Re: [PATCH net-next-2.6] netns: call ops_free right after ops_exit

[Jiri Pirko <jpirko@redhat.com>]

Mon, Apr 26, 2010 at 05:06:07AM CEST, ebiederm@xmission.com wrote:
>Jiri Pirko <jpirko@redhat.com> writes:
>
>> Sun, Apr 25, 2010 at 04:50:34PM CEST, ebiederm@xmission.com wrote:
>>>David Miller <davem@davemloft.net> writes:
>>>
>>>> From: Jiri Pirko <jpirko@redhat.com>
>>>> Date: Sun, 25 Apr 2010 11:26:01 +0200
>>>>
>>>>> There's no need to iterate this twice. We can free net generic
>>>>> variables right after exit is called.
>>>>>
>>>>> Signed-off-by: Jiri Pirko <jpirko@redhat.com>
>>>>
>>>> Are you sure there are no problems with doing this?
>>>>
>>>> What if there are inter-net variable reference dependencies
>>>> or something like that?
>>>>
>>>> I really suspect it is being done this way on purpose, but
>>>> in the end I defer to experts like Eric B. :-)
>>>
>>>I am pretty certain there is a problem.  My memory is fuzzy this
>>>morning but I believe we can have rcu references between various
>>>pieces of the networking stack for a single network namespace.  So we
>>>need to cause all of the network namespace to exit before it is safe
>>>to free those pieces.
>>
>> Hmm, that doesn't make much sense to me. Since the allocated memory in question
>> is used locally, after exit() is called, the memory chunk should not be used by
>> anyone and if it is, I think it's a bug.
>>
>> Earlier, when the memory wasn't allocated automatically (by filling .size)
>> memory was individually freed in exit(). From what I understood from your reply,
>> you are telling this was buggy?
>
>Now I remember clearly.  The use case for not freeing memory
>immediately is the delayed freeing of network devices.  Ideally we
>delay unregistering all of the network devices into
>default_device_exit_batch, when we terminate one or namespaces.
>
>Since network devices are freed outside of their exit routines we need
>to keep their per net memory around until they are freed.
>
>Ultimately all of this is much easier to think about if these chunks of
>memory can be logically thought of as living on struct net and have the
>same lifetime rules.

Ok, I see your point. So scratch this.

Thanks,

	Jirka
>
>Eric