Re: [PATCH v2 1/2] lockdep: Fix OOO unlock when hlocks need merging

[Imre Deak <imre.deak@intel.com>]

On Mon, May 27, 2019 at 05:02:51PM +0200, Peter Zijlstra wrote:
> On Fri, May 24, 2019 at 11:15:08PM +0300, Imre Deak wrote:
> > 
> > 	ww_mutex_lock(&ww_lock_a, &ww_ctx);
> > 
> > 	mutex_lock(&lock_c);
> > 
> > 	ww_mutex_lock(&ww_lock_b, &ww_ctx);
> > 
> > 	mutex_unlock(&lock_c);		(*)
> 
> > triggers the following WARN in __lock_release() when doing the unlock at *:
> > 
> > 	DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth - 1);
> > 
> > The problem is that the WARN check doesn't take into account the merging
> > of ww_lock_a and ww_lock_b which results in decreasing curr->lockdep_depth
> > by 2 not only 1.
> 
> Cute...
> 
> > diff --git a/kernel/locking/lockdep.c b/kernel/locking/lockdep.c
> > index c40fba54e324..967352d32af1 100644
> > --- a/kernel/locking/lockdep.c
> > +++ b/kernel/locking/lockdep.c
> > @@ -3714,7 +3714,7 @@ static int __lock_acquire(struct lockdep_map *lock, unsigned int subclass,
> >  				hlock->references = 2;
> >  			}
> >  
> > -			return 1;
> > +			return 2;
> >  		}
> >  	}
> >  
> > @@ -3920,22 +3920,33 @@ static struct held_lock *find_held_lock(struct task_struct *curr,
> >  }
> >  
> >  static int reacquire_held_locks(struct task_struct *curr, unsigned int depth,
> > -			      int idx)
> > +				int idx, bool *first_merged)
> >  {
> >  	struct held_lock *hlock;
> > +	int first_idx = idx;
> >  
> >  	if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
> >  		return 0;
> >  
> >  	for (hlock = curr->held_locks + idx; idx < depth; idx++, hlock++) {
> > -		if (!__lock_acquire(hlock->instance,
> > +		switch (__lock_acquire(hlock->instance,
> >  				    hlock_class(hlock)->subclass,
> >  				    hlock->trylock,
> >  				    hlock->read, hlock->check,
> >  				    hlock->hardirqs_off,
> >  				    hlock->nest_lock, hlock->acquire_ip,
> > -				    hlock->references, hlock->pin_count))
> > +				    hlock->references, hlock->pin_count)) {
> > +		case 0:
> >  			return 1;
> > +		case 1:
> > +			break;
> > +		case 2:
> > +			*first_merged = idx == first_idx;
> > +			break;
> > +		default:
> > +			WARN_ON(1);
> > +			return 0;
> > +		}
> >  	}
> >  	return 0;
> >  }
> 
> Does it work for you if I change it like so?

Yep, works this way, and yes thought later that canceling *first_merged
for idx!=first_idx was a bit strange (even if it still worked).

> 
> --- a/kernel/locking/lockdep.c
> +++ b/kernel/locking/lockdep.c
> @@ -3712,7 +3712,7 @@ static int __lock_acquire(struct lockdep
>  				hlock->references = 2;
>  			}
>  
> -			return 1;
> +			return 2;
>  		}
>  	}
>  
> @@ -3918,22 +3918,33 @@ static struct held_lock *find_held_lock(
>  }
>  
>  static int reacquire_held_locks(struct task_struct *curr, unsigned int depth,
> -			      int idx)
> +				int idx, unsigned int *merged)
>  {
>  	struct held_lock *hlock;
> +	int first_idx = idx;
>  
>  	if (DEBUG_LOCKS_WARN_ON(!irqs_disabled()))
>  		return 0;
>  
>  	for (hlock = curr->held_locks + idx; idx < depth; idx++, hlock++) {
> -		if (!__lock_acquire(hlock->instance,
> +		switch (__lock_acquire(hlock->instance,
>  				    hlock_class(hlock)->subclass,
>  				    hlock->trylock,
>  				    hlock->read, hlock->check,
>  				    hlock->hardirqs_off,
>  				    hlock->nest_lock, hlock->acquire_ip,
> -				    hlock->references, hlock->pin_count))
> +				    hlock->references, hlock->pin_count)) {
> +		case 0:
>  			return 1;
> +		case 1:
> +			break;
> +		case 2:
> +			*merged += (idx == first_idx);
> +			break;
> +		default:
> +			WARN_ON(1);
> +			return 0;
> +		}
>  	}
>  	return 0;
>  }
> @@ -3944,9 +3955,9 @@ __lock_set_class(struct lockdep_map *loc
>  		 unsigned long ip)
>  {
>  	struct task_struct *curr = current;
> +	unsigned int depth, merged = 0
>  	struct held_lock *hlock;
>  	struct lock_class *class;
> -	unsigned int depth;
>  	int i;
>  
>  	if (unlikely(!debug_locks))
> @@ -3971,14 +3982,14 @@ __lock_set_class(struct lockdep_map *loc
>  	curr->lockdep_depth = i;
>  	curr->curr_chain_key = hlock->prev_chain_key;
>  
> -	if (reacquire_held_locks(curr, depth, i))
> +	if (reacquire_held_locks(curr, depth, i, &merged))
>  		return 0;
>  
>  	/*
>  	 * I took it apart and put it back together again, except now I have
>  	 * these 'spare' parts.. where shall I put them.
>  	 */
> -	if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth))
> +	if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth - merged))
>  		return 0;
>  	return 1;
>  }
> @@ -3986,8 +3997,8 @@ __lock_set_class(struct lockdep_map *loc
>  static int __lock_downgrade(struct lockdep_map *lock, unsigned long ip)
>  {
>  	struct task_struct *curr = current;
> +	unsigned int depth, merged = 0;
>  	struct held_lock *hlock;
> -	unsigned int depth;
>  	int i;
>  
>  	if (unlikely(!debug_locks))
> @@ -4012,7 +4023,7 @@ static int __lock_downgrade(struct lockd
>  	hlock->read = 1;
>  	hlock->acquire_ip = ip;
>  
> -	if (reacquire_held_locks(curr, depth, i))
> +	if (reacquire_held_locks(curr, depth, i, &merged))
>  		return 0;
>  
>  	/*
> @@ -4021,6 +4032,11 @@ static int __lock_downgrade(struct lockd
>  	 */
>  	if (DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth))
>  		return 0;
> +
> +	/* Merging can't happen with unchanged classes.. */
> +	if (DEBUG_LOCKS_WARN_ON(merged))
> +		return 0;
> +
>  	return 1;
>  }
>  
> @@ -4035,8 +4051,8 @@ static int
>  __lock_release(struct lockdep_map *lock, int nested, unsigned long ip)
>  {
>  	struct task_struct *curr = current;
> +	unsigned int depth, merged = 1;
>  	struct held_lock *hlock;
> -	unsigned int depth;
>  	int i;
>  
>  	if (unlikely(!debug_locks))
> @@ -4091,14 +4107,15 @@ __lock_release(struct lockdep_map *lock,
>  	if (i == depth-1)
>  		return 1;
>  
> -	if (reacquire_held_locks(curr, depth, i + 1))
> +	if (reacquire_held_locks(curr, depth, i + 1, &merged))
>  		return 0;
>  
>  	/*
>  	 * We had N bottles of beer on the wall, we drank one, but now
>  	 * there's not N-1 bottles of beer left on the wall...
> +	 * Pouring two of the bottles together is acceptable.
>  	 */
> -	DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth-1);
> +	DEBUG_LOCKS_WARN_ON(curr->lockdep_depth != depth - merged);
>  
>  	/*
>  	 * Since reacquire_held_locks() would have called check_chain_key()