Re: [PATCH] lock/semaphore: Avoid a deadlock within __up()

[Byungchul Park <byungchul.park@lge.com>]

On Tue, Feb 02, 2016 at 09:13:55AM +0100, Ingo Molnar wrote:
> 
> * Byungchul Park <byungchul.park@lge.com> wrote:
> 
> > Since I faced a infinite recursive printk() bug, I've tried to propose
> > patches the title of which is "lib/spinlock_debug.c: prevent a recursive
> > cycle in the debug code". But I noticed the root problem cannot be fixed
> > by that, through some discussion thanks to Sergey and Peter. So I focused
> > on preventing the DEADLOCK.
> > 
> > -----8<-----
> > From 94a66990677735459a7790b637179d8600479639 Mon Sep 17 00:00:00 2001
> > From: Byungchul Park <byungchul.park@lge.com>
> > Date: Tue, 2 Feb 2016 15:35:48 +0900
> > Subject: [PATCH] lock/semaphore: Avoid a deadlock within __up()
> > 
> > When the semaphore __up() is called from within printk() with
> > console_sem.lock, a DEADLOCK can happen, since the wake_up_process() can
> > call printk() again, esp. if defined CONFIG_DEBUG_SPINLOCK. And the
> > wake_up_process() don't need to be within a critical section.
> > 
> > The scenario the bad thing can happen is,
> > 
> > printk
> >   console_trylock
> >   console_unlock
> >     up_console_sem
> >       up
> >         raw_spin_lock_irqsave(&sem->lock, flags)
> >         __up
> >           wake_up_process
> >             try_to_wake_up
> >               raw_spin_lock_irqsave(&p->pi_lock)
> >                 __spin_lock_debug
> >                   spin_dump
> >                     printk
> >                       console_trylock
> >                         raw_spin_lock_irqsave(&sem->lock, flags)
> > 
> >                         *** DEADLOCK ***
> > 
> > Signed-off-by: Byungchul Park <byungchul.park@lge.com>
> > ---
> >  kernel/locking/semaphore.c | 9 +++++++++
> >  1 file changed, 9 insertions(+)
> > 
> > diff --git a/kernel/locking/semaphore.c b/kernel/locking/semaphore.c
> > index b8120ab..d3a28dc 100644
> > --- a/kernel/locking/semaphore.c
> > +++ b/kernel/locking/semaphore.c
> > @@ -259,5 +259,14 @@ static noinline void __sched __up(struct semaphore *sem)
> >  						struct semaphore_waiter, list);
> >  	list_del(&waiter->list);
> >  	waiter->up = true;
> > +
> > +	/*
> > +	 * Trying to acquire this sem->lock in wake_up_process() leads a
> > +	 * DEADLOCK unless we unlock it here. For example, it's possile
> > +	 * in the case that called from within printk() since
> > +	 * wake_up_process() might call printk().
> > +	 */
> > +	raw_spin_unlock_irq(&sem->lock);
> >  	wake_up_process(waiter->task);
> > +	raw_spin_lock_irq(&sem->lock);
> 
> So I'm pretty sad about this solution, as it penalizes every semaphore user - 

Yeh... That was on my mind. Then... What about this alternative?

before
======
up
  spin_lock
  add count
  __up
    wake_up_process
  spin_unlock

thispatch
=========
up
  spin_lock
  add count
  __up
    spin_unlock
    wake_up_process
    spin_lock
  spin_unlock

alternative
===========
up
  spin_lock
  add count
  spin_unlock
  wake_up_process

This alternative does not have additional overhead and seems to be
reasonable, doesn't it? The reason why I proposed patches like this
including this alternative is that I thought it define the critical
section wider than it needs.

> while the deadlock is a really obscure one occuring within the scheduler or a 
> console driver, which are very narrow code paths!
> 
> (Also, please don't shout in comments, unless there's some really good reason to 
> do it.)

Do you mean the upper case e.i. DEADLOCK? Okay I will keep in mind.

> 
> Why doesn't spin_dump() break the console lock instead, if it detects that it's 
> spinning on it, before doing the printk()? It's a likely fail state anyway - and 
> this way we push any intrusive debug oriented action towards the unlikely fail 
> state.
> 
> Alternatively: why not improve down_trylock() to be lockless? The main reason for 
> the lockup is that a trylock op takes the semaphore spinlock unconditionally. 
> Which is fine for legacy code, but could perhaps be improved upon - I think we 
> could in fact do it without turning sem->count into atomics.
> 
> Alternatively #2: move printk() away from semaphores - it's pretty special code 
> anyway and semaphore semanthics are far from obvious.
> 

Thank you for your advice, and these approaches also look good. Could you
answer my question? If you don't think so, I will try it as you advised.

Thanks,
Byungchul

> Thanks,
> 
> 	Ingo