Re: [PATCH 7/8] memcg: get rid of mm_struct::owner

[Michal Hocko <mhocko@kernel.org>]

On Tue 14-07-15 17:18:23, Michal Hocko wrote:
> On Fri 10-07-15 16:05:33, Michal Hocko wrote:
> > JFYI: I've found some more issues while hamerring this more.
> 
> OK so the main issue is quite simple but I have completely missed it when
> thinking about the patch before. clone(CLONE_VM) without CLONE_THREAD is
> really nasty and it will easily lockup the machine with preempt. disabled
> for ever. It goes like this:
> taskA (in memcg A)
>   taskB = clone(CLONE_VM)
> 				taskB
> 				  A -> B	# Both tasks charge to B now
> 				  exit()	# No tasks in B -> can be
> 				  		# offlined now
> 				css_offline()
>   mem_cgroup_try_charge
>     get_mem_cgroup_from_mm
>       rcu_read_lock()
>       do {
>       } while css_tryget_online(mm->memcg)	# will never succeed
>       rcu_read_unlock()
> 
> taskA and taskB are basically independent entities wrt. the life
> cycle (unlike threads which are bound to the group leader). The
> previous code handles this by re-ownering during exit by the monster
> mm_update_next_owner.
> 
> I can see the following options without reintroducing reintroducing
> some form of mm_update_next_owner:
> 
> 1) Do not allow offlining a cgroup if we have active users in it.  This
> would require a callback from the cgroup core to the subsystem called if
> there are no active tasks tracked by the cgroup core. Tracking on the memcg
> side doesn't sound terribly hard - just mark a mm_struct as an alien and
> count the number of aliens during the move in mem_cgroup. mm_drop_memcg
> then drops the counter. We could end up with EBUSY cgroup without any
> visible tasks which is a bit awkward.
> 
> 2) update get_mem_cgroup_from_mm and others to fallback to the parent
> memcg if the current one is offline. This would be in line with charge
> reparenting we used to do. I cannot say I would like this because it
> allows for easy runaway to the root memcg if the hierarchy is not
> configured cautiously. The code would be also quite tricky because each
> direct consumer of mm->memcg would have to be aware of this. This is
> awkward.
> 
> 3) fail mem_cgroup_can_attach if we are trying to migrate a task sharing
> mm_struct with a process outside of the tset. If I understand the
> tset properly this would require all the sharing tasks to be migrated
> together and we would never end up with task_css != &task->mm->css.
> __cgroup_procs_write doesn't seem to support multi pid move currently
> AFAICS, though. cgroup_migrate_add_src, however, seems to be intended
> for this purpose so this should be doable. Without that support we would
> basically disallow migrating these tasks - I wouldn't object if you ask
> me.
> 
> Do you see other options? From the above three options the 3rd one
> sounds the most sane to me and the 1st quite easy to implement. Both will
> require some cgroup core work though. But maybe we would be good enough
> with 3rd option without supporting moving schizophrenic tasks and that
> would be reduced to memcg code.
> 
> Or we can, of course, stay with the current state but I think it would
> be much saner to get rid of the schizophrenia.
> 
> What do you think?

Ideas, thoughs? Anybody?
-- 
Michal Hocko
SUSE Labs