The comment about "race free view of the set of network
namespaces" was a bit hasty. Look (there even can be only
one CPU, as discovered by Alexey Dobriyan and Denis Lunev):

put_net()
  if (atomic_dec_and_test(&net->refcnt))
    /* true */
      __put_net(net);
        queue_work(...);

/*
 * note: the net now has refcnt 0, but still in
 * the global list of net namespaces
 */

== re-schedule ==

register_pernet_subsys(&some_ops);
  register_pernet_operations(&some_ops);
    (*some_ops)->init(net);
      /*
       * we call netlink_kernel_create() here
       * in some places
       */
      netlink_kernel_create();
         sk_alloc();
            get_net(net); /* refcnt = 1 */
         /*
          * now we drop the net refcount not to
          * block the net namespace exit in the
          * future (or this can be done on the
          * error path)
          */
         put_net(sk->sk_net);
             if (atomic_dec_and_test(&...))
                   /*
                    * true. BOOOM! The net is
                    * scheduled for release twice
                    */

When thinking on this problem, I decided, that getting and
putting the net in init callback is wrong. If some init
callback needs to have a refcount-less reference on the struct
net, _it_ has to be careful himself, rather than relying on
the infrastructure to handle this correctly.

In case of netlink_kernel_create(), the problem is that the
sk_alloc() gets the given namespace, but passing the info
that we don't want to get it inside this call is too heavy.

Instead, I propose to crate the socket inside an init_net
namespace and then re-attach it to the desired one right
after the socket is created.

After doing this, we also have to be careful on error paths
not to drop the reference on the namespace, we didn't get
the one on.
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Acked-by: Denis Lunev <den@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>