[Devel] Re: [PATCH] [RFC] network namespace locking rules

[Daniel Lezcano]

Denis V. Lunev wrote:
> 
> ------------------------------------------------------------------------
> 
> Subject:
> Re: [PATCH] [RFC] network namespace locking rules
> From:
> ebiederm at xmission.com (Eric W. Biederman)
> Date:
> Thu, 27 Sep 2007 02:45:34 -0600
> To:
> "Denis V. Lunev" <den at sw.ru>
> 
> To:
> "Denis V. Lunev" <den at sw.ru>
> CC:
> Daniel Lezcano <dlezcano at fr.ibm.com>, Benjamin Thery 
> <benjamin.thery at bull.net>, Kirill Korotaev <dev at sw.ru>
> 
> 
> "Denis V. Lunev" <den at sw.ru> writes:
> 
>> Hello, Eric!
>>
>> Unfortunately, I was wrong that your patch is sane :(
>>
>> It breaks current RTNL socket semantic. Namely, current code relies that
>> - netlink from user-space is queued to RTNL socket if RTNL lock is held
>> - all pending messages in that queue will be processed in rtnl_unlock
> 
> I know we come very close to this but I have a hard time seeing
> this being guaranteed.  We don't hold a lock so I think it is
> possible for a new message to come in via another path on SMP,
> and we miss it in rtnl_unlock.  Although missing that message
> from both paths that grabs rtnl_lock sounds unlikely.
> 

Thanks, I missed this one.
it makes more sense now :)