Re: [PATCH] mm/damon/vaddr: attempt per-vma lock during page table walk

[SeongJae Park <sj@kernel.org>]

On Tue, 12 May 2026 21:59:26 +0800 Kefeng Wang <wangkefeng.wang@huawei.com> wrote:

> 
> 
> On 5/12/2026 9:31 AM, SeongJae Park wrote:
> > On Mon, 11 May 2026 21:25:46 +0800 Kefeng Wang <wangkefeng.wang@huawei.com> wrote:
> > 
> >> Currently, DAMON virtual address operations use mmap_read_lock
> >> during page table walks, which can cause unnecessary contention
> >> under high concurrency.
> >>
> >> Introduce damon_va_walk_page_range() to first attempt acquiring a
> >> per-vma lock. If the VMA is found and the range is fully contained
> >> within it, the page table walk proceeds with the per-vma lock
> >> instead of mmap_read_lock.
> >>
> >> This optimization is particularly effective for damon_va_young()
> >> and damon_va_mkold(), which are frequently called and typically
> >> operate within a single VMA.
> > 
> > Makes sense.  Do you have some measurements?
> 
> In fact, I do not have performance-related tests.
> 
> > 
> >>
> >> Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
> > 
> > Looks good to me.  Nonetheless, because I'm not familiar with per-vma locking,
> > I'd like to wait for Sashiko review.
> > 
> 
> Sashiko review reports a issue about handling VMAs with the VM_PFNMAP
> flag inconsistently[1]，

For other readers who may want to read the Sashiko review and might want to
commeent, it is also available [1] at lore.

> We indeed do not need to handle the vma of
> VM_PFNMAP for damon, so a quick fix is as follows,
> 
> diff --git a/mm/damon/vaddr.c b/mm/damon/vaddr.c
> index eed2ae132a58..d27147603564 100644
> --- a/mm/damon/vaddr.c
> +++ b/mm/damon/vaddr.c
> @@ -251,8 +251,11 @@ static void damon_va_walk_page_range(struct 
> mm_struct *mm, unsigned long start,
>                  goto lock_mmap;
>          }
> 
> -       ops->walk_lock = PGWALK_VMA_RDLOCK_VERIFY;
> -       walk_page_range_vma(vma, start, end, ops, private);
> +       if (!(vma->vm_flags & VM_PFNMAP)) {
> +               ops->walk_lock = PGWALK_VMA_RDLOCK_VERIFY;
> +               walk_page_range_vma(vma, start, end, ops, private);
> +       }
> +
>          vma_end_read(vma);
>          return;
> 
> Any more comments?

Looks good to me, thank you!  Could you send v2 with this change?

> 
> Thanks.
> 
> [1] 
> https://sashiko.dev/#/patchset/20260511132546.1973270-1-wangkefeng.wang@huawei.com?part=1

[1] https://lore.kernel.org/20260512052628.83798C2BCC7@smtp.kernel.org


Thanks,
SJ

[...]