Re: "reuse mergeable anon_vma as parent when fork" causes a crash on s390

[Konstantin Khlebnikov <khlebnikov@yandex-team.ru>]

On 10/10/2019 06.15, Wei Yang wrote:
> On Thu, Oct 10, 2019 at 10:36:01AM +0800, Wei Yang wrote:
>> Hi, Qian, Shakeel
>>
>> Thanks for testing.
>>
>> Sounds I missed some case to handle. anon_vma_clone() now would be called in
>> vma_adjust, which is a different case when it is introduced.
>>
> 
> Well, I have to correct my statement. The reason is we may did something more
> in anon_vma_clone().
> 
> Here is a quick fix, while I need to go through all the cases carefully.

Oops, I've overlooked this case too.

You have to check src->anon_vma
otherwise in  __split_vma or copy_vma dst could pick completely random anon_vma.

Also checking prev will not hurt, just to be sure.

So, something like this should work:

if (!dst->anon_vma && src->anon_vma &&
     prev && pprev && pprev->anon_vma == src->anon_vma)
       dst->anon_vma = prev->anon_vma;

> 
> diff --git a/mm/rmap.c b/mm/rmap.c
> index 12f6c3d7fd9d..2844f442208d 100644
> --- a/mm/rmap.c
> +++ b/mm/rmap.c
> @@ -271,7 +271,7 @@ int anon_vma_clone(struct vm_area_struct *dst, struct vm_area_struct *src)
>           * 1. Parent has vm_prev, which implies we have vm_prev.
>           * 2. Parent and its vm_prev have the same anon_vma.
>           */
> -       if (pprev && pprev->anon_vma == src->anon_vma)
> +       if (!dst->anon_vma && pprev && pprev->anon_vma == src->anon_vma)
>                  dst->anon_vma = prev->anon_vma;
>   
>          list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
> 
>> BTW, do you have the specific test case? So that I could verify my change. The
>> kernel build test doesn't trigger this.
>>
>> Thanks a lot :-)
>>
>> On Wed, Oct 09, 2019 at 03:21:11PM -0700, Shakeel Butt wrote:
>> -- 
>> Wei Yang
>> Help you, Help me
>