[NETNS]: Namespace stop vs 'ip r l' race.

During network namespace stop process kernel side netlink sockets
belonging to a namespace should be closed. They should not prevent
namespace to stop, so they do not increment namespace usage
counter. Though this counter will be put during last sock_put.

The raplacement of the correct netns for init_ns solves the problem
only partial as socket to be stoped until proper stop is a valid
netlink kernel socket and can be looked up by the user processes. This
is not a problem until it resides in initial namespace (no processes
inside this net), but this is not true for init_net.

So, hold the referrence for a socket, remove it from lookup tables and
only after that change namespace and perform a last put.

Signed-off-by: Denis V. Lunev <den@openvz.org>
Tested-by: Alexey Dobriyan <adobriyan@openvz.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
diff --git a/net/netlink/af_netlink.c b/net/netlink/af_netlink.c
index 626a582..6b178e1 100644
--- a/net/netlink/af_netlink.c
+++ b/net/netlink/af_netlink.c
@@ -1396,6 +1396,9 @@
 	}
 	netlink_table_ungrab();
 
+	/* Do not hold an extra referrence to a namespace as this socket is
+	 * internal to a namespace and does not prevent it to stop. */
+	put_net(net);
 	return sk;
 
 out_sock_release:
@@ -1411,7 +1414,19 @@
 {
 	if (sk == NULL || sk->sk_socket == NULL)
 		return;
+
+	/*
+	 * Last sock_put should drop referrence to sk->sk_net. It has already
+	 * been dropped in netlink_kernel_create. Taking referrence to stopping
+	 * namespace is not an option.
+	 * Take referrence to a socket to remove it from netlink lookup table
+	 * _alive_ and after that destroy it in the context of init_net.
+	 */
+	sock_hold(sk);
 	sock_release(sk->sk_socket);
+
+	sk->sk_net = get_net(&init_net);
+	sock_put(sk);
 }
 EXPORT_SYMBOL(netlink_kernel_release);