[PATCH 0/1] hung_task: fix the broken rcu_lock_break() logic

[Oleg Nesterov <oleg@redhat.com>]

On 02/17, Oleg Nesterov wrote:
>
> On 02/17, Tetsuo Handa wrote:
> >
> > Oleg Nesterov wrote:
> > >     int main(void)
> > >     {
> > >             vfork();
> > >             pause();
> > >     }
> >^
> > Wow!
>
> and this reminds me... check_hung_uninterruptible_tasks() is very broken,
> it can crash the kernel itself. I'll resend my ancient fixes once again
> today, maybe this resend will be successful ;)

Please apply.

The patch was sent more than a year ago, I preserved the acks I got.

I was also going to add the PF_FREEZER_SKIP check in check_hung_task(),
but this was already done: f9fab10bbd768b0e5254e53a4a8477a94bfc4b96.
This means I lied, the program above no longer creates the problem
for check_hung_uninterruptible_tasks(). But that was not the point
for make-it-killable anyway.


I simply can't understand why do we have sysctl_hung_task_check_count,
may be someone can explain... OK, we need to do rcu_read_unlock()
from time to time even if need_resched() is not set, that is why
we have HUNG_TASK_BATCHING. Although perhaps it makes sense to do

	-	if (!--batch_count)
	+	if (!--batch_count || need_resched())
			rcu_lock_break();

but max_count? I guess I missed something obvious.

Oleg.