Re: [PATCH v4 3/3] mm/oom_kill: Have the OOM reaper and exit_mmap() traverse the maple tree in opposite orders

[Lorenzo Stoakes <lorenzo.stoakes@oracle.com>]

On Thu, Aug 14, 2025 at 09:55:55PM +0800, zhongjinji@honor.com wrote:
> From: zhongjinji <zhongjinji@honor.com>
>
> When a process is OOM killed, if the OOM reaper and the thread running
> exit_mmap() execute at the same time, both will traverse the vma's maple
> tree along the same path. They may easily unmap the same vma, causing them
> to compete for the pte spinlock. This increases unnecessary load, causing
> the execution time of the OOM reaper and the thread running exit_mmap() to
> increase.

You're not giving any numbers, and this seems pretty niche, you really
exiting that many processes with the reaper running at the exact same time
that this is an issue? Waiting on a spinlock also?

This commit message is very unconvincing.

>
> When a process exits, exit_mmap() traverses the vma's maple tree from low to high
> address. To reduce the chance of unmapping the same vma simultaneously,
> the OOM reaper should traverse vma's tree from high to low address. This reduces
> lock contention when unmapping the same vma.

Are they going to run through and do their work in exactly the same time,
or might one 'run past' the other and you still have an issue?

Seems very vague and timing dependent and again, not convincing.

>
> Signed-off-by: zhongjinji <zhongjinji@honor.com>
> ---
>  include/linux/mm.h | 3 +++
>  mm/oom_kill.c      | 9 +++++++--
>  2 files changed, 10 insertions(+), 2 deletions(-)
>
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index 0c44bb8ce544..b665ea3c30eb 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -923,6 +923,9 @@ static inline void vma_iter_set(struct vma_iterator *vmi, unsigned long addr)
>  #define for_each_vma_range(__vmi, __vma, __end)				\
>  	while (((__vma) = vma_find(&(__vmi), (__end))) != NULL)
>
> +#define for_each_vma_reverse(__vmi, __vma)					\
> +	while (((__vma) = vma_prev(&(__vmi))) != NULL)

Please don't casually add an undocumented public VMA iterator hidden in a
patch doing something else :)

Won't this skip the first VMA? Not sure this is really worth having as a
general thing anyway, it's not many people who want to do this in reverse.

> +
>  #ifdef CONFIG_SHMEM
>  /*
>   * The vma_is_shmem is not inline because it is used only by slow
> diff --git a/mm/oom_kill.c b/mm/oom_kill.c
> index 7ae4001e47c1..602d6836098a 100644
> --- a/mm/oom_kill.c
> +++ b/mm/oom_kill.c
> @@ -517,7 +517,7 @@ static bool __oom_reap_task_mm(struct mm_struct *mm)
>  {
>  	struct vm_area_struct *vma;
>  	bool ret = true;
> -	VMA_ITERATOR(vmi, mm, 0);
> +	VMA_ITERATOR(vmi, mm, ULONG_MAX);
>
>  	/*
>  	 * Tell all users of get_user/copy_from_user etc... that the content
> @@ -527,7 +527,12 @@ static bool __oom_reap_task_mm(struct mm_struct *mm)
>  	 */
>  	set_bit(MMF_UNSTABLE, &mm->flags);
>
> -	for_each_vma(vmi, vma) {
> +	/*
> +	 * When two tasks unmap the same vma at the same time, they may contend for the
> +	 * pte spinlock. To avoid traversing the same vma as exit_mmap unmap, traverse
> +	 * the vma maple tree in reverse order.
> +	 */

Except you won't necessarily avoid anything, as if one walker is faster
than the other they'll run ahead, plus of course they'll have a cross-over
where they share the same PTE anyway.

I feel like maybe you've got a fairly specific situation that indicates an
issue elsewhere and you're maybe solving the wrong problem here?

> +	for_each_vma_reverse(vmi, vma) {
>  		if (vma->vm_flags & (VM_HUGETLB|VM_PFNMAP))
>  			continue;
>
> --
> 2.17.1
>
>