Hi Daniel,

On 31.03.2016 23:52, Daniel Borkmann wrote:

On 03/31/2016 09:48 PM, Hannes Frederic Sowa wrote:
[...]

quoted

Tightest solution would probably be to combine both patches.

bool called_by_tuntap;

old_fp = rcu_dereference_protected(sk->sk_filter, called_by_tuntap ?
lockdep_rtnl_is_held() : lockdep_sock_is_held());

If I understand you correctly with combining them, you mean you'd still
need the API change to pass the bool 'called_by_tuntap' down, right?

I actually decided to simply lock the sockets. So that there will always 
be a clear lock owner during the updates. I think this is cleaner. What 
do you think?

If so, your main difference is, after all, to replace the
sock_owned_by_user()
with the lockdep_sock_is_held() construction instead, correct?

I just didn't do that part because we hold socket lock now.

But then, isn't it already sufficient when you pass the bool itself down
that 'demuxes' in this case between the sock_owned_by_user() vs
lockdep_rtnl_is_held() check?

I just send out the patches, please have a look.

Thanks,
Hannes