Re: [PATCH 3/7] memcg: immigrate charges only when a threadgroup leader is moved

[Michal Hocko <mhocko@suse.cz>]

On Thu 21-05-15 13:22:17, Johannes Weiner wrote:
> On Wed, May 20, 2015 at 10:22:21PM +0200, Michal Hocko wrote:
> > On Wed 20-05-15 19:53:02, Oleg Nesterov wrote:
> > > On 05/20, Michal Hocko wrote:
> > > >
> > > > So I assume the leader simply waits for its threads to finish and it
> > > > stays in the sibling list. __unhash_process seems like it does the final
> > > > cleanup and unlinks the leader from the lists. Which means that
> > > > mm_update_next_owner never sees !group_leader. Is that correct Oleg?
> > > 
> > > Yes, yes, the group leader can't go away until the whole thread-group dies.
> > 
> > OK, then we should have a guarantee that mm->owner is always thread
> > group leader, right?
> > 
> > > But can't we kill mm->owner somehow?
> > 
> > I would be happy about that. But it is not that simple.
> > 
> > > I mean, turn it into something else,
> > > ideally into "struct mem_cgroup *" although I doubt this is possible.
> > 
> > Sounds like a good idea but... it duplicates the cgroup tracking into
> > two places and that asks for troubles. On the other hand we are doing
> > that already because mm->owner might be in a different cgroup than the
> > current. However, this is an inherent problem because CLONE_VM doesn't
> > imply CLONE_THREAD. So in the end it doesn't look much worse IMO.
> > We will loose the "this task is in charge" aspect and that would
> > be a user space visible change but I am not sure how much it is a
> > problem. Maybe somebody is (ab)using this to workaround the restriction
> > that all threads are in the same cgroup.
> 
> If mm->owner is currently always the threadgroup leader, it should be
> fairly straight forward to maintain mm->memcg on all events that move
> any threadgroup leader between cgroups, without having mm->owner, no?

I have a tentative patch for that. It is fairly straightforward and it
even reduces the code size. I plan to post it early next week after it
gets some testing. The primary thing I am worried about is the user
visible behavior change, though.

> It would have a lot of benefits for sure.  The code would be simpler,
> but it would also reduce some of the cost that Mel is observing inside
> __mem_cgroup_count_vm_event(), by reducing one level of indirection.

Agreed!

-- 
Michal Hocko
SUSE Labs

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>