[Question] Is there a race window between __anon_vma_prepare() with page_lock_anon_vma_read() ?

[Question] Is there a race window between __anon_vma_prepare() with page_lock_anon_vma_read() ?

[Miaohe Lin]

Hi all,
I am investigating the rmap code, and I found the below possible race window:

CPU 1								CPU 2
-----								-----
page_lock_anon_vma_read
  rcu_read_lock
  /* We assume anon_vam == root_anon_vma in this case. */
  root_anon_vma = READ_ONCE(anon_vma->root);
  root_anon_vma is *released* somewhere unfortunately.
  down_read_trylock(&root_anon_vma->rwsem)
								__anon_vma_prepare
								  anon_vma_alloc
								    root_anon_vma is *allocated* here.
								    init_rwsem(&anon_vma->rwsem);
  !page_mapped(page)
    up_read(&root_anon_vma->rwsem); -- *Oops!*

root_anon_vma->rwsem is reinitialized after locked. And reinitialized anon_vma->rwsem will be
unlocked without lock first.

I think this could happen due to the subtle SLAB_TYPESAFE_BY_RCU. But only can occur when anon_vma
is root anon_vma or they won't operate on the same rwsem.
Is this will really happen or Am I miss something ? Any reply would be very grateful.
Many Thanks! :)

Re: [Question] Is there a race window between __anon_vma_prepare() with page_lock_anon_vma_read() ?

[Miaohe Lin]

On 2021/5/14 17:16, Miaohe Lin wrote:
> Hi all,
> I am investigating the rmap code, and I found the below possible race window:
> 
> CPU 1								CPU 2
> -----								-----
> page_lock_anon_vma_read
>   rcu_read_lock
>   /* We assume anon_vam == root_anon_vma in this case. */
>   root_anon_vma = READ_ONCE(anon_vma->root);
>   root_anon_vma is *released* somewhere unfortunately.
>   down_read_trylock(&root_anon_vma->rwsem)
> 								__anon_vma_prepare
> 								  anon_vma_alloc
> 								    root_anon_vma is *allocated* here.
> 								    init_rwsem(&anon_vma->rwsem);
>   !page_mapped(page)
>     up_read(&root_anon_vma->rwsem); -- *Oops!*
> 
> root_anon_vma->rwsem is reinitialized after locked. And reinitialized anon_vma->rwsem will be
> unlocked without lock first.
> 
> I think this could happen due to the subtle SLAB_TYPESAFE_BY_RCU. But only can occur when anon_vma
> is root anon_vma or they won't operate on the same rwsem.
> Is this will really happen or Am I miss something ? Any reply would be very grateful.
> Many Thanks! :)

Any reply would be very grateful.
Many thanks!

>