[PATCH] mm/vma: Add VM_WARN_ON for commit_merge

[PATCH] mm/vma: Add VM_WARN_ON for commit_merge

[Liu Ye]

    Add VM_WARN_ON to prevent 'adjust' from accessing NULL pointers
    when 'adjust' is NULL and 'expanded' is false or 'adj_start' is
    not zero.
Signed-off-by: Liu Ye <liuye@kylinos.cn>
---
 mm/vma.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/mm/vma.c b/mm/vma.c
index 31c9c6f51c9f..36b5ac675504 100644
--- a/mm/vma.c
+++ b/mm/vma.c
@@ -641,6 +641,7 @@ static int commit_merge(struct vma_merge_struct *vmg,
 
 	init_multi_vma_prep(&vp, vmg->vma, adjust, remove, remove2);
 
+	VM_WARN_ON(!adjust && (!expanded || adj_start));
 	VM_WARN_ON(vp.anon_vma && adjust && adjust->anon_vma &&
 		   vp.anon_vma != adjust->anon_vma);
 
-- 
2.25.1

Re: [PATCH] mm/vma: Add VM_WARN_ON for commit_merge

[Andrew Morton]

On Thu, 16 Jan 2025 10:50:05 +0800 Liu Ye <liuye@kylinos.cn> wrote:

>     Add VM_WARN_ON to prevent 'adjust' from accessing NULL pointers
>     when 'adjust' is NULL and 'expanded' is false or 'adj_start' is
>     not zero.
>
> ...
>
> --- a/mm/vma.c
> +++ b/mm/vma.c
> @@ -641,6 +641,7 @@ static int commit_merge(struct vma_merge_struct *vmg,
>  
>  	init_multi_vma_prep(&vp, vmg->vma, adjust, remove, remove2);
>  
> +	VM_WARN_ON(!adjust && (!expanded || adj_start));
>  	VM_WARN_ON(vp.anon_vma && adjust && adjust->anon_vma &&
>  		   vp.anon_vma != adjust->anon_vma);
>  

This won't prevent a null deref.  It will emit a warning which
duplicates all the information which we're about to emit from the oops
handler.

Are there any reports of an oops from a NULL deref of `adjust'?

Re: [PATCH] mm/vma: Add VM_WARN_ON for commit_merge

[Lorenzo Stoakes]

Thanks for the patch!

It's not possible for this scenario to happen in practice, so this check just
adds confusion.

There are two callers - the one from vma_expand() which unconditionally sets
expand true, and the one from vma_merge_existing_range() (note that
vma_merge_new_range() ultimately invokes vma_expand()).

The vma_merge_existing_range() case will nearly always set expand, should a
merge be indicated. In the one instance where it will not, adjust will always be
set to a non-NULL value.

Unfortunately this is super unclear, which is a big hint that refactoring is
required - and it's something I've already started working on - so I will take
this as a signal to increase the priority on this.

So this has to be a no unfortunately, but thank you very much highlighting this
as it clearly indicates the need to improve this, which is hugely valuable!

Thanks, Lorenzo

On Thu, Jan 16, 2025 at 10:50:05AM +0800, Liu Ye wrote:
>     Add VM_WARN_ON to prevent 'adjust' from accessing NULL pointers
>     when 'adjust' is NULL and 'expanded' is false or 'adj_start' is
>     not zero.
> Signed-off-by: Liu Ye <liuye@kylinos.cn>
> ---
>  mm/vma.c | 1 +
>  1 file changed, 1 insertion(+)
>
> diff --git a/mm/vma.c b/mm/vma.c
> index 31c9c6f51c9f..36b5ac675504 100644
> --- a/mm/vma.c
> +++ b/mm/vma.c
> @@ -641,6 +641,7 @@ static int commit_merge(struct vma_merge_struct *vmg,
>
>  	init_multi_vma_prep(&vp, vmg->vma, adjust, remove, remove2);
>
> +	VM_WARN_ON(!adjust && (!expanded || adj_start));
>  	VM_WARN_ON(vp.anon_vma && adjust && adjust->anon_vma &&
>  		   vp.anon_vma != adjust->anon_vma);
>
> --
> 2.25.1
>