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

[Paul Menage]

On Wed, Jun 18, 2008 at 1:02 AM, Li Zefan <lizf at 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.

> - 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?

> +
> +       /* 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?

>        {
> +               .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.

Paul
_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers