Re: [PATCH 2/3] mm/page_ref: Ensure page_ref_unfreeze is ordered against prior accesses

[Kirill A. Shutemov]

On Thu, Jun 08, 2017 at 11:38:21AM +0200, Vlastimil Babka wrote:
> On 06/06/2017 07:58 PM, Will Deacon wrote:
> > page_ref_freeze and page_ref_unfreeze are designed to be used as a pair,
> > wrapping a critical section where struct pages can be modified without
> > having to worry about consistency for a concurrent fast-GUP.
> > 
> > Whilst page_ref_freeze has full barrier semantics due to its use of
> > atomic_cmpxchg, page_ref_unfreeze is implemented using atomic_set, which
> > doesn't provide any barrier semantics and allows the operation to be
> > reordered with respect to page modifications in the critical section.
> > 
> > This patch ensures that page_ref_unfreeze is ordered after any critical
> > section updates, by invoking smp_mb__before_atomic() prior to the
> > atomic_set.
> > 
> > Cc: "Kirill A. Shutemov" <kirill.shutemov@xxxxxxxxxxxxxxx>
> > Acked-by: Steve Capper <steve.capper@xxxxxxx>
> > Signed-off-by: Will Deacon <will.deacon@xxxxxxx>
> 
> Undecided if it's really needed. This is IMHO not the classical case
> from Documentation/core-api/atomic_ops.rst where we have to make
> modifications visible before we let others see them? Here the one who is
> freezing is doing it so others can't get their page pin and interfere
> with the freezer's work.

Hm.. I'm not sure I'm getting what you are talking about. 

What would guarantee others to see changes to page before seeing page
unfreezed?

> >  include/linux/page_ref.h | 1 +
> >  1 file changed, 1 insertion(+)
> > 
> > diff --git a/include/linux/page_ref.h b/include/linux/page_ref.h
> > index 610e13271918..74d32d7905cb 100644
> > --- a/include/linux/page_ref.h
> > +++ b/include/linux/page_ref.h
> > @@ -174,6 +174,7 @@ static inline void page_ref_unfreeze(struct page *page, int count)
> >  	VM_BUG_ON_PAGE(page_count(page) != 0, page);
> >  	VM_BUG_ON(count == 0);
> >  
> > +	smp_mb__before_atomic();
> >  	atomic_set(&page->_refcount, count);

I *think* it should be smp_mb(), not __before_atomic(). atomic_set() is
not really atomic. For instance on x86 it's plain WRITE_ONCE() which CPU
would happily reorder.

> >  	if (page_ref_tracepoint_active(__tracepoint_page_ref_unfreeze))
> >  		__page_ref_unfreeze(page, count);
> > 
> 

-- 
 Kirill A. Shutemov