Re: [RFC for stable 5.15 and 5.10] mm/memory: only copy anonymous pages during fork()

[Vlastimil Babka <vbabka@suse.cz>]

On 11/13/24 17:09, David Hildenbrand wrote:
>> ---
>> Hi, we've seen this in our 5.14 based kernel and it involved the out of
>> tree gpfs module, but I believe the same thing can happen in LTS's 5.10
>> and 5.15 without out of tree modules as well. So I'd like your opinion
>> on this fix before I propose it to stable as a non-standard
>> version-specific fix (I don't think we'd want to backport fb3d824d1a46
>> with prerequisities). Thanks.
> 
> I recall seeing+discussing this exact patch already a couple years ago :D
> 
> Ah, here is the 5.15 version
> 
> https://lkml.kernel.org/r/20221028075244.3112566-1-songyuanzheng@huawei.com
> 
> And the 5.10 version
> 
> https://lore.kernel.org/lkml/20221024094911.3054769-1-songyuanzheng@huawei.com/
> 
> 
> ... I could have sworn they got applied.
> 
> ... and in linux-5.10.y I see
> 
> commit 935a8b6202101d7f58fe9cd11287f9cec0d8dd32
> Author: Yuanzheng Song <songyuanzheng@huawei.com>
> Date:   Fri Oct 28 03:07:05 2022 +0000
> 
>      mm/memory: add non-anonymous page check in the copy_present_page()
> 
>      The vma->anon_vma of the child process may be NULL because
>      the entire vma does not contain anonymous pages. In this
>      case, a BUG will occur when the copy_present_page() passes
>      a copy of a non-anonymous page of that vma to the
>      page_add_new_anon_rmap() to set up new anonymous rmap.
> 
> 
> 
> Maybe you missed that the PageAnon() check is simply a couple of lines 
> further down in there?

No I was only looking at the 5.15 branch so far, and seems it was never
applied there, unlike 5.10. Weird. But thanks for the heads up.