Re: [mm] [PATCH 3/4] Memory cgroup hierarchical reclaim (v4)

[Balbir Singh <balbir@linux.vnet.ibm.com>]

Daisuke Nishimura wrote:
> Hi.
> 
> Unfortunately, trying to hold cgroup_mutex at reclaim causes dead lock.
> 
> For example, when attaching a task to some cpuset directory(memory_migrate=on),
> 
>     cgroup_tasks_write (hold cgroup_mutex)
>         attach_task_by_pid
>             cgroup_attach_task
>                 cpuset_attach
>                     cpuset_migrate_mm
>                         :
>                         unmap_and_move
>                             mem_cgroup_prepare_migration
>                                 mem_cgroup_try_charge
>                                     mem_cgroup_hierarchical_reclaim
> 

Did lockdep complain about it?

1. We could probably move away from cgroup_mutex to a memory controller specific
mutex.
2. We could give up cgroup_mutex before migrate_mm, since it seems like we'll
hold the cgroup lock for long and holding it during reclaim will definitely be
visible to users trying to create/delete nodes.

I prefer to do (2), I'll look at the code more closely

> I think similar problem can also happen when removing memcg's directory.
> 

Why removing a directory? memcg (now) marks the directory as obsolete and we
check for obsolete directories and get/put references.

Thanks for the bug report!

-- 
	Balbir