Re: [PATCH] cgroup: bug fix of cgroup_root object was not released after cgroup umounted

[Tejun Heo <tj@kernel.org>]

Hello, Zhang.

On Mon, Dec 14, 2015 at 07:45:54PM +0800, kyler.zhangkui@huawei.com wrote:
> @@ -2147,6 +2147,13 @@ static void cgroup_kill_sb(struct super_block *sb)
>  {
>  	struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
>  	struct cgroup_root *root = cgroup_root_from_kf(kf_root);
> +	struct cgroup *child;
> +	int live_child = 0;
> +
> +	mutex_lock(&cgroup_mutex);
> +	cgroup_for_each_live_child(child, &root->cgrp) {
> +		live_child++;
> +	}
>  
>  	/*
>  	 * If @root doesn't have any mounts or children, start killing it.
> @@ -2155,12 +2162,12 @@ static void cgroup_kill_sb(struct super_block *sb)
>  	 *
>  	 * And don't kill the default root.
>  	 */
> -	if (!list_empty(&root->cgrp.self.children) ||
> -	    root == &cgrp_dfl_root)
> -		cgroup_put(&root->cgrp);
> +	if (live_child || root == &cgrp_dfl_root)
> +	cgroup_put(&root->cgrp);
>  	else
>  		percpu_ref_kill(&root->cgrp.self.refcnt);
>  
> +	mutex_unlock(&cgroup_mutex);
>  	kernfs_kill_sb(sb);

I don't know.  With the above change, we might fail remounting with
the same option for some controllers which keep residual objects
around.  Changing controller associations has never been supported
correctly and I don't think it'll ever be.  I have no objection to
improving it but this has the potential to regress use cases which
would be more common.

Thanks.

-- 
tejun