Re: [PATCH 3/8] md: raid5: use refcount_t for reference counting instead atomic_t

[Shaohua Li <shli@kernel.org>]

On Wed, May 23, 2018 at 07:49:04PM +0200, Peter Zijlstra wrote:
> On Wed, May 23, 2018 at 06:21:19AM -0700, Matthew Wilcox wrote:
> > On Wed, May 09, 2018 at 09:36:40PM +0200, Sebastian Andrzej Siewior wrote:
> > > refcount_t type and corresponding API should be used instead of atomic_t when
> > > the variable is used as a reference counter. This allows to avoid accidental
> > > refcounter overflows that might lead to use-after-free situations.
> > > 
> > > Most changes are 1:1 replacements except for
> > > 	BUG_ON(atomic_inc_return(&sh->count) != 1);
> > > 
> > > which has been turned into
> > >         refcount_inc(&sh->count);
> > >         BUG_ON(refcount_read(&sh->count) != 1);
> > 
> > @@ -5387,7 +5387,8 @@ static struct stripe_head *__get_priority_stripe(struct
> > +r5conf *conf, int group)
> >                 sh->group = NULL;
> >         }
> >         list_del_init(&sh->lru);
> > -       BUG_ON(atomic_inc_return(&sh->count) != 1);
> > +       refcount_inc(&sh->count);
> > +	BUG_ON(refcount_read(&sh->count) != 1);
> >         return sh;
> >  }
> > 
> > 
> > That's the only problematic usage.  And I think what it's really saying is:
> > 
> > 	BUG_ON(refcount_read(&sh->count) != 0);
> > 	refcount_set(&sh->count, 1);
> > 
> > With that, this looks like a reasonable use of refcount_t to me.
> 
> I'm not so sure, look at:
> 
>   r5c_do_reclaim():
> 
> 	if (!list_empty(&sh->lru) &&
> 	    !test_bit(STRIPE_HANDLE, &sh->state) &&
> 	    atomic_read(&sh->count) == 0) {
> 	      r5c_flush_stripe(cond, sh)
> 
> Which does:
> 
>   r5c_flush_stripe():
> 
> 	atomic_inc(&sh->count);
> 
> Which is another inc-from-zero. Also, having sh's with count==0 in a
> list is counter to the concept of refcounts and smells like usage-counts
> to me. For refcount 0 really means deads and gone.
> 
> If this really is supposed to be a refcount, someone more familiar with
> the raid5 should do the patch and write a comprehensive changelog on it.

I don't know what is changed in the refcount, such raid5 change has attempted
before and didn't work. 0 for the stripe count is a valid usage and we do
inc-from-zero in several places.

Thanks,
Shaohua