lkml.org 
[lkml]   [2015]   [May]   [29]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC 3/3] memcg: get rid of mm_struct::owner
Hello,

On Fri, May 29, 2015 at 04:57:39PM +0200, Michal Hocko wrote:
> > > "
> > > It also allows several control groups that are virtually grouped by
> > > mm_struct, to exist independent of the memory controller i.e., without
> > > adding mem_cgroup's for each controller, to mm_struct.
> > > "
> > > suggests it might have been intentional. That being said, I think it was
> >
> > I think he's talking about implmenting different controllers which may
> > want to add their own css pointer in mm_struct now wouldn't need to as
> > the mm is tagged with the owning task from which membership of all
> > controllers can be derived. I don't think that's something we need to
> > worry about. We haven't seen even a suggestion for such a controller
> > and even if that happens we'd be better off adding a separate field
> > for the new controller.
>
> Maybe I've just misunderstood. My understandig was that tasks sharing
> the mm could live in different cgroups while the memory would be bound
> by a shared memcg.

Hmm.... it specifically goes into explaining that it's about having
different controllers sharing the owner field.

"i.e., without adding mem_cgroup's for each controller, to mm_struct."

It seems fairly clear to me.

> > I'm a bit lost on what's cleared defined is actually changing. It's
> > not like userland had firm control over mm->owner. It was already a
> > crapshoot, no?
>
> OK so you creat a task A (leader) which clones several tasks Pn with
> CLONE_VM without CLONE_THREAD. Moving A around would control memcg
> membership while Pn could be moved around freely to control membership
> in other controllers (e.g. cpu to control shares). So it is something
> like moving threads separately.

Sure, it'd behave clearly in certain cases but then again you'd have
cases where how mm->owner changes isn't clear at all when seen from
the userland. e.g. When the original owner goes away, the assignment
of the next owner is essentially arbitrary. That's what I meant by
saying it was already a crapshoot. We should definitely document the
change but this isn't likely to be an issue. CLONE_VM &&
!CLONE_THREAD is an extreme corner case to begin with and even the
behavior there wasn't all that clearly defined.

Thanks.

--
tejun


\
 
 \ /
  Last update: 2015-05-29 18:01    [W:0.050 / U:10.220 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site