Re: mm, something wring in page_lock_anon_vma_read()?

[Hugh Dickins]

On Fri, 19 May 2017, Xishi Qiu wrote:
> On 2017/5/19 16:52, Xishi Qiu wrote:
> > On 2017/5/18 17:46, Xishi Qiu wrote:
> > 
> >> Hi, my system triggers this bug, and the vmcore shows the anon_vma seems be freed.
> >> The kernel is RHEL 7.2, and the bug is hard to reproduce, so I don't know if it
> >> exists in mainline, any reply is welcome!
> >>
> > 
> > When we alloc anon_vma, we will init the value of anon_vma->root,
> > so can we set anon_vma->root to NULL when calling
> > anon_vma_free -> kmem_cache_free(anon_vma_cachep, anon_vma);
> > 
> > anon_vma_free()
> > 	...
> > 	anon_vma->root = NULL;
> > 	kmem_cache_free(anon_vma_cachep, anon_vma);
> > 
> > I find if we do this above, system boot failed, why?
> > 
> 
> If anon_vma was freed, we should not to access the root_anon_vma, because it maybe also
> freed(e.g. anon_vma == root_anon_vma), right?
> 
> page_lock_anon_vma_read()
> 	...
> 	anon_vma = (struct anon_vma *) (anon_mapping - PAGE_MAPPING_ANON);
> 	root_anon_vma = ACCESS_ONCE(anon_vma->root);
> 	if (down_read_trylock(&root_anon_vma->rwsem)) {  // it's not safe
> 	...
> 	if (!atomic_inc_not_zero(&anon_vma->refcount)) {  // check anon_vma was not freed
> 	...
> 	anon_vma_lock_read(anon_vma);  // it's safe
> 	...

You're ignoring the rcu_read_lock() on entry to page_lock_anon_vma_read(),
and the SLAB_DESTROY_BY_RCU (recently renamed SLAB_TYPESAFE_BY_RCU) nature
of the anon_vma_cachep kmem cache.  It is not safe to muck with anon_vma->
root in anon_vma_free(), others could still be looking at it.

Hugh