Re: [PATCH 4/5] freezer cgroup: Fix an RCU warning in cgroup_freezing_or_frozen()

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

>>>> with CONFIG_PROVE_RCU, a warning can be triggered when we
>>>> resume from suspend:
>>>>
>>>> ...
>>>> include/linux/cgroup.h:533 invoked rcu_dereference_check() without protection!
>>>> ...
>>>>
>>>> task_freezer() calls task_subsys_state(), which needs to be
>>>> protected by rcu_read_lock or cgroup_mutex.
>>>>
>>>> Signed-off-by: Li Zefan <lizf@cn.fujitsu.com>
>>>> ---
>>>>  kernel/cgroup_freezer.c |    2 ++
>>>>  1 files changed, 2 insertions(+), 0 deletions(-)
>>>>
>>>> diff --git a/kernel/cgroup_freezer.c b/kernel/cgroup_freezer.c
>>>> index 5038f4c..ac76983 100644
>>>> --- a/kernel/cgroup_freezer.c
>>>> +++ b/kernel/cgroup_freezer.c
>>>> @@ -53,6 +53,7 @@ int cgroup_freezing_or_frozen(struct task_struct *task)
>>>>  	struct freezer *freezer;
>>>>  	enum freezer_state state;
>>>>  
>>>> +	rcu_read_lock();
>>>>  	task_lock(task);
>>>>  	freezer = task_freezer(task);
>>>>  	if (!freezer->css.cgroup->parent)
>>>> @@ -60,6 +61,7 @@ int cgroup_freezing_or_frozen(struct task_struct *task)
>>>>  	else
>>>>  		state = freezer->state;
>>>>  	task_unlock(task);
>>>> +	rcu_read_unlock();
>>>>  
>>>>  	return (state == CGROUP_FREEZING) || (state == CGROUP_FROZEN);
>>>>  }
>>> Hmm cgroup_attach_task() does hold task_lock() over setting
>>> tsk->cgroups, so doesn't that also pin the task to the cgroup and thus
>>> the cgroup itself?
>> So you are advocating for the rcu_dereference check including the
>> task lock, correct?
> 
> I think that might be correct yes, although I would prefer confirmation
> from someone who actually knows kernel/cgroup.c ;-)
> 

You are right in that taking task_lock() is sufficient (I forgot
this lock rule), but it's not true that whatever locks are held
in the ->attach method can pin a task's cgroup.

So the right fix is including task_lock in rcu_deref check in
task_subsys_state(). I'll send a new fix.