vxlan: fix race between flush and incoming learning

It is possible for a packet to arrive during vxlan_stop(), and
have a dynamic entry created. Close this by checking if device
is up.

 CPU1                             CPU2
vxlan_stop
  vxlan_flush
     hash_lock acquired
                                  vxlan_encap_recv
                                     vxlan_snoop
                                        waiting for hash_lock
     hash_lock relased
  vxlan_flush done
                                        hash_lock acquired
                                        vxlan_fdb_create

This is a day-one bug in vxlan goes back to 3.7.
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/vxlan.c

@@ -571,7 +571,6 @@ static void vxlan_snoop(struct net_device *dev,
 {
 	struct vxlan_dev *vxlan = netdev_priv(dev);
 	struct vxlan_fdb *f;
-	int err;
 
 	f = vxlan_find_mac(vxlan, src_mac);
 	if (likely(f)) {
@@ -588,7 +587,10 @@ static void vxlan_snoop(struct net_device *dev,
 	} else {
 		/* learned new entry */
 		spin_lock(&vxlan->hash_lock);
-		err = vxlan_fdb_create(vxlan, src_mac, src_ip,
+
+		/* close off race between vxlan_flush and incoming packets */
+		if (netif_running(dev))
+			vxlan_fdb_create(vxlan, src_mac, src_ip,
 					 NUD_REACHABLE,
 					 NLM_F_EXCL|NLM_F_CREATE,
 					 vxlan->dst_port,