Re: [PATCH 2/2] kthread, cgroup: close race window where new kthreads can be migrated to non-root cgroups

[Oleg Nesterov <oleg@redhat.com>]

On 03/16, Tejun Heo wrote:
>
> > --- x/kernel/kthread.c
> > +++ x/kernel/kthread.c
> > @@ -226,6 +226,7 @@
> >  	ret = -EINTR;
> >  	if (!test_bit(KTHREAD_SHOULD_STOP, &self->flags)) {
> >  		__kthread_parkme(self);
> > +		current->flags &= ~PF_IDONTLIKECGROUPS;
> >  		ret = threadfn(data);
> >  	}
> >  	do_exit(ret);
> > @@ -537,7 +538,7 @@
> >  	set_cpus_allowed_ptr(tsk, cpu_all_mask);
> >  	set_mems_allowed(node_states[N_MEMORY]);
> >
> > -	current->flags |= PF_NOFREEZE;
> > +	current->flags |= (PF_NOFREEZE | PF_IDONTLIKECGROUPS);
> >
> >  	for (;;) {
> >  		set_current_state(TASK_INTERRUPTIBLE);
> > --- x/kernel/cgroup/cgroup.c
> > +++ x/kernel/cgroup/cgroup.c
> > @@ -2429,7 +2429,7 @@
> >  	 * trapped in a cpuset, or RT worker may be born in a cgroup
> >  	 * with no rt_runtime allocated.  Just say no.
> >  	 */
> > -	if (tsk == kthreadd_task || (tsk->flags & PF_NO_SETAFFINITY)) {
> > +	if (tsk->flags & (PF_NO_SETAFFINITY | PF_IDONTLIKECGROUPS)) {
> >  		ret = -EINVAL;
> >  		goto out_unlock_rcu;
> >  	}
>
> Absolutely.  If we're willing to spend a PF flag on it, we can
> properly wait for it too instead of failing it.

Or we can add another "unsigned no_cgroups:1" bit into task_struct,
not sure.

Anyway, I do not understand the PF_NO_SETAFFINITY check in
__cgroup_procs_write(). task_can_attach() checks it too, so cgroups
can't change the affinity. Imo something explicit like no_cgroups
makes more sense.

Oleg.