[PATCH] mempool: split statistics from debug

[Stephen Hemminger]

On Sun, 30 Oct 2022 15:04:18 +0100
Morten Brørup <mb at smartsharesystems.com> wrote:

> > From: Morten Brørup [mailto:mb at smartsharesystems.com]
> > Sent: Sunday, 30 October 2022 12.55
> > 
> > Split statistics from debug, to make mempool statistics available
> > without
> > the performance cost of continuously validating the cookies in the
> > mempool
> > elements.  
> 
> mempool_perf_autotest shows that the rate_persec drops to a third (-66 %) when enabling full mempool debug - quite prohibitive!
> 
> With this patch, the performance cost is much lower. I don't have detailed test results, but the initial tests without mempool cache show a performance drop of -11 %. Significant, but not prohibitive.


One trick to avoid conditional in fast path would be to add a dummy stats[] per core.

Another would be to move the fast path get/put stats into the mempool_cache.
The current model has stats[] per cpu always on another cache line.

diff --git a/lib/mempool/rte_mempool.h b/lib/mempool/rte_mempool.h
index 1f5707f46a21..87905b7286a6 100644
--- a/lib/mempool/rte_mempool.h
+++ b/lib/mempool/rte_mempool.h
@@ -236,8 +236,10 @@ struct rte_mempool {
        struct rte_mempool_memhdr_list mem_list; /**< List of memory chunks */
 
 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
-       /** Per-lcore statistics. */
-       struct rte_mempool_debug_stats stats[RTE_MAX_LCORE];
+       /** Per-lcore statistics.
+        * Allocate one additional per-cpu slot for non-DPDK threads
+        */
+       struct rte_mempool_debug_stats stats[RTE_MAX_LCORE + 1];
 #endif
 }  __rte_cache_aligned;
 
@@ -302,10 +304,7 @@ struct rte_mempool {
  */
 #ifdef RTE_LIBRTE_MEMPOOL_DEBUG
 #define RTE_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;        \
-               }                                               \
+       (mp)->stats[rte_lcore_id()].name += n;
        } while (0)
 #else
 #define RTE_MEMPOOL_STAT_ADD(mp, name, n) do {} while (0)