veth: Fix veth_get_stats()

veth_get_stats() can be called in parallel on several cpus.

It's better to not reset dev->stats as it could give wrong result on
one cpu. Use temporary variables, then store the final results.

Also, we should loop on every possible cpus, not only online cpus,
or cpu hotplug can suddenly give wrong veth stats.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/veth.c

@@ -210,32 +210,29 @@ rx_drop:
 static struct net_device_stats *veth_get_stats(struct net_device *dev)
 {
 	struct veth_priv *priv;
-	struct net_device_stats *dev_stats;
 	int cpu;
-	struct veth_net_stats *stats;
+	struct veth_net_stats *stats, total = {0};
 
 	priv = netdev_priv(dev);
-	dev_stats = &dev->stats;
-
-	dev_stats->rx_packets = 0;
-	dev_stats->tx_packets = 0;
-	dev_stats->rx_bytes = 0;
-	dev_stats->tx_bytes = 0;
-	dev_stats->tx_dropped = 0;
-	dev_stats->rx_dropped = 0;
 
-	for_each_online_cpu(cpu) {
+	for_each_possible_cpu(cpu) {
 		stats = per_cpu_ptr(priv->stats, cpu);
 
-		dev_stats->rx_packets += stats->rx_packets;
-		dev_stats->tx_packets += stats->tx_packets;
-		dev_stats->rx_bytes += stats->rx_bytes;
-		dev_stats->tx_bytes += stats->tx_bytes;
-		dev_stats->tx_dropped += stats->tx_dropped;
-		dev_stats->rx_dropped += stats->rx_dropped;
+		total.rx_packets += stats->rx_packets;
+		total.tx_packets += stats->tx_packets;
+		total.rx_bytes   += stats->rx_bytes;
+		total.tx_bytes   += stats->tx_bytes;
+		total.tx_dropped += stats->tx_dropped;
+		total.rx_dropped += stats->rx_dropped;
 	}
-
-	return dev_stats;
+	dev->stats.rx_packets = total.rx_packets;
+	dev->stats.tx_packets = total.tx_packets;
+	dev->stats.rx_bytes   = total.rx_bytes;
+	dev->stats.tx_bytes   = total.tx_bytes;
+	dev->stats.tx_dropped = total.tx_dropped;
+	dev->stats.rx_dropped = total.rx_dropped;
+
+	return &dev->stats;
 }
 
 static int veth_open(struct net_device *dev)