On Sat, Jan 30, 2010 at 1:53 PM, Alexey Dobriyan [off-list ref] wrote:

On Fri, Jan 29, 2010 at 05:33:26PM +0100, Eric Dumazet wrote:

quoted

@@ -3807,21 +3807,24 @@ static int __init ipsec_pfkey_init(void)

      if (err != 0)
              goto out;

-     err = sock_register(&pfkey_family_ops);
-     if (err != 0)
-             goto out_unregister_key_proto;
      err = xfrm_register_km(&pfkeyv2_mgr);
      if (err != 0)
-             goto out_sock_unregister;
+             goto out_unregister_key_proto;
+
      err = register_pernet_subsys(&pfkey_net_ops);
      if (err != 0)
              goto out_xfrm_unregister_km;
+
+     err = sock_register(&pfkey_family_ops);
+     if (err != 0)
+             goto out_unregister_pernet;
 out:
      return err;
+
+out_unregister_pernet:
+     unregister_pernet_subsys(&pfkey_net_ops);
 out_xfrm_unregister_km:
      xfrm_unregister_km(&pfkeyv2_mgr);
-out_sock_unregister:
-     sock_unregister(PF_KEY);
 out_unregister_key_proto:
      proto_unregister(&key_proto);
      goto out;

ACK analysis, except this is not enough.

Here is patch which survived netns start/stop/modprobe/rmmod cycles.

       Alexey, who still doesn't get why bug reproduces so easily for bug reporter.

Luca, please confirm.

Seems to work fine.

[PATCH] af_key: fix netns ops ordering on module load/unload

1. After sock_register() returns, it's possible to create sockets,
  even if module still not initialized fully (blame generic module code
  for that!)
2. Consequently, pfkey_create() can be called with pfkey_net_id still not
  initialized which will BUG_ON in net_generic():
       kernel BUG at include/net/netns/generic.h:43!
3. During netns shutdown, netns ops should be unregistered after
  key manager unregistered because key manager calls can be triggered
  from xfrm_user module:

       general protection fault: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC
       pfkey_broadcast+0x111/0x210 [af_key]
       pfkey_send_notify+0x16a/0x300 [af_key]
       km_state_notify+0x41/0x70
       xfrm_flush_sa+0x75/0x90 [xfrm_user]
4. Unregister netns ops after socket ops just in case and for symmetry.

Reported by Luca Tettamanti.

Signed-off-by: Alexey Dobriyan <redacted>

Tested-by: Luca Tettamanti <kronos.it@gmail.com>

thanks,
Luca