Re: [PATCH] cpuset: current_cpuset_is_being_rebound() need rcu lock

[KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>]

> On Fri, Mar 12, 2010 at 09:03:03AM +0900, KOSAKI Motohiro wrote:
> > > On Thu, Mar 11, 2010 at 02:46:15PM +0900, KOSAKI Motohiro wrote:
> > > > rcu lockdep detected cpuset have wrong rcu usage.
> > > > the fixing is trivial. but I wonder why don't cpuset_being_rebound assignment
> > > > and read need a memory barrier pairing?
> > > 
> > > The fix is in -tip, commit 99ee4ca746dda71326db7645463b4075ac1d665c.
> > > 
> > > This is an initialization-time use of rcu_dereference(), so no other
> > > task has a reference to this data.  Hence it is constant.  Other uses
> > > of this code operate on shared data structures, which might change at
> > > any time.
> > 
> > thanks. I haven't notice such commit.
> > 
> > I think you are talking about task_cs(current) accessing and you are right
> > in such point.
> > but I'm talking cpuset_being_rebound global variable.
> > 
> > update_tasks_nodemask() has following code
> > 
> > static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
> >                                  struct ptr_heap *heap)
> > {
> >         cpuset_being_rebound = cs;		/* start transaction */
> >         cgroup_scan_tasks(&scan);
> >         cpuset_being_rebound = NULL;		/* end transaction */
> > }
> 
> Hmmm...  What commit are you looking at in what tree?  I instead see
> a much larger function body for update_tasks_nodemask().  I am looking
> in a number of places, including 522dba7134d6b2e5821d3457f7941ec34f668e6d
> in Linus's git tree.

Ahh, I'm sorry. I wrote essential piece of the function instead
full cut-n-paste.

Plus, my previous explanation was too little and unclear. I don't think
your commit is wrong. it is definitely right. I only talked about I think
current_cpuset_is_being_rebound() has another sick.