Re: [PATCH V2 1/3] sched/numa: Apply the scan delay to every vma instead of tasks

[Peter Zijlstra]

On Wed, Feb 01, 2023 at 01:32:20PM +0530, Raghavendra K T wrote:
> From: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>
> 
>  Avoid scanning new or very short-lived VMAs.
> 
> (Raghavendra: Add initialization in vm_area_dup())

Given this is a performance centric patch -- some sort of qualification
/ justification would be much appreciated.

Also, perhaps explain the rationale for the actual heuristics chosen.

> Signed-off-by: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>
> Signed-off-by: Raghavendra K T <raghavendra.kt@xxxxxxx>
> ---
>  include/linux/mm.h       |  9 +++++++++
>  include/linux/mm_types.h |  7 +++++++
>  kernel/fork.c            |  2 ++
>  kernel/sched/fair.c      | 17 +++++++++++++++++
>  4 files changed, 35 insertions(+)
> 
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index 974ccca609d2..74d9df1d8982 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -611,6 +611,14 @@ struct vm_operations_struct {
>  					  unsigned long addr);
>  };
>  
> +#ifdef CONFIG_NUMA_BALANCING
> +#define vma_numab_init(vma) do { (vma)->numab = NULL; } while (0)
> +#define vma_numab_free(vma) do { kfree((vma)->numab); } while (0)
> +#else
> +static inline void vma_numab_init(struct vm_area_struct *vma) {}
> +static inline void vma_numab_free(struct vm_area_struct *vma) {}
> +#endif /* CONFIG_NUMA_BALANCING */

I'm tripping over the inconsistency of macros and functions here. I'd
suggest making both cases functions.


> diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
> index 500e536796ca..e84f95a77321 100644
> --- a/include/linux/mm_types.h
> +++ b/include/linux/mm_types.h
> @@ -435,6 +435,10 @@ struct anon_vma_name {
>  	char name[];
>  };
>  
> +struct vma_numab {
> +	unsigned long next_scan;
> +};

I'm not sure what a numab is; contraction of new-kebab, something else?

While I appreciate short names, they'd ideally also make sense. If we
cannot come up with a better one, perhaps elucidate the reader with a
comment.

> +
>  /*
>   * This struct describes a virtual memory area. There is one of these
>   * per VM-area/task. A VM area is any part of the process virtual memory
> @@ -504,6 +508,9 @@ struct vm_area_struct {

> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index e4a0b8bd941c..060b241ce3c5 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -3015,6 +3015,23 @@ static void task_numa_work(struct callback_head *work)
>  		if (!vma_is_accessible(vma))
>  			continue;
>  
> +		/* Initialise new per-VMA NUMAB state. */
> +		if (!vma->numab) {
> +			vma->numab = kzalloc(sizeof(struct vma_numab), GFP_KERNEL);
> +			if (!vma->numab)
> +				continue;
> +
> +			vma->numab->next_scan = now +
> +				msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
> +		}
> +
> +		/*
> +		 * After the first scan is complete, delay the balancing scan
> +		 * for new VMAs.
> +		 */
> +		if (mm->numa_scan_seq && time_before(jiffies, vma->numab->next_scan))
> +			continue;

I think I sorta see why, but I'm thinking it would be good to include
more of the why in that comment.