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State | Superseded, archived |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