Re: [PATCH] memcg: don't do cleanup manually if mem_cgroup_css_online() fails

[Li Zefan <lizefan-hv44wF8Li93QT0dZR+AlfA@public.gmane.org>]

On 2013/4/2 21:32, Michal Hocko wrote:
> On Tue 02-04-13 16:22:23, Glauber Costa wrote:
>> On 04/02/2013 04:16 PM, Michal Hocko wrote:
>>> On Tue 02-04-13 15:35:28, Li Zefan wrote:
>>> [...]
>>>> @@ -6247,16 +6247,7 @@ mem_cgroup_css_online(struct cgroup *cont)
>>>>  
>>>>  	error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
>>>>  	mutex_unlock(&memcg_create_mutex);
>>>> -	if (error) {
>>>> -		/*
>>>> -		 * We call put now because our (and parent's) refcnts
>>>> -		 * are already in place. mem_cgroup_put() will internally
>>>> -		 * call __mem_cgroup_free, so return directly
>>>> -		 */
>>>> -		mem_cgroup_put(memcg);
>>>> -		if (parent->use_hierarchy)
>>>> -			mem_cgroup_put(parent);
>>>> -	}
>>>> +
>>>>  	return error;
>>>>  }
>>>
>>> The mem_cgroup_put(parent) part is incorrect because mem_cgroup_put goes
>>> up the hierarchy already but I do not think mem_cgroup_put(memcg) should
>>> go away as well. Who is going to free the last reference then?
>>>
>>> Maybe I am missing something but we have:
>>> cgroup_create
>>>   css = ss->css_alloc(cgrp)
>>>     mem_cgroup_css_alloc
>>>       atomic_set(&memcg->refcnt, 1)
>>>   online_css(ss, cgrp)
>>>     mem_cgroup_css_online
>>>       error = memcg_init_kmem		# fails
>>>   goto err_destroy
>>> err_destroy:
>>>   cgroup_destroy_locked(cgrp)
>>>     offline_css
>>>       mem_cgroup_css_offline
>>>
>>> no mem_cgroup_put on the way.
>>>
>>
>> static void mem_cgroup_css_free(struct cgroup *cont)
>> {
>>         struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
>>
>>         kmem_cgroup_destroy(memcg);
>>
>>         mem_cgroup_put(memcg);
>> }
>>
>> kernel/cgroup.c:
>> err_free_all:
>>         for_each_subsys(root, ss) {
>>                 if (cgrp->subsys[ss->subsys_id])
>>                         ss->css_free(cgrp);
>>         }
> 
> But we do not get to that path after online_css fails because that one
> jumps to err_destroy. So this is not it. Maybe css_free gets called from
> cgroup_diput but I got lost in the indirection.
> 

Just like rmdir a cgroup. cgroup_destroy_locked() will be called, and it
makes cgroup dentry go down to 0, and then cgroup_diput() is called, which
calls mem_cgroup_css_free().