Re: [RFC] [PATCH] cgroup: add "procs" control file

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

Paul Menage wrote:
> On Wed, Jun 18, 2008 at 1:02 AM, Li Zefan <lizf@cn.fujitsu.com> wrote:
>> - What to do if the attaching of a thread failed? continue to attach
>>  other threads, or stop and return error?
> 
> I think this is something that will have to be handled in the design
> of transactional cgroup attach.
> 

Is the following proposal feasable?
- call can_attach() for each thread before attaching them into the new group.
  This works for cpuset, doesn't it?
- the above may not always reasonable, for example for Kosaki-san's task cgroup.
  in this case, we require the subsystem to provide a can_attach_thread_group(),
  like:

static int task_cgroup_can_attach_group(struct cgroup_subsys *ss,
				  struct cgroup *cgrp, struct task_struct *tsk)
{
	struct task_cgroup *taskcg = task_cgroup_from_cgrp(cgrp);
	struct task_struct *t;
	int ret = 0;
	int nr_threads = 1;

	for (t = next_thread(tsk); t != tsk; t = next_thread(t)
		nr_threads++;

	spin_lock(&taskcg->lock);
	if (taskcg->nr_tasks + nr_threads > taskcg->max_tasks)
		ret = -EBUSY;
	spin_unlock(&taskcg->lock);

	return ret;
}

>> - When a sub-thread of a process is in the cgroup, but not its thread
>>  cgroup leader, what to do when 'cat procs'? just skip those threads?
> 
> Sounds reasonable. I think that in general the procs file is more
> useful as a write API than a read API anyway, for the reasons you
> indicate there.
> 
> 
>> +       tsk = attach_get_task(cgrp, pidbuf);
>> +       if (IS_ERR(tsk))
>> +               return PTR_ERR(tsk);
>> +
>> +       /* attach thread group leader */
> 
> Should we check that this is in fact a thread group leader?
> 

No need actually, I added this check originally and then removed it, but
forgot to remove the comment.

>> +
>> +       /* attach all sub-threads */
>> +       rcu_read_lock();
> 
> cgroup_attach_task() calls synchronize_rcu(), so it doesn't seem
> likely that rcu_read_lock() is useful here, and might even deadlock?
> 
> What are you trying to protect against with the RCU lock?
> 

Ah yes, bad here. I am trying to protect the thread list.

>>        {
>> +               .name = "procs",
> 
> Maybe call it "cgroup.procs" to avoid name clashes in future? We had a
> debate a while back where I tried to get the cgroup files like "tasks"
> and "notify_on_release" prefixed with "cgroup." , which were argued
> against on grounds of backwards compatibility. But there's no
> compatibility issue here. The only question is whether it's too ugly
> to have the legacy filenames without a prefix and the new ones with a
> prefix.
> 

Yes it's ugly.. Is possible name clash of "procs" a kind of breaking
compatibility that should be avoid in any case?