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

[Glauber Costa <glommer@parallels.com>]

On 04/02/2013 12:34 PM, Li Zefan wrote:
> On 2013/4/2 16:07, Glauber Costa wrote:
>> On 04/02/2013 12:03 PM, Li Zefan wrote:
>>> On 2013/4/2 15:35, Li Zefan wrote:
>>>> If memcg_init_kmem() returns -errno when a memcg is being created,
>>>> mem_cgroup_css_online() will decrement memcg and its parent's refcnt,
>>>
>>>> (but strangely there's no mem_cgroup_put() for mem_cgroup_get() called
>>>> in memcg_propagate_kmem()).
>>>
>>> The comment in memcg_propagate_kmem() suggests it knows mem_cgroup_css_free()
>>> will be called in failure, while mem_cgroup_css_online() doesn't know.
>>>
>> This is a bit suspicious. At first your analysis seems fair, but I've
>> extensively tested memcg teardown process with kmemcg (and even
>> uncovered some bugs at that), and it works when and how expected.
>>
> 
> Because this bug is in a failure path, and seems the only way to get into
> this path is -ENOMEM.
> 
Yes, but I tend to test that with manually introduced error codes.

For what is worth, I just did it. And indeed, by ignoring kmemcg
initialization and failing with ENOMEM here triggers a bug. Your patch
fixes it. I tested both failing all non-root, and letting the first
succeed and failing the second if the parent is use_hierarchy. Both
cases have a bug initially that you fix.

If that is allowed to proceed, kmemcg initialization and teardown works
as expected.

>> Also, note that this teardown code long predates kmemcg.
>>
> 
> Maybe this bug was introduced when ss->create() was changed to ss->css_alloc()
> and ss->css_online(), and before that change ss->destroy() won't be called
> if ss->create() failed.
> 
Yes, this is possible,