lkml.org 
[lkml]   [2008]   [May]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [-mm][PATCH 4/4] Add memrlimit controller accounting and control (v4)
* Paul Menage <menage@google.com> [2008-05-15 00:39:45]:

> On Thu, May 15, 2008 at 12:03 AM, Balbir Singh
> <balbir@linux.vnet.ibm.com> wrote:
> >
> > I want to focus on this conclusion/assertion, since it takes care of
> > most of the locking related discussion above, unless I missed
> > something.
> >
> > My concern with using mmap_sem, is that
> >
> > 1. It's highly contended (every page fault, vma change, etc)
>
> But the only *new* cases of taking the mmap_sem that this would
> introduce would be:
>
> - on a failed vm limit charge

Why a failed charge? Aren't we talking of moving all charge/uncharge
under mmap_sem?

> - when a task exit/exec causes an mm ownership change

Yes, in the mm_owner_changed callbacks

> - when a task moves between two cgroups in the memrlimit hierarchy.
>

Yes, this would nest cgroup_mutex and mmap_sem. Not sure if that would
be a bad side-effect.

> All of these should be rare events, so I don't think the additional
> contention is a worry.

We do make several of all charge calls under the mmap_sem, but not
all of them. So the additional contention might not be all that bad.

>
> > 2. It's going to make the locking hierarchy deeper and complex
>
> Yes, potentially. But if the upside of that is that we eliminate a
> lock/unlock on a shared lock on every mmap/munmap call, it might well
> be worth it.
>
> > 3. It's not appropriate to call all the accounting callbacks with
> > the mmap_sem() held, since the undo operations _can get_ complicated
> > at the caller.
> >
>
> Can you give an example?

Some paths of the uncharge are not under mmap_sem. Undoing the
operation there seemed complex.

>
> > I would prefer introducing a new lock, so that other subsystems are
> > not affected.
> >
>
> For getting the first cut of the memrlimit controller working this may
> well make sense. But it would be nice to avoid it longer-term.

OK, so here's what I am going to try and do

Refactor the code to try and use mmap_sem and see what I come up
with. Basically use mmap_sem for all charge/uncharge operations as
well use mmap_sem in read_mode in the move_task() and
mm_owner_changed() callbacks. That should take care of the race
conditions discussed, unless I missed something.
Try and instrument insert_vm_struct() for charge/uncharge

--
Warm Regards,
Balbir Singh
Linux Technology Center
IBM, ISTL


\
 
 \ /
  Last update: 2008-05-15 10:29    [from the cache]
©2003-2011 Jasper Spaans