Andrei Vagin [off-list ref] writes:

On Thu, Jul 21, 2016 at 9:40 AM, Eric W. Biederman [off-list ref] wrote:

quoted

index 2c2eb1b629b1..a489f192d619 100644

--- a/net/core/net_namespace.c
+++ b/net/core/net_namespace.c

@@ -266,6 +266,16 @@ struct net *get_net_ns_by_id(struct net *net, int id)
        return peer;
 }

+static bool inc_net_namespaces(struct user_namespace *ns)
+{
+       return inc_ucount(ns, UCOUNT_NET_NAMESPACES);
+}
+
+static void dec_net_namespaces(struct user_namespace *ns)
+{
+       dec_ucount(ns, UCOUNT_NET_NAMESPACES);
+}
+
 /*
  * setup_net runs the initializers for the network namespace object.
  */

@@ -276,6 +286,9 @@ static __net_init int setup_net(struct net *net, struct user_namespace *user_ns)
        int error = 0;
        LIST_HEAD(net_exit_list);

+       if (!inc_net_namespaces(user_ns))
+               return -ENFILE;

I think you need to move this check after initilizing  net->passive.
When setup_net returns an error, net_drop_ns is called:

Good point.  Ouch!

void net_drop_ns(void *p)
{
        struct net *ns = p;
        if (ns && atomic_dec_and_test(&ns->passive))
                net_free(ns);
}

Actually, I think it would be better to make this check before
net_alloc().

You are probably right.  I seem to be trying to be entirely too clever
putting that in setup_net so I can cover the initial network namespace.
Which really does not need to be counted.  As clearly I also goofed up
the decrement on error case as well.

Eric