[Devel] [PATCH 10/12] L2 network namespace: playing with pass-through device

Daniel Lezcano daniel.lezcano at free.fr
Wed Dec 13 01:36:25 PST 2006


Herbert Poetzl wrote:
> On Tue, Dec 12, 2006 at 04:50:50PM +0100, Daniel Lezcano wrote:
>> Dmitry Mishin wrote:
>>> On Tuesday 12 December 2006 17:19, Daniel Lezcano wrote:
>>>> Dmitry Mishin wrote:
>>>>
>>>>>>>> Why do yo need to have a child list and sibling list ?
>>>>>>> Because of the level2<->level3 hierarchy, for example.
>>>>>> This hierarchy doesn't exist with ns->parent ? Do you have an example
>>>>>> when the hierarchy should be used ? I mean when we need to browse from
>>>>>> l2 -> l3 ?
>>>>> For example, to check that new ifaddr is already used by child l3 namespace.
>>>> The devinet isolation does already do that, you can not add a new ifaddr 
>>>> if it already exists. Do you have another example ?
>>> Could devinet isolation provide ifaddrs list with namespaces?
>>> What will be with child namespaces if you decide to destroy parent namespace?
>>> If we decide to destroy them, than how we could get their list?
>>> It is a question of flexibility and easy management.
>>> Why do you want to remove this code? 
>> I don't want to especially remove this code, I just want to understand 
>> what it does and why. If it appears to be useless, let's remove it, if 
>> it appears to be useful, let's keep it.
>>
>> By the way, what is the meaning on destroying the namespaces directly, 
>> is it not the kref mechanism which needs to do that ? For example, if 
>> you  create a l2 namespace and after you create l3 namespaces. You want 
>> to destroy the l2 namespace, the l2 namespace should stay "zombie" until 
>> all the l3 namespaces exit. If you need to wipe out all the namespaces, 
>> you should destroy all the related namespaces' ressources, like killing 
>> all processes inside it. The namespaces will "put" their respective kref 
>> and will trigger the freeing of the ressources.
> 
> networking (mostly sockets) will probably require
> some mechanism to 'zap' them, ignoring the defined
> timeouts. otherwise the spaces could hang around
> for quite a while waiting for some response, which
> might never come ...

Yes, exact. We will need a specific socket cleanup by namespace in order 
  to do network migration. This is the only case I see to 'zap' the sockets.
The sockets should never be flushed in other cases. For example, you 
launch an application into a network namespace, it sends 10MB to a peer 
and exits. The network namespace should stay "alive" until all orphans 
sockets have flushed their buffers to the peer. This behavior is 
perfectly handled by the kref mechanism because sock_release will "put" 
the network namespace and that will trigger the network namespace 
destruction.

> but that should not be _that_ important right now

I think this should be addressed later for the network checkpoint/restart.



More information about the Containers mailing list