[NETNS]: icmp(v6)_sk should not pin a namespace.

So, change icmp(v6)_sk creation/disposal to the scheme used in the
netlink for rtnl, i.e. create a socket in the context of the init_net
and assign the namespace without getting a referrence later.

Also use sk_release_kernel instead of sock_release to properly destroy
such sockets.

Signed-off-by: Denis V. Lunev <den@openvz.org>
Acked-by: Daniel Lezcano <dlezcano@fr.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
diff --git a/net/ipv4/icmp.c b/net/ipv4/icmp.c
index 7c62a0d..97d97ad 100644
--- a/net/ipv4/icmp.c
+++ b/net/ipv4/icmp.c
@@ -1146,14 +1146,8 @@
 {
 	int i;
 
-	for_each_possible_cpu(i) {
-		struct sock *sk;
-
-		sk = __icmp_sk[i];
-		if (sk == NULL)
-			continue;
-		sock_release(sk->sk_socket);
-	}
+	for_each_possible_cpu(i)
+		sk_release_kernel(__icmp_sk[i]);
 	kfree(__icmp_sk);
 	__icmp_sk = NULL;
 }
@@ -1176,6 +1170,8 @@
 			goto fail;
 
 		__icmp_sk[i] = sk = sock->sk;
+		sk_change_net(sk, &init_net);
+
 		sk->sk_allocation = GFP_ATOMIC;
 
 		/* Enough space for 2 64K ICMP packets, including