Re: That's pretty much it for 3.5.0

[John Fastabend <john.r.fastabend@intel.com>]

On 7/17/2012 2:02 PM, David Miller wrote:
> From: John Fastabend <john.r.fastabend@intel.com>
> Date: Tue, 17 Jul 2012 13:50:16 -0700
>
>> On 7/17/2012 12:24 PM, David Miller wrote:
>>> From: John Fastabend <john.r.fastabend@intel.com>
>>> Date: Tue, 17 Jul 2012 12:09:53 -0700
>>>
>>>> although we don't have an early_init hook for netprio_cgroup so this
>>>> is probably not correct.
>>>
>>> The dependency is actually on net_dev_init (a subsys_initcall) rather
>>> than a pure_initcall.
>>>
>>> net_dev_init is what registers the netdev_net_ops, which in turn
>>> initializes the netdev list in namespaces such as &init_net
>>>
>>
>> Ah right thanks sorry for the thrash. I guess we need to check if the
>> netdev list in the init_net namespace is initialized.
>
> It's a hack, but we could export and then test dev_boot_phase == 0,
> and if that test is true then skip the init_net device walk in the
> cgroup code.
>
> But I don't like that very much.
>
> The things this code cares about can't even be an issue until
> net_dev_init() runs.
>
> There is a comment warning not to do this in linux/init.h, but we
> could change the module_init() in netprio_cgroup.c to some level which
> runs after subsys_inticall().  When built as a module, linux/init.h
> will translate this into module_init() which is basically the behavior
> we want.
>

Perhaps the easiest way is to check net->count this should be zero
until setup_net is called.

if (!atomic_read(&init_net.count))
	return ret;