Re: [PATCH stable-4.4 1/3] mm: memcontrol: fix cgroup creation failure after many small jobs

[Johannes Weiner]

On Mon, Aug 15, 2016 at 02:46:19PM +0200, Michal Hocko wrote:
> On Mon 15-08-16 08:34:07, Johannes Weiner wrote:
> > Hi Michal, thanks for doing this. There is only one issue I can see:
> > 
> > On Fri, Aug 12, 2016 at 11:56:17AM +0200, Michal Hocko wrote:
> > > @@ -4171,17 +4211,27 @@ static struct mem_cgroup *mem_cgroup_alloc(void)
> > >  	if (!memcg)
> > >  		return NULL;
> > >  
> > > +	memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
> > > +				 1, MEM_CGROUP_ID_MAX,
> > > +				 GFP_KERNEL);
> > > +	if (memcg->id.id < 0)
> > > +		goto out_free;
> > > +
> > >  	memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
> > >  	if (!memcg->stat)
> > > -		goto out_free;
> > > +		goto out_idr;
> > >  
> > >  	if (memcg_wb_domain_init(memcg, GFP_KERNEL))
> > >  		goto out_free_stat;
> > >  
> > > +	idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
> > 
> > This publishes the memcg object too early. Before 4.5, the memcg is
> > not fully initialized in mem_cgroup_alloc(). You have to move the
> > idr_replace() down to that function (and idr_remove() on free_out).
> 
> You are right. I am just wondering whether it matters. Nobody should see
> the id so nobody will be looking it up, no?

Page cache shadow entries refer to these IDs weakly. It's possible to
refault with a recently recycled memcg ID and crash. That's why we do
the whole alloc(NULL) -> replace(memcg) dance in the first place.