Re: [PATCH] mm, swap: Remove unnecessary smp_rmb() in swap_type_to_swap_info()

[Peter Zijlstra]

On Thu, May 13, 2021 at 09:59:46PM -0400, Daniel Jordan wrote:
> On Thu, May 13, 2021 at 02:46:10PM +0200, Peter Zijlstra wrote:
> > Ah, I think I see what you meant to say, it would perhaps help if you
> > write it like so:
> > 
> > 
> > diff --git a/mm/swapfile.c b/mm/swapfile.c
> > index 149e77454e3c..94735248dcd2 100644
> > --- a/mm/swapfile.c
> > +++ b/mm/swapfile.c
> > @@ -99,11 +99,10 @@ atomic_t nr_rotate_swap = ATOMIC_INIT(0);
> >  
> >  static struct swap_info_struct *swap_type_to_swap_info(int type)
> >  {
> > -	if (type >= READ_ONCE(nr_swapfiles))
> > +	if (type >= MAX_SWAPFILES)
> >  		return NULL;
> >  
> > -	smp_rmb();	/* Pairs with smp_wmb in alloc_swap_info. */
> > -	return READ_ONCE(swap_info[type]);
> > +	return READ_ONCE(swap_info[type]); /* rcu_dereference() */
> >  }
> >  
> >  static inline unsigned char swap_count(unsigned char ent)
> > @@ -2869,14 +2868,11 @@ static struct swap_info_struct *alloc_swap_info(void)
> >  	}
> >  	if (type >= nr_swapfiles) {
> >  		p->type = type;
> > -		WRITE_ONCE(swap_info[type], p);
> >  		/*
> > -		 * Write swap_info[type] before nr_swapfiles, in case a
> > -		 * racing procfs swap_start() or swap_next() is reading them.
> > -		 * (We never shrink nr_swapfiles, we never free this entry.)
> > +		 * Publish the swap_info_struct.
> >  		 */
> > -		smp_wmb();
> > -		WRITE_ONCE(nr_swapfiles, nr_swapfiles + 1);
> > +		smp_store_release(&swap_info[type], p); /* rcu_assign_pointer() */
> > +		nr_swapfiles++;
> 
> Yes, this does help, I didn't understand why smp_wmb stayed around in
> the original post.
> 
> I think the only access smp_store_release() orders is p->type.  Wouldn't
> it be kinda inconsistent to only initialize that one field before
> publishing when many others would be done at the end of
> alloc_swap_info() after the fact?  p->type doesn't seem special.  For
> instance, get_swap_page_of_type() touches si->lock soon after it calls
> swap_type_to_swap_info(), so there could be a small window where there's
> a non-NULL si with an uninitialized lock.
> 
> It's not as if this is likely to be a problem in practice, it would just
> make it harder to understand why smp_store_release is there.  Maybe all
> we need is a WRITE_ONCE, or if it's really necessary for certain fields
> to be set before publication then move them up and explain?

You also care about the zero fill from kvzalloc(). Without the
smp_store_release() the zero-fill from the memset() might only be
visible 'late'.

Unless that also isn't a problem?