Re: [PATCH v4] mm: use per_vma lock for MADV_DONTNEED

["Lai, Yi" <yi1.lai@linux.intel.com>]

On Tue, Aug 05, 2025 at 09:48:32AM +1200, Barry Song wrote:
> On Mon, Aug 4, 2025 at 7:57 PM David Hildenbrand <david@redhat.com> wrote:
> >
> > On 04.08.25 02:58, Lai, Yi wrote:
> > > Hi Barry Song,
> > >
> > > Greetings!
> > >
> > > I used Syzkaller and found that there is general protection fault in __pte_offset_map_lock in linux-next next-20250801.
> > >
> > > After bisection and the first bad commit is:
> > > "
> > > a6fde7add78d mm: use per_vma lock for MADV_DONTNEED
> > > "
> > >
> > > All detailed into can be found at:
> > > https://github.com/laifryiee/syzkaller_logs/tree/main/250803_193026___pte_offset_map_lock
> > > Syzkaller repro code:
> > > https://github.com/laifryiee/syzkaller_logs/tree/main/250803_193026___pte_offset_map_lock/repro.c
> > > Syzkaller repro syscall steps:
> [...]
> >
> > Skimming over the reproducer, we seem to have racing MADV_DONTNEED and
> > MADV_COLLAPSE on the same anon area, but the problem only shows up once
> > we tear down that MM.
> >
> 
> This seems to be where the race happens.
> Hi Lai, can you also double check if the below diff fixes the problem?
> 
> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> index 374a6a5193a7..6b40bdfd224c 100644
> --- a/mm/khugepaged.c
> +++ b/mm/khugepaged.c
> @@ -1172,11 +1172,11 @@ static int collapse_huge_page(struct mm_struct
> *mm, unsigned long address,
>         if (result != SCAN_SUCCEED)
>                 goto out_up_write;
>         /* check if the pmd is still valid */
> +       vma_start_write(vma);
>         result = check_pmd_still_valid(mm, address, pmd);
>         if (result != SCAN_SUCCEED)
>                 goto out_up_write;
> 
> -       vma_start_write(vma);
>         anon_vma_lock_write(vma->anon_vma);
> 
>         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, address,
>

Applied this change on top of both linux-next next-20250801 and
next-20250804 separately. Issue cannot be reproduced using the
reproducer.
 
> > If I would have to guess, I'd assume it's related to PT_RECLAIM
> > reclaiming empty page tables during MADV_DONTNEED -- but the kconfig
> > does not indicate that CONFIG_PT_RECLAIM was set.
> >
> > --
> > Cheers,
> >
> > David / dhildenb
> >
> 
> Thanks
> Barry