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

[Vladimir Davydov]

On Fri, Jul 10, 2015 at 02:45:20PM +0200, Michal Hocko wrote:
> On Fri 10-07-15 10:54:00, Vladimir Davydov wrote:
> > On Thu, Jul 09, 2015 at 04:09:41PM +0200, Michal Hocko wrote:
> > > On Wed 08-07-15 20:32:51, Vladimir Davydov wrote:
> > > > On Wed, Jul 08, 2015 at 02:27:51PM +0200, Michal Hocko wrote:
> > [...]
> > > > > @@ -474,7 +519,7 @@ static inline void mem_cgroup_count_vm_event(struct mm_struct *mm,
> > > > >  		return;
> > > > >  
> > > > >  	rcu_read_lock();
> > > > > -	memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
> > > > > +	memcg = rcu_dereference(mm->memcg);
> > > > >  	if (unlikely(!memcg))
> > > > >  		goto out;
> > > > >  
> > > > 
> > > > If I'm not mistaken, mm->memcg equals NULL for any task in the root
> > > > memory cgroup
> > > 
> > > right
> > > 
> > > > (BTW, it it's true, it's worth mentioning in the comment
> > > > to mm->memcg definition IMO). As a result, we won't account the stats
> > > > for such tasks, will we?
> > > 
> > > well spotted! This is certainly a bug. There are more places which are
> > > checking for mm->memcg being NULL and falling back to root_mem_cgroup. I
> > > think it would be better to simply use root_mem_cgroup right away. We
> > > can setup init_mm.memcg = root_mem_cgroup during initialization and be
> > > done with it. What do you think? The diff is in the very end of the
> > > email (completely untested yet).
> > 
> > I'd prefer initializing init_mm.memcg to root_mem_cgroup. This way we
> > wouldn't have to check whether mm->memcg is NULL or not here and there,
> > which would make the code cleaner IMO.
> 
> So the patch I've posted will not work as a simple boot test told me. We
> are initializing root_mem_cgroup too late. This will be more complicated.
> I will leave this idea outside of this patch series and will come up
> with a separate patch which will clean this up later. I will update the
> doc discouraging any use of mm->memcg outside of memcg and use accessor
> functions instead. There is only one currently (mm/debug.c) and this is
> used only to print the pointer which is safe.

Why can't we make root_mem_cgroup statically allocated? AFAICS it's a
common practice - e.g. see blkcg_root, root_task_group.

Thanks,
Vladimir
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/