Re: [PATCH] spinlock_debug: Fix various data races

[Marco Elver]

On Thu, 21 Nov 2019, Ingo Molnar wrote:

> 
> * Marco Elver <elver@xxxxxxxxxx> wrote:
> 
> >  static inline void debug_spin_lock_after(raw_spinlock_t *lock)
> >  {
> > -	lock->owner_cpu = raw_smp_processor_id();
> > -	lock->owner = current;
> > +	WRITE_ONCE(lock->owner_cpu, raw_smp_processor_id());
> > +	WRITE_ONCE(lock->owner, current);
> >  }
> 
> debug_spin_lock_after() runs inside the spinlock itself - why do these 
> writes have to be WRITE_ONCE()?
> 
> > @@ -197,8 +197,8 @@ static inline void debug_write_unlock(rwlock_t *lock)
> >  	RWLOCK_BUG_ON(lock->owner != current, lock, "wrong owner");
> >  	RWLOCK_BUG_ON(lock->owner_cpu != raw_smp_processor_id(),
> >  							lock, "wrong CPU");
> > -	lock->owner = SPINLOCK_OWNER_INIT;
> > -	lock->owner_cpu = -1;
> > +	WRITE_ONCE(lock->owner, SPINLOCK_OWNER_INIT);
> > +	WRITE_ONCE(lock->owner_cpu, -1);
> >  }
> 
> This too is running inside the critical section of the spinlock - why are 
> the WRITE_ONCE() calls necessary?

Although the writes are inside the critical section, they are read
concurrently outside the critical section, e.g. in
debug_spin_lock_before(). In other words, the WRITE_ONCE pair with the
READ_ONCE that are *outside* the critical section.

Thanks,
-- Marco

> Thanks,
> 
> 	Ingo