Re: [RFC 2/4] net: netfilter conntrack - add per-net functionality for DCCP protocol

[Pavel Emelyanov <xemul@openvz.org>]

Cyrill Gorcunov wrote:
> On Tue, Mar 10, 2009 at 1:33 PM, Daniel Lezcano <daniel.lezcano@free.fr> wrote:
>> Daniel Lezcano wrote:
>>> Cyrill Gorcunov wrote:
>>>> Module specific data moved into per-net site and being allocated/freed
>>>> during net namespace creation/deletion.
>>>>
>>>> Signed-off-by: Cyrill Gorcunov <gorcunov@openvz.org>
>>>> ---
>>>>  net/netfilter/nf_conntrack_proto_dccp.c |  148
>>>> ++++++++++++++++++++++++--------
>>>>  1 file changed, 111 insertions(+), 37 deletions(-)
>>>>
>>>>
>>>>   static int __init nf_conntrack_proto_dccp_init(void)
>>>>  {
>>>>     int err;
>>>>  -    err = nf_conntrack_l4proto_register(&dccp_proto4);
>>>> +    err = register_pernet_gen_device(&dccp_net_id, &dccp_net_ops);
>>>>
>> [ cut ]
>>> Shouldn't it be register_pernet_gen_subsys ?
> 
> No, I believe. By using  register_pernet_gen_device I'm allowed to
> not modify 'struct net' and friends and keep all I need in my own
> pointer retrieved thru per-net gen-device id I've registered.

I believe Daniel means, that we need the register_xxx_get_subsys call
for subsystems, rather than devices, that will behave according to the
generic net pointers.

Daniel, am I right with this suggestion?

>> If you use register_pernet_gen_device, your subsystem will be deleted before
>> the network devices and potentially you can receive a packet even if your
>> subsystem is already freed.
>>
>> Eric did a fix for tcp and icmp a few weeks ago. I thing its explanation is
>> better than mine :)
>> it is the commit 6eb0777228f31932fc941eafe8b08848466630a1 for net-2.6
>>
>> Thanks.
>>  -- Daniel
>>
> 
> Thanks a lot Daniel, will check!
>