	skb_pull(skb, ETH_HLEN);
> @@ -81,6 +82,28 @@ static int br_dev_stop(struct net_device
>  	return 0;
>  }
>  


--
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
Eric Dumazet March 2, 2010, 7:43 a.m. UTC | #2 
Le lundi 01 mars 2010 à 16:16 -0800, Stephen Hemminger a écrit :
> +static void br_dev_free(struct net_device *dev)
> +{
> +	struct net_bridge *br = netdev_priv(dev);
> +
> +	free_percpu(br->stats);
> +}
> +
>  void br_dev_setup(struct net_device *dev)
>  {
>  	random_ether_addr(dev->dev_addr);
>  	ether_setup(dev);
>  
>  	dev->netdev_ops = &br_netdev_ops;
> -	dev->destructor = free_netdev;
> +	dev->destructor = br_dev_free;
>  	SET_ETHTOOL_OPS(dev, &br_ethtool_ops);

Isnt free_netdev() missing after this change ?



--
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
Eric Dumazet March 2, 2010, 7:51 a.m. UTC | #3 
Le lundi 01 mars 2010 à 16:16 -0800, Stephen Hemminger a écrit :
> --- a/net/bridge/br_if.c	2010-03-01 08:22:23.476657998 -0800
> +++ b/net/bridge/br_if.c	2010-03-01 15:30:47.733227819 -0800
> @@ -185,6 +185,12 @@ static struct net_device *new_bridge_dev
>  	br = netdev_priv(dev);
>  	br->dev = dev;
>  
> +	br->stats = alloc_percpu(sizeof(struct br_cpu_netstats));
> +	if (!br->stats) {
> +		free_netdev(dev);
> +		return NULL;
> +	}
> +

Strange... this should be :

	nr->stats = alloc_percpu(struct br_cpu_netstats);

Or even better, ask percpu allocator an aligned chunk 
(2 or 4 longs) instead of (1 long)

	nr->stats = __alloc_percpu(sizeof(struct br_cpu_netstats),
				   4 * sizeof(long));


--
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
stephen hemminger March 2, 2010, 5:22 p.m. UTC | #4 
On Tue, 02 Mar 2010 07:01:30 +0100
Eric Dumazet <eric.dumazet@gmail.com> wrote:

> Le lundi 01 mars 2010 à 16:16 -0800, Stephen Hemminger a écrit :
> > The shared packet statistics are a potential source of slow down
> > on bridged traffic. Convert to per-cpu array, but only keep those
> > statistics which change per-packet.
> > 
> > Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
> > 
> > ---
> > 
> >  net/bridge/br_device.c  |   43 ++++++++++++++++++++++++++++++++++++++-----
> >  net/bridge/br_if.c      |    6 ++++++
> >  net/bridge/br_input.c   |    5 +++--
> >  net/bridge/br_private.h |    8 ++++++++
> >  4 files changed, 55 insertions(+), 7 deletions(-)
> > 
> > --- a/net/bridge/br_device.c	2010-03-01 08:22:23.476657998 -0800
> > +++ b/net/bridge/br_device.c	2010-03-01 15:31:36.737227465 -0800
> > @@ -26,11 +26,12 @@ netdev_tx_t br_dev_xmit(struct sk_buff *
> >  	const unsigned char *dest = skb->data;
> >  	struct net_bridge_fdb_entry *dst;
> >  	struct net_bridge_mdb_entry *mdst;
> > +	struct br_cpu_netstats *brstats = this_cpu_ptr(br->stats);
> >  
> > -	BR_INPUT_SKB_CB(skb)->brdev = dev;
> > +	brstats->tx_packets++;
> > +	brstats->tx_bytes += skb->len;
> 
> 
> On TX path, this is not really necessary, since we already dirtied
> txq->lock before calling br_dev_xmit(), we can use txq->tx_packets and
> txq->tx_bytes for free ?

Bridge is already using lockless transmit LLTX, so tx_lock is not touched.
--
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
Herbert Xu March 16, 2010, 2:48 a.m. UTC | #5 
Stephen Hemminger <shemminger@vyatta.com> wrote:
>
> Bridge is already using lockless transmit LLTX, so tx_lock is not touched.

LLTX doesn't actually buy you anything since you're still going
through a single qdisc.  To get the full benefits of the per-cpu
counters you need to implement multiqueue support in the bridge.

Cheers,
diff mbox

Patch

--- a/net/bridge/br_device.c	2010-03-01 08:22:23.476657998 -0800
+++ b/net/bridge/br_device.c	2010-03-01 15:31:36.737227465 -0800
@@ -26,11 +26,12 @@  netdev_tx_t br_dev_xmit(struct sk_buff *
 	const unsigned char *dest = skb->data;
 	struct net_bridge_fdb_entry *dst;
 	struct net_bridge_mdb_entry *mdst;
+	struct br_cpu_netstats *brstats = this_cpu_ptr(br->stats);
 
-	BR_INPUT_SKB_CB(skb)->brdev = dev;
+	brstats->tx_packets++;
+	brstats->tx_bytes += skb->len;
 
-	dev->stats.tx_packets++;
-	dev->stats.tx_bytes += skb->len;
+	BR_INPUT_SKB_CB(skb)->brdev = dev;
 
 	skb_reset_mac_header(skb);
 	skb_pull(skb, ETH_HLEN);
@@ -81,6 +82,28 @@  static int br_dev_stop(struct net_device
 	return 0;
 }
 
+static struct net_device_stats *br_get_stats(struct net_device *dev)
+{
+	struct net_bridge *br = netdev_priv(dev);
+	struct net_device_stats *stats = &dev->stats;
+	unsigned int cpu;
+
+	stats->tx_bytes = stats->tx_packets = 0;
+	stats->rx_bytes = stats->rx_packets = 0;
+
+	for_each_online_cpu(cpu) {
+		const struct br_cpu_netstats *bstats
+			= per_cpu_ptr(br->stats, cpu);
+
+		stats->tx_bytes   += bstats->tx_bytes;
+		stats->tx_packets += bstats->tx_packets;
+		stats->rx_bytes   += bstats->rx_bytes;
+		stats->rx_packets += bstats->rx_packets;
+	}
+
+	return stats;
+}
+
 static int br_change_mtu(struct net_device *dev, int new_mtu)
 {
 	struct net_bridge *br = netdev_priv(dev);
@@ -180,19 +203,27 @@  static const struct net_device_ops br_ne
 	.ndo_open		 = br_dev_open,
 	.ndo_stop		 = br_dev_stop,
 	.ndo_start_xmit		 = br_dev_xmit,
+	.ndo_get_stats		 = br_get_stats,
 	.ndo_set_mac_address	 = br_set_mac_address,
 	.ndo_set_multicast_list	 = br_dev_set_multicast_list,
 	.ndo_change_mtu		 = br_change_mtu,
 	.ndo_do_ioctl		 = br_dev_ioctl,
 };
 
+static void br_dev_free(struct net_device *dev)
+{
+	struct net_bridge *br = netdev_priv(dev);
+
+	free_percpu(br->stats);
+}
+
 void br_dev_setup(struct net_device *dev)
 {
 	random_ether_addr(dev->dev_addr);
 	ether_setup(dev);
 
 	dev->netdev_ops = &br_netdev_ops;
-	dev->destructor = free_netdev;
+	dev->destructor = br_dev_free;
 	SET_ETHTOOL_OPS(dev, &br_ethtool_ops);
 	dev->tx_queue_len = 0;
 	dev->priv_flags = IFF_EBRIDGE;
--- a/net/bridge/br_if.c	2010-03-01 08:22:23.476657998 -0800
+++ b/net/bridge/br_if.c	2010-03-01 15:30:47.733227819 -0800
@@ -185,6 +185,12 @@  static struct net_device *new_bridge_dev
 	br = netdev_priv(dev);
 	br->dev = dev;
 
+	br->stats = alloc_percpu(sizeof(struct br_cpu_netstats));
+	if (!br->stats) {
+		free_netdev(dev);
+		return NULL;
+	}
+
 	spin_lock_init(&br->lock);
 	INIT_LIST_HEAD(&br->port_list);
 	spin_lock_init(&br->hash_lock);
--- a/net/bridge/br_input.c	2010-03-01 08:22:23.476657998 -0800
+++ b/net/bridge/br_input.c	2010-03-01 15:32:45.882471626 -0800
@@ -23,9 +23,11 @@  const u8 br_group_address[ETH_ALEN] = { 
 static int br_pass_frame_up(struct sk_buff *skb)
 {
 	struct net_device *indev, *brdev = BR_INPUT_SKB_CB(skb)->brdev;
+	struct net_bridge *br = netdev_priv(brdev);
+	struct br_cpu_netstats *brstats = this_cpu_ptr(br->stats);
 
-	brdev->stats.rx_packets++;
-	brdev->stats.rx_bytes += skb->len;
+	brstats->rx_packets++;
+	brstats->rx_bytes += skb->len;
 
 	indev = skb->dev;
 	skb->dev = brdev;
--- a/net/bridge/br_private.h	2010-03-01 08:22:23.476657998 -0800
+++ b/net/bridge/br_private.h	2010-03-01 15:31:03.437228864 -0800
@@ -135,6 +135,14 @@  struct net_bridge
 	spinlock_t			lock;
 	struct list_head		port_list;
 	struct net_device		*dev;
+
+	struct br_cpu_netstats __percpu {
+		unsigned long	rx_packets;
+		unsigned long	tx_packets;
+		unsigned long	rx_bytes;
+		unsigned long	tx_bytes;
+	} *stats;
+
 	spinlock_t			hash_lock;
 	struct hlist_head		hash[BR_HASH_SIZE];
 	unsigned long			feature_mask;