[RFD][PATCH] memcg: Move Usage at Task Move

[Serge E. Hallyn]

Quoting KAMEZAWA Hiroyuki (kamezawa.hiroyu at jp.fujitsu.com):
> On Wed, 11 Jun 2008 01:48:20 -0700
> "Paul Menage" <menage at google.com> wrote:
> 
> > On Wed, Jun 11, 2008 at 1:27 AM, KAMEZAWA Hiroyuki
> > <kamezawa.hiroyu at jp.fujitsu.com> wrote:
> > > Sorry. try another sentense..
> > >
> > > I think cgroup itself is designed to be able to be used without middleware.
> > 
> > True, but it shouldn't be hostile to middleware, since I think that
> > automated use will be much more common. (And certainly if you count
> > the number of servers :-) )
> > 
> > > IOW, whether using middleware or not is the matter of users not of developpers.
> > > There will be a system that system admin controlles all and move tasks by hand.
> > > ex)...personal notebooks etc..
> > >
> > 
> > You think so? I think that at the very least users will be using tools
> > based around config scripts, rule engines and libcgroup, if not a
> > persistent daemon.
> > 
> I believe some users will never use middlewares because of their special
> usage of linux.
> 
> 
> 
> > >> If the common mode for middleware starting a new cgroup is fork() /
> > >> move / exec() then after the fork(), the child will be sharing pages
> > >> with the main daemon process. So the move will pull all the daemon's
> > >> memory into the new cgroup
> > >>
> > > My patch (this patch) just moves Private Anon page to new cgroup. (of mapcount=1)
> > 
> > OK, well that makes it more reasonable regarding the above problem.
> > But I can still see problems if, say, a single thread moves into a new
> > cgroup, you move the entire memory. Perhaps you should only do so if
> > the mm->owner changes task?
> > 
> 
> Thank you for pointing out. I'll add mm->owner check.
> 
> BTW, should we have a cgroup for SYSVIPC resource controller and devide it
> from memory resource controller ?  I think that per-task on-demand usage
> accounting is not suitable for shmem (and hugepage).
> per-creater (caller of shmget()) accounting seems to be better for me.
> 
> Just a question:
> What happens when a thread (not thread-group-leader) changes its ns by
> ns-cgroup ? not-allowed ?

I don't quite understand the question.  I assume you're asking whether
your cgroup, when composed with ns, will refuse a task in cgroup /cg/1/2
from being able to

	mkdir /cg/1/2/3
	echo $$ > /cg/1/2/3/tasks

or

	unshare(CLONE_NEWNS)

which the ns cgroup would allow, and what your cgroup would do in that
case.  If your question ("not-allowed ?") is about ns cgroup behavior
then please rephrase.

thanks,
-serge