Re: [PATCH][rc1] cgroup: catch bad css refcnt at css_put

[KAMEZAWA Hiroyuki]

On Tue, 29 Sep 2009 08:50:33 +0800
Li Zefan <lizf@xxxxxxxxxxxxxx> wrote:

> KAMEZAWA Hiroyuki wrote:
> > This is a patch for checking css->refcnt's sanity at css_put().
> > 
> > BTW, I noticed that...css->refcnt may overflow if used with memcg...
> > Now, refcnt is incremented per a page. Paul, do you have any idea ?
> 
> atomic64_t ?
> 
maybe. atomic_long_t ?

> But for 4K pagesize, it won't overflow until when the referenced
> memory is > 8T?
> 
you're right. But there tends to be a few users who use unbelievable amounts
of memory in the world.
(Such user uses memcg or not is another problem ;)

> > (Ah, yes. "don't use css->refcnt per page" is maybe reasonable but
> >  it will be big change..)
> > 
> > ==
> > __css_put() doesn't check a bug as refcnt goes to minus.
> > I think it should be caught. This patch adds a check for it.
> > 
> > Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx>
> 
> Acked-by: Li Zefan <lizf@xxxxxxxxxxxxxx>
> 
> > ---
> >  kernel/cgroup.c |    5 ++++-
> >  1 file changed, 4 insertions(+), 1 deletion(-)
> > 
> > Index: linux-2.6.32-rc1/kernel/cgroup.c
> > ===================================================================
> > --- linux-2.6.32-rc1.orig/kernel/cgroup.c
> > +++ linux-2.6.32-rc1/kernel/cgroup.c
> > @@ -3708,8 +3708,10 @@ static void check_for_release(struct cgr
> >  void __css_put(struct cgroup_subsys_state *css)
> >  {
> >  	struct cgroup *cgrp = css->cgroup;
> > +	int val;
> >  	rcu_read_lock();
> > -	if (atomic_dec_return(&css->refcnt) == 1) {
> > +	val = atomic_dec_return(&css->refcnt);
> > +	if (val == 1) {
> >  		if (notify_on_release(cgrp)) {
> >  			set_bit(CGRP_RELEASABLE, &cgrp->flags);
> >  			check_for_release(cgrp);
> > @@ -3717,6 +3719,7 @@ void __css_put(struct cgroup_subsys_stat
> >  		cgroup_wakeup_rmdir_waiter(cgrp);
> >  	}
> >  	rcu_read_unlock();
> > +	WARN_ON(val < 1);
> 
> When we run into this, it'll probably fill up the syslog quickly,
> so I think WARN_ON_ONCE() is a bit better.
> 
Hmm, ok. I'll rewrite.

Thanks,
-Kame


> >  }
> >  
> 

--
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/