[PATCH 1/2] cgroups: enhance task_cgroup()

[PATCH 1/2] cgroups: enhance task_cgroup()

[Lai Jiangshan]

task_cgroup() calls cgroup_subsys_state().
and we must use rcu_read_lock() to protect cgroup_subsys_state().
so we must use rcu_read_lock() to protect task_cgroup().

but it'll not so friendly to caller: the callers of task_cgroup() have
held cgroup_lock(). it means that struct cgroup will not be freed.

So this patch add rcu_read_lock() in task_cgroup() to enhance task_cgroup().
And we do NOT NEED FIX task_cgroup()'s callers, and cgroup_lock()
can protect task_cgroup().

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
---
diff --git a/include/linux/cgroup.h b/include/linux/cgroup.h
index 1164963..22901ff 100644
--- a/include/linux/cgroup.h
+++ b/include/linux/cgroup.h
@@ -359,6 +360,10 @@
 static inline struct cgroup* task_cgroup(struct task_struct *task,
 					       int subsys_id)
 {
-	return task_subsys_state(task, subsys_id)->cgroup;
+	struct cgroup *ret;
+	rcu_read_lock();
+	ret = task_subsys_state(task, subsys_id)->cgroup;
+	rcu_read_unlock();
+	return ret;
 }
 
 int cgroup_clone(struct task_struct *tsk, struct cgroup_subsys *ss,

Re: [PATCH 1/2] cgroups: enhance task_cgroup()

[Paul Menage]

On Fri, Nov 21, 2008 at 12:49 AM, Lai Jiangshan <laijs@cn.fujitsu.com> wrote:
>
> task_cgroup() calls cgroup_subsys_state().

No, it calls task_subsys_state()

> and we must use rcu_read_lock() to protect cgroup_subsys_state().
> so we must use rcu_read_lock() to protect task_cgroup().
>
> but it'll not so friendly to caller: the callers of task_cgroup() have
> held cgroup_lock(). it means that struct cgroup will not be freed.
>
> So this patch add rcu_read_lock() in task_cgroup() to enhance task_cgroup().
> And we do NOT NEED FIX task_cgroup()'s callers, and cgroup_lock()
> can protect task_cgroup().

Is there a reason to add an implicit rcu_read_lock() in task_cgroup()
and not directly in task_subsys_state() ?

Paul

>
> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
> ---
> diff --git a/include/linux/cgroup.h b/include/linux/cgroup.h
> index 1164963..22901ff 100644
> --- a/include/linux/cgroup.h
> +++ b/include/linux/cgroup.h
> @@ -359,6 +360,10 @@
>  static inline struct cgroup* task_cgroup(struct task_struct *task,
>                                               int subsys_id)
>  {
> -       return task_subsys_state(task, subsys_id)->cgroup;
> +       struct cgroup *ret;
> +       rcu_read_lock();
> +       ret = task_subsys_state(task, subsys_id)->cgroup;
> +       rcu_read_unlock();
> +       return ret;
>  }
>
>  int cgroup_clone(struct task_struct *tsk, struct cgroup_subsys *ss,
>
>
>

Re: [PATCH 1/2] cgroups: enhance task_cgroup()

[Li Zefan]

Paul Menage wrote:
> On Fri, Nov 21, 2008 at 12:49 AM, Lai Jiangshan <laijs@cn.fujitsu.com> wrote:
>> task_cgroup() calls cgroup_subsys_state().
> 
> No, it calls task_subsys_state()
> 
>> and we must use rcu_read_lock() to protect cgroup_subsys_state().
>> so we must use rcu_read_lock() to protect task_cgroup().
>>
>> but it'll not so friendly to caller: the callers of task_cgroup() have
>> held cgroup_lock(). it means that struct cgroup will not be freed.
>>
>> So this patch add rcu_read_lock() in task_cgroup() to enhance task_cgroup().
>> And we do NOT NEED FIX task_cgroup()'s callers, and cgroup_lock()
>> can protect task_cgroup().
> 
> Is there a reason to add an implicit rcu_read_lock() in task_cgroup()
> and not directly in task_subsys_state() ?
> 

I don't think it's a good idea to add rcu_read_lock() in task_cgroup() or
task_subsys_state(). The returned pointer of the 2 functions should be
protected by rcu_read_lock or cgroup_lock, so the caller should still
hold the lock by itself.

An example of this is:

static int mem_cgroup_charge_common(...)
{
	...
		rcu_read_lock();
		mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
		if (unlikely(!mem)) {
			rcu_read_unlock();
			return 0;
		}
		/*
		 * For every charge from the cgroup, increment reference count
		 */
		css_get(&mem->css);
		rcu_read_unlock();
	...
}

> Paul
> 
>> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
>> ---
>> diff --git a/include/linux/cgroup.h b/include/linux/cgroup.h
>> index 1164963..22901ff 100644
>> --- a/include/linux/cgroup.h
>> +++ b/include/linux/cgroup.h
>> @@ -359,6 +360,10 @@
>>  static inline struct cgroup* task_cgroup(struct task_struct *task,
>>                                               int subsys_id)
>>  {
>> -       return task_subsys_state(task, subsys_id)->cgroup;
>> +       struct cgroup *ret;
>> +       rcu_read_lock();
>> +       ret = task_subsys_state(task, subsys_id)->cgroup;
>> +       rcu_read_unlock();
>> +       return ret;
>>  }
>>
>>  int cgroup_clone(struct task_struct *tsk, struct cgroup_subsys *ss,
>>
>>
>>

Re: [PATCH 1/2] cgroups: enhance task_cgroup()

[Lai Jiangshan]

Paul Menage wrote:
> On Fri, Nov 21, 2008 at 12:49 AM, Lai Jiangshan <laijs@cn.fujitsu.com> wrote:
>> task_cgroup() calls cgroup_subsys_state().
> 
> No, it calls task_subsys_state()
> 
>> and we must use rcu_read_lock() to protect cgroup_subsys_state().
>> so we must use rcu_read_lock() to protect task_cgroup().
>>
>> but it'll not so friendly to caller: the callers of task_cgroup() have
>> held cgroup_lock(). it means that struct cgroup will not be freed.
>>
>> So this patch add rcu_read_lock() in task_cgroup() to enhance task_cgroup().
>> And we do NOT NEED FIX task_cgroup()'s callers, and cgroup_lock()
>> can protect task_cgroup().
> 
> Is there a reason to add an implicit rcu_read_lock() in task_cgroup()
> and not directly in task_subsys_state() ?

Yes.

The caller have held the cgroup_lock() when it calls task_cgroup().
After we add an implicit rcu_read_lock() in task_cgroup(),
we don't need rcu_read_lock()/task_lock() for using task_cgroup().

For cgroup_exit() will change tsk->cgroups, if we don't
add an implicit rcu_read_lock() in task_cgroup(), we have to fix 7
places which using task_cgroup().


task_subsys_state() is different, it is used in fast path,
If we add an implicit rcu_read_lock() in task_subsys_state(),
we still need rcu_read_lock()/task_lock() for using it,
so it's redundant rcu_read_lock(), and slower the fast path a little.


Lai.

> 
> Paul
>