Re: [PATCH 2/4] cgroup:add css_is_populated

[KAMEZAWA Hiroyuki]

On Mon, 19 Jan 2009 18:23:03 -0800
Paul Menage <menage@xxxxxxxxxx> wrote:

> On Mon, Jan 19, 2009 at 6:02 PM, KAMEZAWA Hiroyuki
> <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:
> > Ah, this is related to CSS ID scanning. No real problem to current codes.
> >
> > Now, in my patch, CSS ID is attached just after create().
> >
> > Then, "scan by CSS ID" can find a cgroup which is not populated yet.
> > I just wanted to skip them for avoiding mess.
> >
> > For example, css_tryget() can succeed against css which belongs to not-populated
> > cgroup. If creation of cgroup fails, it's destroyed and freed without RCU synchronize.
> > This may breaks CSS ID scanning's assumption that "we're safe under rcu_read_lock".
> > And allows destroy css while css->refcnt > 1.
> 
> So for the CSS ID case, we could solve this by not populating
> css_id->css until creation is guaranteed to have succeeded? (We'd
> still allocated the css_id along with the subsystem, just not complete
> its initialization). cgroup.c already knows about and hides the
> details of css_id, so this wouldn't be hard.
> 

Hmm, moving this call to after populate is not good because
id > max_id cannot be handled ;)

> +               if (ss->use_id)
> +                       if (alloc_css_id(ss, parent, cgrp))
> +                               goto err_destroy;
> +               /* At error, ->destroy() callback has to free assigned ID. */
>        }

Should I delay to set css_id->css pointer to valid value until the end of
populate() ? (add populage_css_id() call after cgroup_populate_dir()).

I'd like to write add-on patch to the patch [1/4]. (or update it.)
css_id->css == NULL case is handled now, anyway.



> The question is whether other users of css_tryget() might run into
> this problem, without using CSS IDs. But currently no-one's using
> css_tryget() apart from you, so that's a problem we can solve as it
> arises.
> 
yes ;)

> I think we're safe from css_get() being used in this case, since
> css_get() can only be used on css references obtained from a locked
> task, or from other pointers that are known to be ref-counted, which
> would be impossible if css_tryget() can't succeed on a css in the
> partially-completed state.
> 
> 

Thanks, I'll try some other patch. Maybe it's ok to delay updating css_id->css
pointer.

-Kame

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/