[PATCH] bridge: per-cpu packet statistics

[PATCH] bridge: per-cpu packet statistics

[Stephen Hemminger]

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;
 
-	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;

Re: [PATCH] bridge: per-cpu packet statistics

[Eric Dumazet]

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;
>  }
>  

Re: [PATCH] bridge: per-cpu packet statistics

[Eric Dumazet]

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 ?

Re: [PATCH] bridge: per-cpu packet statistics

[Eric Dumazet]

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));

Re: [PATCH] bridge: per-cpu packet statistics

[Eric Dumazet]

From: Stephen Hemminger <shemminger@vyatta.com>

Le lundi 01 mars 2010 à 16:16 -0800, Stephen Hemminger a écrit :

> +	for_each_online_cpu(cpu) {
> +		const struct br_cpu_netstats *bstats
> +			= per_cpu_ptr(br->stats, cpu);
> +
> +		stats->rx_bytes   += bstats->rx_bytes;
> +		stats->rx_packets += bstats->rx_packets;
> +	}

And last point, we should use for_each_possible_cpu() here

Here is your patch with all my comments integrated :

1) Use txq->{tx_bytes|tx_packets} counter

2) alloc_percpu(struct ...) instead of alloc_percpu(sizeof(struct ...))

3) free_netdev() in destructor

4) for_each_possible_cpu() instead of for_each_online_cpu()

5) br_get_stats() use local variables for the sake of concurrent users

Next step would be to use multiqueue :)

Thanks

[PATCH] bridge: per-cpu packet statistics

The shared packet statistics are a potential source of slow down
on bridged traffic. Convert to per-cpu array for rx_packets/rx_bytes RX
accounting, and use txq tx_packets/tx_bytes for TX accounting

Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
---
 net/bridge/br_device.c  |   39 +++++++++++++++++++++++++++++++++++---
 net/bridge/br_if.c      |    6 +++++
 net/bridge/br_input.c   |    6 +++--
 net/bridge/br_private.h |    6 +++++
 4 files changed, 52 insertions(+), 5 deletions(-)

diff --git a/net/bridge/br_device.c b/net/bridge/br_device.c
index eb7062d..e73c42c 100644
--- a/net/bridge/br_device.c
+++ b/net/bridge/br_device.c
@@ -22,6 +22,8 @@
 /* net device transmit always called with no BH (preempt_disabled) */
 netdev_tx_t br_dev_xmit(struct sk_buff *skb, struct net_device *dev)
 {
+	int qidx = skb_get_queue_mapping(skb);
+	struct netdev_queue *txq = netdev_get_tx_queue(dev, qidx);
 	struct net_bridge *br = netdev_priv(dev);
 	const unsigned char *dest = skb->data;
 	struct net_bridge_fdb_entry *dst;
@@ -29,8 +31,8 @@ netdev_tx_t br_dev_xmit(struct sk_buff *skb, struct net_device *dev)
 
 	BR_INPUT_SKB_CB(skb)->brdev = dev;
 
-	dev->stats.tx_packets++;
-	dev->stats.tx_bytes += skb->len;
+	txq->tx_packets++;
+	txq->tx_bytes += skb->len;
 
 	skb_reset_mac_header(skb);
 	skb_pull(skb, ETH_HLEN);
@@ -81,6 +83,28 @@ static int br_dev_stop(struct net_device *dev)
 	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 long rx_bytes = 0, rx_packets = 0;
+	unsigned int cpu;
+
+	dev_txq_stats_fold(dev, stats);
+
+	for_each_possible_cpu(cpu) {
+		const struct br_cpu_netstats *bstats
+			= per_cpu_ptr(br->stats, cpu);
+
+		rx_bytes   += bstats->rx_bytes;
+		rx_packets += bstats->rx_packets;
+	}
+	stats->rx_bytes   = rx_bytes;
+	stats->rx_packets = 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 +204,28 @@ static const struct net_device_ops br_netdev_ops = {
 	.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);
+	free_netdev(dev);
+}
+
 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;
diff --git a/net/bridge/br_if.c b/net/bridge/br_if.c
index b6a3872..b7cdd2e 100644
--- a/net/bridge/br_if.c
+++ b/net/bridge/br_if.c
@@ -185,6 +185,12 @@ static struct net_device *new_bridge_dev(struct net *net, const char *name)
 	br = netdev_priv(dev);
 	br->dev = dev;
 
+	br->stats = alloc_percpu(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);
diff --git a/net/bridge/br_input.c b/net/bridge/br_input.c
index 53b3985..7a5a5b4 100644
--- a/net/bridge/br_input.c
+++ b/net/bridge/br_input.c
@@ -23,9 +23,11 @@ const u8 br_group_address[ETH_ALEN] = { 0x01, 0x80, 0xc2, 0x00, 0x00, 0x00 };
 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;
diff --git a/net/bridge/br_private.h b/net/bridge/br_private.h
index 9191198..06b30af 100644
--- a/net/bridge/br_private.h
+++ b/net/bridge/br_private.h
@@ -135,6 +135,12 @@ struct net_bridge
 	spinlock_t			lock;
 	struct list_head		port_list;
 	struct net_device		*dev;
+
+	struct br_cpu_netstats __percpu {
+		unsigned long	rx_bytes;
+		unsigned long	rx_packets;
+	} *stats;
+
 	spinlock_t			hash_lock;
 	struct hlist_head		hash[BR_HASH_SIZE];
 	unsigned long			feature_mask;

Re: [Bridge] [PATCH] bridge: per-cpu packet statistics

[Stephen Hemminger]

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.

Re: [Bridge] [PATCH] bridge: per-cpu packet statistics

[Stephen Hemminger]

On Tue, 02 Mar 2010 10:02:59 +0100
Eric Dumazet <eric.dumazet@gmail.com> wrote:

> From: Stephen Hemminger <shemminger@vyatta.com>
> 
> Le lundi 01 mars 2010 à 16:16 -0800, Stephen Hemminger a écrit :
> 
> > +	for_each_online_cpu(cpu) {
> > +		const struct br_cpu_netstats *bstats
> > +			= per_cpu_ptr(br->stats, cpu);
> > +
> > +		stats->rx_bytes   += bstats->rx_bytes;
> > +		stats->rx_packets += bstats->rx_packets;
> > +	}
> 
> And last point, we should use for_each_possible_cpu() here
> 
> Here is your patch with all my comments integrated :
> 
> 1) Use txq->{tx_bytes|tx_packets} counter
> 
> 2) alloc_percpu(struct ...) instead of alloc_percpu(sizeof(struct ...))
> 
> 3) free_netdev() in destructor
> 
> 4) for_each_possible_cpu() instead of for_each_online_cpu()
> 
> 5) br_get_stats() use local variables for the sake of concurrent users

No need, bridge has no queue!

[PATCH] bridge: per-cpu packet statistics (v2)

[Stephen Hemminger]

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>

---
Take Eric's bug fix changes. But keep the per cpu transmit
per cpu statistics.  There is no point in going multiqueue for
a pseudo-interface which is lockless and queue less.

 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	2010-03-02 09:26:23.208191713 -0800
+++ b/net/bridge/br_device.c	2010-03-02 09:29:38.811818410 -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 +82,31 @@ 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;
+	struct br_cpu_netstats sum = { 0 };
+	unsigned int cpu;
+
+	for_each_possible_cpu(cpu) {
+		const struct br_cpu_netstats *bstats
+			= per_cpu_ptr(br->stats, cpu);
+
+		sum.tx_bytes   += bstats->tx_bytes;
+		sum.tx_packets += bstats->tx_packets;
+		sum.rx_bytes   += bstats->rx_bytes;
+		sum.rx_packets += bstats->rx_packets;
+	}
+
+	stats->tx_bytes   = sum.tx_bytes;
+	stats->tx_packets = sum.tx_packets;
+	stats->rx_bytes   = sum.rx_bytes;
+	stats->rx_packets = sum.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 +206,28 @@ 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);
+	free_netdev(dev);
+}
+
 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	2010-03-02 09:26:23.188192219 -0800
+++ b/net/bridge/br_if.c	2010-03-02 09:31:19.801199305 -0800
@@ -185,6 +185,12 @@ static struct net_device *new_bridge_dev
 	br = netdev_priv(dev);
 	br->dev = dev;
 
+	br->stats = alloc_percpu(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	2010-03-02 09:26:23.196192436 -0800
+++ b/net/bridge/br_input.c	2010-03-02 09:27:04.424067504 -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	2010-03-02 09:26:23.180193181 -0800
+++ b/net/bridge/br_private.h	2010-03-02 09:27:04.424067504 -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;

Re: [PATCH] bridge: per-cpu packet statistics (v2)

[Eric Dumazet]

Le mardi 02 mars 2010 à 09:58 -0800, Stephen Hemminger a écrit :
> +	struct br_cpu_netstats __percpu {
> +		unsigned long	rx_packets;
> +		unsigned long	tx_packets;
> +		unsigned long	rx_bytes;
> +		unsigned long	tx_bytes;
> +	} *stats;
> +

Ok, please group rx stats together instead of interleave, this to reduce
possibility of touching two cache lines in tx path (or rx path)

	struct br_cpu_netstats __percpu {
> 		unsigned long	rx_packets;
> +		unsigned long	rx_bytes;
> 
> +		unsigned long	tx_packets;
> +		unsigned long	tx_bytes;
> +	} *stats;
> +

[Bridge] [PATCH] bridge: per-cpu packet statistics (v3)

[Stephen Hemminger]

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>

---
Move tx and rx to be next to each other per Eric's suggestion

 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	2010-03-02 10:48:44.527817663 -0800
+++ b/net/bridge/br_device.c	2010-03-02 10:48:48.287817348 -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 +82,31 @@ 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;
+	struct br_cpu_netstats sum = { 0 };
+	unsigned int cpu;
+
+	for_each_possible_cpu(cpu) {
+		const struct br_cpu_netstats *bstats
+			= per_cpu_ptr(br->stats, cpu);
+
+		sum.tx_bytes   += bstats->tx_bytes;
+		sum.tx_packets += bstats->tx_packets;
+		sum.rx_bytes   += bstats->rx_bytes;
+		sum.rx_packets += bstats->rx_packets;
+	}
+
+	stats->tx_bytes   = sum.tx_bytes;
+	stats->tx_packets = sum.tx_packets;
+	stats->rx_bytes   = sum.rx_bytes;
+	stats->rx_packets = sum.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 +206,28 @@ 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);
+	free_netdev(dev);
+}
+
 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	2010-03-02 10:48:44.507817575 -0800
+++ b/net/bridge/br_if.c	2010-03-02 10:48:48.287817348 -0800
@@ -185,6 +185,12 @@ static struct net_device *new_bridge_dev
 	br = netdev_priv(dev);
 	br->dev = dev;
 
+	br->stats = alloc_percpu(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	2010-03-02 10:48:44.515818544 -0800
+++ b/net/bridge/br_input.c	2010-03-02 10:48:48.287817348 -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	2010-03-02 10:48:44.503817627 -0800
+++ b/net/bridge/br_private.h	2010-03-02 10:49:10.632566819 -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	rx_bytes;
+		unsigned long	tx_packets;
+		unsigned long	tx_bytes;
+	} *stats;
+
 	spinlock_t			hash_lock;
 	struct hlist_head		hash[BR_HASH_SIZE];
 	unsigned long			feature_mask;

Re: [Bridge] [PATCH] bridge: per-cpu packet statistics (v3)

[Eric Dumazet]

Le mardi 02 mars 2010 à 15:32 -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>
> 

Thanks Stephen !

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>

Re: [Bridge] [PATCH] bridge: per-cpu packet statistics (v3)

[David Miller]

From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Wed, 03 Mar 2010 07:09:04 +0100

> Le mardi 02 mars 2010 à 15:32 -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>
>> 
> 
> Thanks Stephen !
> 
> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>

This is fine but needs to go into the next merge window, not
this one.

Bug fixes only now.

Re: [Bridge] [PATCH] bridge: per-cpu packet statistics

[Herbert Xu]

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,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [Bridge] [PATCH] bridge: per-cpu packet statistics

[Herbert Xu]

Stephen Hemminger <shemminger@vyatta.com> wrote:
>
> No need, bridge has no queue!

OK, ignore my previous comment :)
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [Bridge] [PATCH] bridge: per-cpu packet statistics (v3)

[David Miller]

From: Stephen Hemminger <shemminger@vyatta.com>
Date: Tue, 2 Mar 2010 15:32:09 -0800

> 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>

Applied.