[dpdk-dev] [PATCH] mempool: avoid memory waste with large pagesize

[Olivier MATZ]

On 03/09/2016 03:29 AM, Stephen Hemminger wrote:
> If page size is large (like 64K on ARM) and object size is small
> then don't waste lots of memory by rounding up to page size.
> Instead, round up so that 1 or more objects all fit in a page.
> 
> This preserves the requirement that an object must not a page
> or virt2phys would break, and makes sure 62K is not wasted per mbuf.

You should specify that it only affects runs with "--no-huge".


> --- a/lib/librte_mempool/rte_mempool.c
> +++ b/lib/librte_mempool/rte_mempool.c
> @@ -300,18 +300,24 @@ rte_mempool_calc_obj_size(uint32_t elt_size, uint32_t flags,
>  	if (! rte_eal_has_hugepages()) {
>  		/*
>  		 * compute trailer size so that pool elements fit exactly in
> -		 * a standard page
> +		 * a standard page. If elements are smaller than a page
> +		 * then allow multiple elements per page
>  		 */
> -		int page_size = getpagesize();
> -		int new_size = page_size - sz->header_size - sz->elt_size;
> -		if (new_size < 0 || (unsigned int)new_size < sz->trailer_size) {
> +		unsigned page_size = getpagesize();
> +		uint32_t orig_size, new_size;
> +
> +		orig_size = sz->header_size + sz->elt_size;
> +		new_size = rte_align32pow2(orig_size);
> +		if (new_size > page_size) {
>  			printf("When hugepages are disabled, pool objects "
>  			       "can't exceed PAGE_SIZE: %d + %d + %d > %d\n",
>  			       sz->header_size, sz->elt_size, sz->trailer_size,
>  			       page_size);
>  			return 0;
>  		}
> -		sz->trailer_size = new_size;
> +
> +		sz->trailer_size = (new_size - orig_size)
> +			/ (page_size / new_size);
>  	}

Looks it does not work, did I miss something?

Examples:

# start with --no-huge
mp = rte_mempool_create("test", 128,
	35, 0, 0, NULL, NULL, NULL, NULL, SOCKET_ID_ANY, 0);
rte_mempool_dump(stdout, mp);

shows:
  header_size=64
  elt_size=40
  trailer_size=0
  total_obj_size=104       <<<<< should be 128?


# start with --no-huge
mp = rte_mempool_create("test", 128,
	191, 0, 0, NULL, NULL, NULL, NULL, SOCKET_ID_ANY,
	MEMPOOL_F_NO_CACHE_ALIGN);
rte_mempool_dump(stdout, mp);

shows:
  header_size=8
  elt_size=192
  trailer_size=3
  total_obj_size=203       <<<<< should be 256?


The RFC I've just submitted also aims to fix this issue
(but differently).

Regards,
Olivier