Re: [COUNTERPATCH] mm: avoid overflowing preempt_count() in mmu_take_all_locks()

[Andrea Arcangeli <aarcange@redhat.com>]

On Thu, Apr 01, 2010 at 06:49:52PM +0200, Peter Zijlstra wrote:
> On Thu, 2010-04-01 at 18:45 +0200, Peter Zijlstra wrote:
> > On Thu, 2010-04-01 at 18:18 +0200, Andrea Arcangeli wrote:
> > > On Thu, Apr 01, 2010 at 06:12:34PM +0200, Peter Zijlstra wrote:
> > > > One thing we can do there is to mutex_trylock() if we get the lock, see
> > > > if we've got the right object, if the trylock fails we can do the
> > > > refcount thing and sleep. That would allow the fast-path to remain a
> > > > single atomic.
> > > 
> > > But then how do you know which anon_vma_unlink has to decrease the
> > > refcount and which not? That info should need to be stored in the
> > > kernel stack, can't be stored in the vma. I guess it's feasible but
> > > passing that info around sounds more tricky than the trylock itself
> > > (adding params to those functions with int &refcount). 
> > 
> > I was thinking of something like:
> > 
> > struct anon_vma *page_lock_anon_vma(struct page *page)
> > {
> > 	struct anon_vma *anon_vma = NULL;
> > 	unsigned long anon_mapping;
> > 
> > 	rcu_read_lock();
> > 	anon_mapping = (unsigned long) ACCESS_ONCE(page->mapping);
> >         if ((anon_mapping & PAGE_MAPPING_FLAGS) != PAGE_MAPPING_ANON)              
> >                 goto out;                                                          
> >         if (!page_mapped(page))                                                    
> >                 goto out;                                                          
> > 
> >         anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);         
> >         if (!mutex_trylock(&anon_vma->lock)) {
> > 		if (atomic_inc_unless_zero(&anon_vma->ref)) {
> > 			rcu_read_unlock();
> > 			mutex_lock(&anon_vma->lock);
> > 			atomic_dec(&anon_vma->ref); /* ensure the lock pins it */
> > 		} else
> > 			anon_vma = NULL;
> > 	}
> > 	rcu_read_unlock();
> > 	
> >         return anon_vma;
> > }
> > 
> > void page_unlock_anon_vma(struct anon_vma *anon_vma)
> > {
> > 	mutex_unlock(&anon_vma->lock);	
> > }
> > 
> > Then anybody reaching ref==0 would only need to sync against the lock
> > before freeing.
> 
> Ah, there is a race where the dec after lock makes it 0, we could catch
> that by making it -1 and free in unlock_anon_vma().

You'd simply need to atomic_dec_and_test instead of atomic_dec above,
then you free it there above and return NULL.

The bug is to ever call atomic_dec instead of always
atomic_dec_and_test all over the place. I doubt you could fix it with
a -1.