neighbor: Call __ipv4_neigh_lookup_noref in neigh_xmit

Commit cd9ff4d changed the key for IFF_POINTOPOINT devices to
INADDR_ANY but neigh_xmit which is used for MPLS encapsulations was not
updated to use the altered key. The result is that every packet Tx does
a lookup on the gateway address which does not find an entry, a new one
is created only to find the existing one in the table right before the
insert since arp_constructor was updated to reset the primary key. This
is seen in the allocs and destroys counters:
    ip -s -4 ntable show | head -10 | grep alloc

which increase for each packet showing the unnecessary overhread.

Fix by having neigh_xmit use __ipv4_neigh_lookup_noref for NEIGH_ARP_TABLE.

Fixes: cd9ff4d ("ipv4: Make neigh lookup keys for loopback/point-to-point devices be INADDR_ANY")
Reported-by: Alan Maguire <alan.maguire@oracle.com>
Signed-off-by: David Ahern <dsahern@gmail.com>
Tested-by: Alan Maguire <alan.maguire@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

net/core/neighbour.c

@@ -31,6 +31,7 @@
 #include <linux/times.h>
 #include <net/net_namespace.h>
 #include <net/neighbour.h>
+#include <net/arp.h>
 #include <net/dst.h>
 #include <net/sock.h>
 #include <net/netevent.h>
@@ -2984,7 +2985,13 @@ int neigh_xmit(int index, struct net_device *dev,
 		if (!tbl)
 			goto out;
 		rcu_read_lock_bh();
-		neigh = __neigh_lookup_noref(tbl, addr, dev);
+		if (index == NEIGH_ARP_TABLE) {
+			u32 key = *((u32 *)addr);
+
+			neigh = __ipv4_neigh_lookup_noref(dev, key);
+		} else {
+			neigh = __neigh_lookup_noref(tbl, addr, dev);
+		}
 		if (!neigh)
 			neigh = __neigh_create(tbl, addr, dev, false);
 		err = PTR_ERR(neigh);