Re: [dpdk-dev] [PATCH 1/2] mempool: indicate the usages of multi memzones

[Olivier MATZ <olivier.matz@6wind.com>]

Hi Hemant,

On Wed, Dec 06, 2017 at 06:01:12PM +0530, Hemant Agrawal wrote:
> This is required for the optimizations w.r.t hw mempools.
> They will use different kind of optimizations if the buffers
> are from single contiguous memzone.
> 
> Signed-off-by: Hemant Agrawal <hemant.agrawal@nxp.com>
> ---
>  lib/librte_mempool/rte_mempool.c | 7 +++++--
>  lib/librte_mempool/rte_mempool.h | 5 +++++
>  2 files changed, 10 insertions(+), 2 deletions(-)
> 
> diff --git a/lib/librte_mempool/rte_mempool.c b/lib/librte_mempool/rte_mempool.c
> index d50dba4..9d3737c 100644
> --- a/lib/librte_mempool/rte_mempool.c
> +++ b/lib/librte_mempool/rte_mempool.c
> @@ -387,13 +387,16 @@ rte_mempool_populate_iova(struct rte_mempool *mp, char *vaddr,
>  	total_elt_sz = mp->header_size + mp->elt_size + mp->trailer_size;
>  
>  	/* Detect pool area has sufficient space for elements */
> -	if (mp->flags & MEMPOOL_F_CAPA_PHYS_CONTIG) {
> -		if (len < total_elt_sz * mp->size) {
> +	if (len < total_elt_sz * mp->size) {
> +		if (mp->flags & MEMPOOL_F_CAPA_PHYS_CONTIG) {
>  			RTE_LOG(ERR, MEMPOOL,
>  				"pool area %" PRIx64 " not enough\n",
>  				(uint64_t)len);
>  			return -ENOSPC;
>  		}
> +	} else {
> +		/* Memory will be allocated from multiple memzones */
> +		mp->flags |= MEMPOOL_F_MULTI_MEMZONE;
>  	}
>  
>  	memhdr = rte_zmalloc("MEMPOOL_MEMHDR", sizeof(*memhdr), 0);
> diff --git a/lib/librte_mempool/rte_mempool.h b/lib/librte_mempool/rte_mempool.h
> index 721227f..394a4fe 100644
> --- a/lib/librte_mempool/rte_mempool.h
> +++ b/lib/librte_mempool/rte_mempool.h
> @@ -292,6 +292,11 @@ struct rte_mempool {
>   */
>  #define MEMPOOL_F_CAPA_BLK_ALIGNED_OBJECTS 0x0080
>  
> +/* Indicates that the mempool buffers are allocated from multiple memzones
> + * the buffer may or may not be physically contiguous.
> + */
> +#define MEMPOOL_F_MULTI_MEMZONE 0x0100
> +
>  /**
>   * @internal When debug is enabled, store some statistics.
>   *
> -- 
> 2.7.4
> 

I'm not confortable with adding more and more flags, as I explained
here: http://dpdk.org/ml/archives/dev/2017-December/083909.html

It makes the generic code very complex, and probably buggy (many
flags are incompatible with other flags).

I'm thinking about moving the populate_* functions in the drivers
(this is described a bit more in the link above). What do you think
about this approach?

Thanks,
Olivier