Re: [dpdk-dev] [PATCH v2] lib/mempool: distinguish debug counters from cache and pool

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

Hi Joyce,

On Thu, Mar 18, 2021 at 07:20:22PM +0800, Joyce Kong wrote:
> If cache is enabled, objects will be retrieved/put from/to cache,
> subsequently from/to the common pool. Now the debug stats calculate
> the objects retrieved/put from/to cache and pool together, it is
> better to distinguish the data number from local cache and common
> pool.

This is indeed a very useful information, thanks for proposing this.

Please see some comments below.

> Signed-off-by: Joyce Kong <joyce.kong@arm.com>
> ---
>  lib/librte_mempool/rte_mempool.c | 12 ++++++
>  lib/librte_mempool/rte_mempool.h | 64 ++++++++++++++++++++++----------
>  2 files changed, 57 insertions(+), 19 deletions(-)
> 
> diff --git a/lib/librte_mempool/rte_mempool.c b/lib/librte_mempool/rte_mempool.c
> index afb1239c8..9cb69367a 100644
> --- a/lib/librte_mempool/rte_mempool.c
> +++ b/lib/librte_mempool/rte_mempool.c
> @@ -1244,8 +1244,14 @@ rte_mempool_dump(FILE *f, struct rte_mempool *mp)
>  	for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
>  		sum.put_bulk += mp->stats[lcore_id].put_bulk;
>  		sum.put_objs += mp->stats[lcore_id].put_objs;
> +		sum.put_objs_cache += mp->stats[lcore_id].put_objs_cache;
> +		sum.put_objs_pool += mp->stats[lcore_id].put_objs_pool;
> +		sum.put_objs_flush += mp->stats[lcore_id].put_objs_flush;
>  		sum.get_success_bulk += mp->stats[lcore_id].get_success_bulk;
>  		sum.get_success_objs += mp->stats[lcore_id].get_success_objs;
> +		sum.get_success_objs_cache += mp->stats[lcore_id].get_success_objs_cache;
> +		sum.get_success_objs_pool += mp->stats[lcore_id].get_success_objs_pool;
> +		sum.get_success_objs_refill += mp->stats[lcore_id].get_success_objs_refill;
>  		sum.get_fail_bulk += mp->stats[lcore_id].get_fail_bulk;
>  		sum.get_fail_objs += mp->stats[lcore_id].get_fail_objs;
>  		sum.get_success_blks += mp->stats[lcore_id].get_success_blks;
> @@ -1254,8 +1260,14 @@ rte_mempool_dump(FILE *f, struct rte_mempool *mp)
>  	fprintf(f, "  stats:\n");
>  	fprintf(f, "    put_bulk=%"PRIu64"\n", sum.put_bulk);
>  	fprintf(f, "    put_objs=%"PRIu64"\n", sum.put_objs);
> +	fprintf(f, "    put_objs_cache=%"PRIu64"\n", sum.put_objs_cache);
> +	fprintf(f, "    put_objs_pool=%"PRIu64"\n", sum.put_objs_pool);
> +	fprintf(f, "    put_objs_flush=%"PRIu64"\n", sum.put_objs_flush);
>  	fprintf(f, "    get_success_bulk=%"PRIu64"\n", sum.get_success_bulk);
>  	fprintf(f, "    get_success_objs=%"PRIu64"\n", sum.get_success_objs);
> +	fprintf(f, "    get_success_objs_cache=%"PRIu64"\n", sum.get_success_objs_cache);
> +	fprintf(f, "    get_success_objs_pool=%"PRIu64"\n", sum.get_success_objs_pool);
> +	fprintf(f, "    get_success_objs_refill=%"PRIu64"\n", sum.get_success_objs_refill);
>  	fprintf(f, "    get_fail_bulk=%"PRIu64"\n", sum.get_fail_bulk);
>  	fprintf(f, "    get_fail_objs=%"PRIu64"\n", sum.get_fail_objs);
>  	if (info.contig_block_size > 0) {
> diff --git a/lib/librte_mempool/rte_mempool.h b/lib/librte_mempool/rte_mempool.h
> index c551cf733..29d80d97e 100644
> --- a/lib/librte_mempool/rte_mempool.h
> +++ b/lib/librte_mempool/rte_mempool.h
> @@ -66,12 +66,18 @@ extern "C" {
>   * A structure that stores the mempool statistics (per-lcore).
>   */
>  struct rte_mempool_debug_stats {
> -	uint64_t put_bulk;         /**< Number of puts. */
> -	uint64_t put_objs;         /**< Number of objects successfully put. */
> -	uint64_t get_success_bulk; /**< Successful allocation number. */
> -	uint64_t get_success_objs; /**< Objects successfully allocated. */
> -	uint64_t get_fail_bulk;    /**< Failed allocation number. */
> -	uint64_t get_fail_objs;    /**< Objects that failed to be allocated. */
> +	uint64_t put_bulk;		  /**< Number of puts. */
> +	uint64_t put_objs;		  /**< Number of objects successfully put. */
> +	uint64_t put_objs_cache;	  /**< Number of objects successfully put to cache. */
> +	uint64_t put_objs_pool;		  /**< Number of objects successfully put to pool. */
> +	uint64_t put_objs_flush;	  /**< Number of flushing objects from cache to pool. */
> +	uint64_t get_success_bulk;	  /**< Successful allocation number. */
> +	uint64_t get_success_objs;	  /**< Objects successfully allocated. */
> +	uint64_t get_success_objs_cache;  /**< Objects successfully allocated from cache. */
> +	uint64_t get_success_objs_pool;	  /**< Objects successfully allocated from pool. */
> +	uint64_t get_success_objs_refill; /**< Number of refilling objects from pool to cache. */
> +	uint64_t get_fail_bulk;		  /**< Failed allocation number. */
> +	uint64_t get_fail_objs;		  /**< Objects that failed to be allocated. */

What about having instead the following new stats:

- put_common_pool_bulk: number of bulks enqueued in common pool
- put_common_pool_objs: number of objects enqueued in common pool
- get_common_pool_bulk: number of bulks dequeued from common pool
- get_common_pool_objs: number of objects dequeued from common pool

It looks easier to me to understand, compared to
flush/refill. Especially, having 'objs' in the name makes me think that
it counts a number of objects, but it's not the case.

>  	/** Successful allocation number of contiguous blocks. */
>  	uint64_t get_success_blks;
>  	/** Failed allocation number of contiguous blocks. */
> @@ -270,22 +276,34 @@ struct rte_mempool {
>   *   Number to add to the object-oriented statistics.
>   */
>  #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
> -#define __MEMPOOL_STAT_ADD(mp, name, n) do {                    \
> -		unsigned __lcore_id = rte_lcore_id();           \
> -		if (__lcore_id < RTE_MAX_LCORE) {               \
> +#define __MEMPOOL_STAT_ADD(mp, name, n) do {			\
> +		unsigned __lcore_id = rte_lcore_id();		\
> +		if (__lcore_id < RTE_MAX_LCORE) {		\
>  			mp->stats[__lcore_id].name##_objs += n;	\
> -			mp->stats[__lcore_id].name##_bulk += 1;	\
> -		}                                               \
> -	} while(0)
> -#define __MEMPOOL_CONTIG_BLOCKS_STAT_ADD(mp, name, n) do {                    \
> -		unsigned int __lcore_id = rte_lcore_id();       \
> -		if (__lcore_id < RTE_MAX_LCORE) {               \
> +			mp->stats[__lcore_id].name##_bulk += 1; \
> +		}						\
> +	} while (0)
> +#define __MEMPOOL_OBJS_STAT_ADD(mp, name1, name2, n) do {	\
> +		unsigned __lcore_id = rte_lcore_id();		\
> +		if (__lcore_id < RTE_MAX_LCORE)			\
> +			mp->stats[__lcore_id].name1##_objs_##name2 += n;	\
> +	} while (0)
> +#define __MEMPOOL_OBJS_STAT_SUB(mp, name1, name2, n) do {	\
> +		unsigned __lcore_id = rte_lcore_id();		\
> +		if (__lcore_id < RTE_MAX_LCORE)			\
> +			mp->stats[__lcore_id].name1##_objs_##name2 -= n;	\
> +	} while (0)
> +#define __MEMPOOL_CONTIG_BLOCKS_STAT_ADD(mp, name, n) do {	\
> +		unsigned int __lcore_id = rte_lcore_id();	\
> +		if (__lcore_id < RTE_MAX_LCORE) {		\
>  			mp->stats[__lcore_id].name##_blks += n;	\
>  			mp->stats[__lcore_id].name##_bulk += 1;	\
> -		}                                               \
> +		}						\
>  	} while (0)

There are too many stats macros. I think that the original
__MEMPOOL_STAT_ADD() macro can be reused if using the stats name I'm
suggesting.

Else, if not using names ending with _objs and _bulks, it would be
better to rework the macro to something more generic, like this:

  #define __MEMPOOL_STAT_ADD(mp, name, n) do {	\
  		unsigned __lcore_id = rte_lcore_id();		\
  		if (__lcore_id < RTE_MAX_LCORE)			\
  			mp->stats[__lcore_id].name += n;	\
  	} while (0)

In this case, it could replace both __MEMPOOL_STAT_ADD() and
__MEMPOOL_CONTIG_BLOCKS_STAT_ADD(), and could be in a separate patch,
before this one. It would replace the 6 macro calls by 12, but would
be clearer because one can see the real field name.
Eventually the macro could take the lcore_id as a parameter.

>  #else
> -#define __MEMPOOL_STAT_ADD(mp, name, n) do {} while(0)
> +#define __MEMPOOL_STAT_ADD(mp, name, n) do {} while (0)
> +#define __MEMPOOL_OBJS_STAT_ADD(mp, name1, name2, n) do {} while (0)
> +#define __MEMPOOL_OBJS_STAT_SUB(mp, name1, nmae2, n) do {} while (0)
>  #define __MEMPOOL_CONTIG_BLOCKS_STAT_ADD(mp, name, n) do {} while (0)
>  #endif
>  
> @@ -1305,10 +1323,13 @@ __mempool_generic_put(struct rte_mempool *mp, void * const *obj_table,
>  
>  	/* Add elements back into the cache */
>  	rte_memcpy(&cache_objs[0], obj_table, sizeof(void *) * n);
> -
> +	__MEMPOOL_OBJS_STAT_ADD(mp, put, cache, n);
>  	cache->len += n;
>  
>  	if (cache->len >= cache->flushthresh) {
> +		__MEMPOOL_OBJS_STAT_SUB(mp, put, cache, cache->len - cache->size);

I don't think it is a good idea to decrease a statistic counter.
If using put_common_pool_bulk/put_common_pool_objs, I think the code could
go in rte_mempool_ops_enqueue_bulk() (and in rte_mempool_ops_dequeue_bulk() for
'get' stats).


> +		__MEMPOOL_OBJS_STAT_ADD(mp, put, pool, cache->len - cache->size);
> +		__MEMPOOL_OBJS_STAT_ADD(mp, put, flush, 1);
>  		rte_mempool_ops_enqueue_bulk(mp, &cache->objs[cache->size],
>  				cache->len - cache->size);
>  		cache->len = cache->size;
> @@ -1318,6 +1339,7 @@ __mempool_generic_put(struct rte_mempool *mp, void * const *obj_table,
>  
>  ring_enqueue:
>  
> +	__MEMPOOL_OBJS_STAT_ADD(mp, put, pool, n);
>  	/* push remaining objects in ring */
>  #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
>  	if (rte_mempool_ops_enqueue_bulk(mp, obj_table, n) < 0)
> @@ -1437,6 +1459,7 @@ __mempool_generic_get(struct rte_mempool *mp, void **obj_table,
>  			goto ring_dequeue;
>  		}
>  
> +		__MEMPOOL_OBJS_STAT_ADD(mp, get_success, refill, 1);
>  		cache->len += req;
>  	}
>  
> @@ -1447,6 +1470,7 @@ __mempool_generic_get(struct rte_mempool *mp, void **obj_table,
>  	cache->len -= n;
>  
>  	__MEMPOOL_STAT_ADD(mp, get_success, n);
> +	__MEMPOOL_OBJS_STAT_ADD(mp, get_success, cache, n);
>  
>  	return 0;
>  
> @@ -1457,8 +1481,10 @@ __mempool_generic_get(struct rte_mempool *mp, void **obj_table,
>  
>  	if (ret < 0)
>  		__MEMPOOL_STAT_ADD(mp, get_fail, n);
> -	else
> +	else {
>  		__MEMPOOL_STAT_ADD(mp, get_success, n);
> +		__MEMPOOL_OBJS_STAT_ADD(mp, get_success, pool, n);
> +	}
>  
>  	return ret;
>  }
> -- 
> 2.30.0
>