Re: [PATCH] kthreads: Fix startup synchronization boot crash

[Ingo Molnar <mingo@elte.hu>]

* Oleg Nesterov <oleg@redhat.com> wrote:

> On 09/01, Ingo Molnar wrote:
> >
> > * Oleg Nesterov <oleg@redhat.com> wrote:
> >
> > > On 09/01, Ingo Molnar wrote:
> > > >
> > > > * Oleg Nesterov <oleg@redhat.com> wrote:
> > > >
> > > > > Yes, this should work. But I _think_ we can make the better fix...
> > > > >
> > > > > I'll try to make the patch soon. Afaics we don't need
> > > > > kthreadd_task_init_done.
> > > >
> > > > ok.
> > >
> > > Just in case, the patch is ready. [...]
> >
> > yes - that's roughly the cleanup i referred to in the commit log.
> >
> > way too late for -rc8 though - the minimal fix i did _might_ be
> > eligible.
> >
> > agreed?
> 
> Agreed. Then I will sent the patch on top of this change.
> 
> But. May be your minimal patch needs a small tweak ?
> 
> rest_init()->complete(&kthreadd_task_init_done) assumes that exactly
> _one_ caller of kthread_create() can race with kernel_thread(kthreadd).
> Perhap we need complete_all() ?
> 
> 
> But I must admit, now I don't understand what happens,
> 
> 	The modification of that variable is protected by the BKL, but
> 	the _ordering_ of the initial task (which becomes the idle
> 	thread of CPU0) and the init task (which is spawned by the
> 	initial task) is not synchronized.
> 
> 	So we can occasionally end up init running sooner than
> 	rest_init()
> 
> How? rest_init() can't be preempted and it holds BKL. And 
> kernel_init() takes BKL before anything else. Confused...

it cannot be preempted but it can schedule anywhere - and the BKL 
will be dropped silently.

This is one of the biggest dangers of the BKL - rescheduling 
_somewhere_ in a huge codepath might change timings and 'breaks up 
the critical path' - breaking ancient assumptions.

	Ingo