Re: [RFC][PATCH 1/6] memcg: fix pre_destory handler

[Li Zefan <lizf@cn.fujitsu.com>]

> +static bool memcg_is_obsolete(struct mem_cgroup *mem)
> +{

Will this function be called with mem->css.refcnt == 0? If yes, then
this function is racy.

cg = mem->css.cgroup
				cgroup_diput()
				  mem_cgroup_destroy()
				    mem->css.cgroup = NULL;
				  kfree(cg);
if (!cg || cgroup_is_removed(cg)...)

(accessing invalid cg)

> +	struct cgroup *cg = mem->css.cgroup;
> +	/*
> +	 * "Being Removed" means pre_destroy() handler is called.
> +	 * After  "pre_destroy" handler is called, memcg should not
> +	 * have any additional charges.
> +	 * This means there are small races for mis-accounting. But this
> +	 * mis-accounting should happen only under swap-in opration.
> +	 * (Attachin new task will fail if cgroup is under rmdir()).
> +	 */
> +
> +	if (!cg || cgroup_is_removed(cg) || cgroup_is_being_removed(cg))
> +		return true;
> +	return false;
> +}
> +

...

>  static void mem_cgroup_destroy(struct cgroup_subsys *ss,
>  				struct cgroup *cont)
>  {
> -	mem_cgroup_free(mem_cgroup_from_cont(cont));
> +	struct mem_cgroup *mem = mem_cgroup_from_cont(cont):
> +	mem_cgroup_free(mem);
> +	/* forget */
> +	mem->css.cgroup = NULL;

mem might already be destroyed by mem_cgroup_free(mem).

>  }