Re: [Bug] race condition at rebind_subsystems()

[Jing-Ting Wu]

On Fri, 2022-07-15 at 06:47 -1000, Tejun Heo wrote:
> (resending, I messed up the message header, sorry)
> 
> Hello,
> 
> On Fri, Jul 15, 2022 at 01:59:38PM +0200, Michal Koutný wrote:
> > The css->rstat_css_node should not be modified if there are
> > possible RCU
> > readers elsewhere.
> > One way to fix this would be to insert synchronize_rcu() after
> > list_del_rcu() and before list_add_rcu().
> > (A further alternative (I've heard about) would be to utilize
> > 'nulls'
> > RCU lists [1] to make the move between lists detectable.)
> > 
> > But as I'm looking at it from distance, it may be simpler and
> > sufficient
> > to just take cgroup_rstat_lock around the list migration (the
> > nesting
> > under cgroup_mutex that's held with rebind_subsystems() is fine).
> 
> synchronize_rcu() prolly is the better fit here given how that
> list_node's
> usage, but yeah, great find.
> 
> Thanks.
> 

Hi Michal and Tejun,


Thanks for your suggestion.
Accroding your description, is the following patch corrent?

--- a/kernel/cgroup/cgroup.c
+++ b/kernel/cgroup/cgroup.c
@@ -1813,6 +1813,7 @@
 
 		if (ss->css_rstat_flush) {
 			list_del_rcu(&css->rstat_css_node);
+			synchronize_rcu();
 			list_add_rcu(&css->rstat_css_node,
 				     &dcgrp->rstat_css_list);
 		}


If the patch is correct, we will add this patch to our stability test.
And we will continue to observe whether the problem is solved.
Thank you.




Best regards,
Jing-Ting Wu