ipv4: Cap ADVMSS metric in the FIB rather than the routing cache.

It makes no sense to execute this limit test every time we create a
routing cache entry.

We can't simply error out on these things since we've silently
accepted and truncated them forever.

Signed-off-by: David S. Miller <davem@davemloft.net>
diff --git a/net/ipv4/fib_semantics.c b/net/ipv4/fib_semantics.c
index e5b7182..415f823 100644
--- a/net/ipv4/fib_semantics.c
+++ b/net/ipv4/fib_semantics.c
@@ -779,9 +779,14 @@
 			int type = nla_type(nla);
 
 			if (type) {
+				u32 val;
+
 				if (type > RTAX_MAX)
 					goto err_inval;
-				fi->fib_metrics[type - 1] = nla_get_u32(nla);
+				val = nla_get_u32(nla);
+				if (type == RTAX_ADVMSS && val > 65535 - 40)
+					val = 65535 - 40;
+				fi->fib_metrics[type - 1] = val;
 			}
 		}
 	}