vlan: Precise RX stats accounting
authorEric Dumazet <eric.dumazet@gmail.com>
Tue, 17 Nov 2009 04:53:09 +0000 (04:53 +0000)
committerDavid S. Miller <davem@davemloft.net>
Wed, 18 Nov 2009 07:51:55 +0000 (23:51 -0800)
With multi queue devices, its possible that several cpus call
vlan RX routines simultaneously for the same vlan device.

We update RX stats counter without any locking, so we can
get slightly wrong counters.

One possible fix is to use percpu counters, to get precise
accounting and also get guarantee of no cache line ping pongs
between cpus.

Note: this adds 16 bytes (32 bytes on 64bit arches) of percpu
data per vlan device.

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/8021q/vlan.h
net/8021q/vlan_core.c
net/8021q/vlan_dev.c

index 68f9290e68373e5977fb526275dedd9ab5bc1b57..5685296017e903a2b8ef93cb285e96980f875b81 100644 (file)
@@ -16,6 +16,21 @@ struct vlan_priority_tci_mapping {
        struct vlan_priority_tci_mapping        *next;
 };
 
+
+/**
+ *     struct vlan_rx_stats - VLAN percpu rx stats
+ *     @rx_packets: number of received packets
+ *     @rx_bytes: number of received bytes
+ *     @multicast: number of received multicast packets
+ *     @rx_errors: number of errors
+ */
+struct vlan_rx_stats {
+       unsigned long rx_packets;
+       unsigned long rx_bytes;
+       unsigned long multicast;
+       unsigned long rx_errors;
+};
+
 /**
  *     struct vlan_dev_info - VLAN private device data
  *     @nr_ingress_mappings: number of ingress priority mappings
@@ -29,6 +44,7 @@ struct vlan_priority_tci_mapping {
  *     @dent: proc dir entry
  *     @cnt_inc_headroom_on_tx: statistic - number of skb expansions on TX
  *     @cnt_encap_on_xmit: statistic - number of skb encapsulations on TX
+ *     @vlan_rx_stats: ptr to percpu rx stats
  */
 struct vlan_dev_info {
        unsigned int                            nr_ingress_mappings;
@@ -45,6 +61,7 @@ struct vlan_dev_info {
        struct proc_dir_entry                   *dent;
        unsigned long                           cnt_inc_headroom_on_tx;
        unsigned long                           cnt_encap_on_xmit;
+       struct vlan_rx_stats                    *vlan_rx_stats;
 };
 
 static inline struct vlan_dev_info *vlan_dev_info(const struct net_device *dev)
index 971d3755ae874ee5b37c4a11e75addc10c842929..e75a2f3b10af7721610af6adf0c42e5c7fec2192 100644 (file)
@@ -31,7 +31,7 @@ EXPORT_SYMBOL(__vlan_hwaccel_rx);
 int vlan_hwaccel_do_receive(struct sk_buff *skb)
 {
        struct net_device *dev = skb->dev;
-       struct net_device_stats *stats;
+       struct vlan_rx_stats     *rx_stats;
 
        skb->dev = vlan_dev_info(dev)->real_dev;
        netif_nit_deliver(skb);
@@ -40,15 +40,17 @@ int vlan_hwaccel_do_receive(struct sk_buff *skb)
        skb->priority = vlan_get_ingress_priority(dev, skb->vlan_tci);
        skb->vlan_tci = 0;
 
-       stats = &dev->stats;
-       stats->rx_packets++;
-       stats->rx_bytes += skb->len;
+       rx_stats = per_cpu_ptr(vlan_dev_info(dev)->vlan_rx_stats,
+                              smp_processor_id());
+
+       rx_stats->rx_packets++;
+       rx_stats->rx_bytes += skb->len;
 
        switch (skb->pkt_type) {
        case PACKET_BROADCAST:
                break;
        case PACKET_MULTICAST:
-               stats->multicast++;
+               rx_stats->multicast++;
                break;
        case PACKET_OTHERHOST:
                /* Our lower layer thinks this is not local, let's make sure.
index 9159659421395fc22a675f837238763f81083e93..de0dc6bacbe817b2074dfa45010ed1b7c8e64357 100644 (file)
@@ -140,7 +140,7 @@ int vlan_skb_recv(struct sk_buff *skb, struct net_device *dev,
                  struct packet_type *ptype, struct net_device *orig_dev)
 {
        struct vlan_hdr *vhdr;
-       struct net_device_stats *stats;
+       struct vlan_rx_stats *rx_stats;
        u16 vlan_id;
        u16 vlan_tci;
 
@@ -163,9 +163,10 @@ int vlan_skb_recv(struct sk_buff *skb, struct net_device *dev,
                goto err_unlock;
        }
 
-       stats = &skb->dev->stats;
-       stats->rx_packets++;
-       stats->rx_bytes += skb->len;
+       rx_stats = per_cpu_ptr(vlan_dev_info(dev)->vlan_rx_stats,
+                              smp_processor_id());
+       rx_stats->rx_packets++;
+       rx_stats->rx_bytes += skb->len;
 
        skb_pull_rcsum(skb, VLAN_HLEN);
 
@@ -180,7 +181,7 @@ int vlan_skb_recv(struct sk_buff *skb, struct net_device *dev,
                break;
 
        case PACKET_MULTICAST:
-               stats->multicast++;
+               rx_stats->multicast++;
                break;
 
        case PACKET_OTHERHOST:
@@ -200,7 +201,7 @@ int vlan_skb_recv(struct sk_buff *skb, struct net_device *dev,
 
        skb = vlan_check_reorder_header(skb);
        if (!skb) {
-               stats->rx_errors++;
+               rx_stats->rx_errors++;
                goto err_unlock;
        }
 
@@ -731,6 +732,11 @@ static int vlan_dev_init(struct net_device *dev)
                subclass = 1;
 
        vlan_dev_set_lockdep_class(dev, subclass);
+
+       vlan_dev_info(dev)->vlan_rx_stats = alloc_percpu(struct vlan_rx_stats);
+       if (!vlan_dev_info(dev)->vlan_rx_stats)
+               return -ENOMEM;
+
        return 0;
 }
 
@@ -740,6 +746,8 @@ static void vlan_dev_uninit(struct net_device *dev)
        struct vlan_dev_info *vlan = vlan_dev_info(dev);
        int i;
 
+       free_percpu(vlan->vlan_rx_stats);
+       vlan->vlan_rx_stats = NULL;
        for (i = 0; i < ARRAY_SIZE(vlan->egress_priority_map); i++) {
                while ((pm = vlan->egress_priority_map[i]) != NULL) {
                        vlan->egress_priority_map[i] = pm->next;
@@ -775,6 +783,31 @@ static u32 vlan_ethtool_get_flags(struct net_device *dev)
        return dev_ethtool_get_flags(vlan->real_dev);
 }
 
+static struct net_device_stats *vlan_dev_get_stats(struct net_device *dev)
+{
+       struct net_device_stats *stats = &dev->stats;
+
+       dev_txq_stats_fold(dev, stats);
+
+       if (vlan_dev_info(dev)->vlan_rx_stats) {
+               struct vlan_rx_stats *p, rx = {0};
+               int i;
+
+               for_each_possible_cpu(i) {
+                       p = per_cpu_ptr(vlan_dev_info(dev)->vlan_rx_stats, i);
+                       rx.rx_packets += p->rx_packets;
+                       rx.rx_bytes   += p->rx_bytes;
+                       rx.rx_errors  += p->rx_errors;
+                       rx.multicast  += p->multicast;
+               }
+               stats->rx_packets = rx.rx_packets;
+               stats->rx_bytes   = rx.rx_bytes;
+               stats->rx_errors  = rx.rx_errors;
+               stats->multicast  = rx.multicast;
+       }
+       return stats;
+}
+
 static const struct ethtool_ops vlan_ethtool_ops = {
        .get_settings           = vlan_ethtool_get_settings,
        .get_drvinfo            = vlan_ethtool_get_drvinfo,
@@ -797,6 +830,7 @@ static const struct net_device_ops vlan_netdev_ops = {
        .ndo_change_rx_flags    = vlan_dev_change_rx_flags,
        .ndo_do_ioctl           = vlan_dev_ioctl,
        .ndo_neigh_setup        = vlan_dev_neigh_setup,
+       .ndo_get_stats          = vlan_dev_get_stats,
 #if defined(CONFIG_FCOE) || defined(CONFIG_FCOE_MODULE)
        .ndo_fcoe_ddp_setup     = vlan_dev_fcoe_ddp_setup,
        .ndo_fcoe_ddp_done      = vlan_dev_fcoe_ddp_done,
@@ -820,6 +854,7 @@ static const struct net_device_ops vlan_netdev_accel_ops = {
        .ndo_change_rx_flags    = vlan_dev_change_rx_flags,
        .ndo_do_ioctl           = vlan_dev_ioctl,
        .ndo_neigh_setup        = vlan_dev_neigh_setup,
+       .ndo_get_stats          = vlan_dev_get_stats,
 #if defined(CONFIG_FCOE) || defined(CONFIG_FCOE_MODULE)
        .ndo_fcoe_ddp_setup     = vlan_dev_fcoe_ddp_setup,
        .ndo_fcoe_ddp_done      = vlan_dev_fcoe_ddp_done,