Re: [PATCH 1/4] mm/rmap: per anon_vma lock

[Yuanhan Liu <yuanhan.liu@linux.intel.com>]

On Fri, Nov 01, 2013 at 09:43:29AM +0100, Peter Zijlstra wrote:
> On Fri, Nov 01, 2013 at 03:54:24PM +0800, Yuanhan Liu wrote:
> > @@ -497,15 +495,20 @@ static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
> >   * anon_vma_interval_tree_post_update_vma().
> >   *
> >   * The entire update must be protected by exclusive mmap_sem and by
> > - * the root anon_vma's mutex.
> > + * the anon_vma's mutex.
> >   */
> >  static inline void
> >  anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
> >  {
> >  	struct anon_vma_chain *avc;
> >  
> > -	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
> > -		anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
> > +	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) {
> > +		struct anon_vma *anon_vma = avc->anon_vma;
> > +
> > +		anon_vma_lock_write(anon_vma);
> > +		anon_vma_interval_tree_remove(avc, &anon_vma->rb_root);
> > +		anon_vma_unlock_write(anon_vma);
> > +	}
> >  }
> >  
> >  static inline void
> > @@ -513,8 +516,13 @@ anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
> >  {
> >  	struct anon_vma_chain *avc;
> >  
> > -	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
> > -		anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
> > +	list_for_each_entry(avc, &vma->anon_vma_chain, same_vma) {
> > +		struct anon_vma *anon_vma = avc->anon_vma;
> > +
> > +		anon_vma_lock_write(anon_vma);
> > +		anon_vma_interval_tree_insert(avc, &anon_vma->rb_root);
> > +		anon_vma_unlock_write(anon_vma);
> > +	}
> >  }
> >  
> >  static int find_vma_links(struct mm_struct *mm, unsigned long addr,
> > @@ -781,7 +789,6 @@ again:			remove_next = 1 + (end > next->vm_end);
> >  	if (anon_vma) {
> >  		VM_BUG_ON(adjust_next && next->anon_vma &&
> >  			  anon_vma != next->anon_vma);
> > -		anon_vma_lock_write(anon_vma);
> >  		anon_vma_interval_tree_pre_update_vma(vma);
> >  		if (adjust_next)
> >  			anon_vma_interval_tree_pre_update_vma(next);
> > @@ -845,7 +852,6 @@ again:			remove_next = 1 + (end > next->vm_end);
> >  		anon_vma_interval_tree_post_update_vma(vma);
> >  		if (adjust_next)
> >  			anon_vma_interval_tree_post_update_vma(next);
> > -		anon_vma_unlock_write(anon_vma);
> >  	}
> >  	if (mapping)
> >  		mutex_unlock(&mapping->i_mmap_mutex);
> 
> AFAICT this isn't correct at all. We used to protect the vma interval
> tree with the root lock, now we don't.

We still use lock to protect anon_vma interval tree, but we lock our own
interval tree this time.

> All we've got left is the
> mmap_sem, but anon_vma chains can cross address-spaces and thus we're up
> some creek without no paddle.

Yep, however, you still need acquire the address-space crossed anon_vma's lock
to modify something.

Say, here is a chart: http://people.freedesktop.org/~yliu/anon_vma.png.
Let's take the 3rd chart as example. And assume we will unlink vma C.

And the steps are:
- lock c, and remove avc between c and C
- lock b, and remove avc between b and C
- lock a, and remove avc between a and C

Thanks.

	--yliu