Hello, On Thu, Feb 13, 2014 at 04:27:45PM +0100, Michal Hocko wrote: > > Further testing showed that an ordered workqueue for cgroup_destroy_wq > > is not always good enough: percpu_ref_kill_and_confirm's call_rcu_sched > > stage on the way can mess up the order before reaching the workqueue. > > This whole code path is so complicated by different types of delayed > work that I am not wondering that we have missed that :/ Yeah, I know. Good part of the complexity comes from RCU -> wq bouncing. I wonder whether we just should bite the bullet and add something along the line of call_rcu_work(). The other part is percpu ref shutdown. For me that part is easier to swallow, as the benefits are quite clear. > > Instead, when offlining a memcg, call mem_cgroup_reparent_charges() on > > all its children (and grandchildren, in the correct order) to have their > > charges reparented first. > > That is basically what I was suggesting > http://marc.info/?l=linux-mm&m=139178386407184&w=2 as #1 option. I > cannot say I would like it and I think that reparenting LRUs in > css_offline and then reparent the remaining charges from css_free is a > better solution but let's keep this for later. I'm kinda wishing the reparenting things works out. Even if that involves a bit of overhead at offline, I think it'd be worthwhile to be able to follow the same object lifetime rules as other controllers, as long as the overhead is reasonable. Thanks. -- tejun -- To unsubscribe from this list: send the line "unsubscribe cgroups" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
