On Wed, 29 May 2013 10:08:53 +0800
Cong Wang [off-list ref] wrote:

On Tue, 2013-05-28 at 08:22 -0700, Stephen Hemminger wrote:

quoted

On Tue, 28 May 2013 19:07:22 +0800
Cong Wang [off-list ref] wrote:

quoted

From: Cong Wang <redacted>

In commit 553675fb5e9ce3d71a (vxlan: listen on multiple ports),
we use kfree_rcu() to free ->vn_sock, but a) there is no use
of RCU API to access this filed, b) RCU is not enough to do refcnt
here, because in vxlan_leave_group() we drop RTNL lock before
locking the socket, it could be possible that this field is
freed during this period.

So, instead making things complex, just do basic refcnt for
the ->vn_sock, like we do for others.

...

quoted

Not needed all access is under RTNL

I know, this is why I had a patch (not posted) which adds the missing
rtnl_dereference(), but even if we had these, it is still not correct.

As I explained in the changelog, vxlan_leave_group() has a problem,
because it releases rtnl lock before locking the socket, _and_ it is
called after vxlan_dellink() which schedules a work to cleanup the
struct. Therefore the ->vn_sock could be freed right after rtnl lock is
released.

Am I miss anything?

Ignoring your IPv6 code for now...

With IPV4:
   refcnt is incremented when socket is incremented in newlink (RTNL held).
   refcnt is decremented in by dellink (RTNL held) and socket is deleted from list
   leave_group doesn't happen until work queue is fired.

rtnl_dereference is fine, but hardly necessary when the call hierarchy is so obvious.

The problem you describe won't be fixed by just converting it to atomic,
I think you need add a dev_hold()/dev_put to vxlan_stop to prevent
device from being deleted when rtnl_lock is dropped.