Re: [PATCH 3/4] mm/slub: attempt to find layouts up to 1/2 waste in calculate_order()

[Feng Tang]

On Fri, Sep 08, 2023 at 10:53:06PM +0800, Vlastimil Babka wrote:
> The main loop in calculate_order() currently tries to find an order with
> at most 1/4 waste. If that's impossible (for particular large object
> sizes), there's a fallback that will try to place one object within
> slab_max_order.
> 
> If we expand the loop boundary to also allow up to 1/2 waste as the last
> resort, we can remove the fallback and simplify the code, as the loop
> will find an order for such sizes as well. Note we don't need to allow
> more than 1/2 waste as that will never happen - calc_slab_order() would
> calculate more objects to fit, reducing waste below 1/2.
> 
> Sucessfully finding an order in the loop (compared to the fallback) will
> also have the benefit in trying to satisfy min_objects, because the
> fallback was passing 1. Thus the resulting slab orders might be larger
> (not because it would improve waste, but to reduce pressure on shared
> locks), which is one of the goals of calculate_order().
> 
> For example, with nr_cpus=1 and 4kB PAGE_SIZE, slub_max_order=3, before
> the patch we would get the following orders for these object sizes:
> 
>  2056 to 10920 - order-3 as selected by the loop
> 10928 to 12280 - order-2 due to fallback, as <1/4 waste is not possible
> 12288 to 32768 - order-3 as <1/4 waste is again possible
> 
> After the patch:
> 
> 2056 to 32768 - order-3, because even in the range of 10928 to 12280 we
>                 try to satisfy the calculated min_objects.
> 
> As a result the code is simpler and gives more consistent results.
 
Current code already tries the fraction "1" in the follwing 2 fallback
calls of calc_slab_order(), so trying fraction "2" makes sense to me.

Reviewed-by: Feng Tang <feng.tang@xxxxxxxxx>

Thanks,
Feng

> Signed-off-by: Vlastimil Babka <vbabka@xxxxxxx>
> ---
>  mm/slub.c | 14 ++++----------
>  1 file changed, 4 insertions(+), 10 deletions(-)
> 
> diff --git a/mm/slub.c b/mm/slub.c
> index 5c287d96b212..f04eb029d85a 100644
> --- a/mm/slub.c
> +++ b/mm/slub.c
> @@ -4171,23 +4171,17 @@ static inline int calculate_order(unsigned int size)
>  	 * the order can only result in same or less fractional waste, not more.
>  	 *
>  	 * If that fails, we increase the acceptable fraction of waste and try
> -	 * again.
> +	 * again. The last iteration with fraction of 1/2 would effectively
> +	 * accept any waste and give us the order determined by min_objects, as
> +	 * long as at least single object fits within slub_max_order.
>  	 */
> -	for (unsigned int fraction = 16; fraction >= 4; fraction /= 2) {
> +	for (unsigned int fraction = 16; fraction > 1; fraction /= 2) {
>  		order = calc_slab_order(size, min_objects, slub_max_order,
>  					fraction);
>  		if (order <= slub_max_order)
>  			return order;
>  	}
>  
> -	/*
> -	 * We were unable to place multiple objects in a slab. Now
> -	 * lets see if we can place a single object there.
> -	 */
> -	order = calc_slab_order(size, 1, slub_max_order, 1);
> -	if (order <= slub_max_order)
> -		return order;
> -
>  	/*
>  	 * Doh this slab cannot be placed using slub_max_order.
>  	 */
> -- 
> 2.42.0
> 
>