Re: [PATCH 0/1] kthreads: simplify !kthreadd_task logic, kill kthreadd_task_init_done

[Oleg Nesterov <oleg@redhat.com>]

On 09/18, Linus Torvalds wrote:
>
> On Fri, 18 Sep 2009, Oleg Nesterov wrote:
> > >  	rcu_scheduler_starting();
> > > -	kernel_thread(kernel_init, NULL, CLONE_FS | CLONE_SIGHAND);
> > > -	numa_default_policy();
> > >  	pid = kernel_thread(kthreadd, NULL, CLONE_FS | CLONE_FILES);
> > >  	kthreadd_task = find_task_by_pid_ns(pid, &init_pid_ns);
> > > +	kernel_thread(kernel_init, NULL, CLONE_FS | CLONE_SIGHAND);
> >
> > We can't do this. kernel_init() should run with ->pid == 1, we are
> > going to exec /sbin/init.
>
> Umm - why not just add CLONE_PID to the 'kthreadd' creation, then?

Not sure I understand... We don't have CLONE_PID, and kthreadd()
shouldn't have pid == 0 anyway.

But this is not needed. We can do alloc_pid() before 'kthreadd' creation,
then free_pid() before kernel_thread(kernel_init).

However, I am not sure we should do something like this. The patch
I sent is very simple, in essence it is one-liner which only does

	-       wake_up_process(kthreadd_task);
	+	if (kthreadd_task)
	+		wake_up_process(kthreadd_task);

this way we should not worry about the ordering, kthread_create()
can be called at any time, even before kthreadd creation (not that
I think this is really useful though).


But personally I still can't understand what happens. If we were
preempted somewhere before "kthreadd_task = find_task_by_pid_ns()"
initialization, then schedule_debug() should complain? rest_init()
runs under preempt_disable().

Oleg.