Re: [PATCH 1/6] cgroup: initialize cgrp->dentry before css_alloc()

[Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>]

On 2013/1/23 8:27, Tejun Heo wrote:
> (cc'ing Andrew as scheduler folks are difficult to get response from
>  these days and I can't think of anyone else to bother :)
> 
> Hello, Li.
> 
> The cgroup part looks good to me but it would be great if the
> descriptions are more detailed, especially, about why the change is
> beneficial or what it's aiming at.  I take it that the shed changes
> are necessary to facilitate the later cgroup changes?  Can you please
> elaborate how?
> 

Scheduler changes made in patch #2 are preparation for #3 and #4.

- for #3:

After ss->css_alloc(), there's a small window that tg->css.cgroup is NULL.
With the change, tg won't be seen before ss->css_online(), so the scheduler
won't see NULL tg->css.cgroup.

- for #4:

If tg is unregistered and removed from global list in css_free(), and if
we kfree cgroup right after ss->css_free(), the scheduler can access
invalid tg->css.cgroup, because tg is also protected by RCU.

Without patch #2:

cgroup_rmdir()
  no ss->css_offline()
diput()
  syncronize_rcu()
  ss->css_free()    <-- remove tg from global list,
			and free tg via call_rcu()
  kfree_rcu(cgroup) <-- wait rcu read section

With the change:

cgroup_rmdir()
  ss->css_offline() <-- remove tg from global list
diput()
  synchronize_rcu() <-- wait rcu read section
  ss->css_free();   <-- free tg
  kfree(cgroup);

> The scheduler part of changes are mostly mechanical, so it would be
> great if we can get ack from scheduler people and route these together.
> 
> Thanks.
>