Re: [PATCH 1/2] signal: simplify deadlock-avoidance in lock_task_sighand()

[Oleg Nesterov <oleg@redhat.com>]

On 09/22, Steven Rostedt wrote:
>
> On Mon, 22 Sep 2014 18:44:37 +0200
> Oleg Nesterov <oleg@redhat.com> wrote:
>
> > __lock_task_sighand() does local_irq_save() to prevent the potential
> > deadlock, we can use preempt_disable() with the same effect. And in
> > this case we can do preempt_disable/enable + rcu_read_lock/unlock only
> > once outside of the main loop and simplify the code. This also shaves
> > 112 bytes from signal.o.
> >
> > With this patch the main loop runs with preemption disabled, but this
> > should be fine because restart is very unlikely: it can only happen if
> > we race with de_thread() and ->sighand is shared. And the latter is only
> > possible if CLONE_SIGHAND was used without CLONE_THREAD, most probably
> > nobody does this nowadays.
> >
> > Signed-off-by: Oleg Nesterov <oleg@redhat.com>
> > ---
> >  kernel/signal.c |   31 +++++++++++++------------------
> >  1 files changed, 13 insertions(+), 18 deletions(-)
> >
> > diff --git a/kernel/signal.c b/kernel/signal.c
> > index 8f0876f..61a1f55 100644
> > --- a/kernel/signal.c
> > +++ b/kernel/signal.c
> > @@ -1261,30 +1261,25 @@ struct sighand_struct *__lock_task_sighand(struct task_struct *tsk,
> >  					   unsigned long *flags)
> >  {
> >  	struct sighand_struct *sighand;
> > -
> > +	/*
> > +	 * We are going to do rcu_read_unlock() under spin_lock_irqsave().
> > +	 * Make sure we can not be preempted after rcu_read_lock(), see
> > +	 * rcu_read_unlock() comment header for details.
> > +	 */
> > +	preempt_disable();
>
> The sad part is, this is going to break -rt.

Hmm, why??

> That
> is, is -rt susceptible to this deadlock as well?

In fact this deadlock is not really possible in any case, scheduler locks
should be fine under ->siglock (for example, signal_wake_up() is called
under this lock).

But, the comment above rcu_read_unlock() says:

	Given that the set of locks acquired by rt_mutex_unlock() might change
	at any time, a somewhat more future-proofed approach is to make sure
	that that preemption never happens ...

so this patch doesn't try to change the rules.

But perhaps we can simply remove this preempt_disable/enable?

Or. We can shift rcu_read_unlock() from lock_task_sighand() to
unlock_task_sighand(). This way we can avoid preempt_disable too, but
I'd prefer to not do this.

Oleg.