HPA stats: Allow some derived stats.

However, we put them in their own struct, to avoid the messiness that the arena
has (mixing derived and non-derived stats in the arena_stats_t).

include/jemalloc/internal/hpa.h

@@ -6,12 +6,22 @@
 #include "jemalloc/internal/pai.h"
 #include "jemalloc/internal/psset.h"
 
+typedef struct hpa_shard_nonderived_stats_s hpa_shard_nonderived_stats_t;
+struct hpa_shard_nonderived_stats_s {
+	/*
+	 * The number of times we've purged a hugepage.  Each eviction purges a
+	 * single hugepage.
+	 *
+	 * Guarded by the grow mutex.
+	 */
+	uint64_t nevictions;
+};
+
 /* Completely derived; only used by CTL. */
 typedef struct hpa_shard_stats_s hpa_shard_stats_t;
 struct hpa_shard_stats_s {
 	psset_stats_t psset_stats;
-	/* The stat version of the nevictions counter. */
+	hpa_shard_nonderived_stats_t nonderived_stats;
-	uint64_t nevictions;
 };
 
 typedef struct hpa_shard_s hpa_shard_t;
@@ -73,12 +83,10 @@ struct hpa_shard_s {
 	emap_t *emap;
 
 	/*
-	 * The number of times we've purged a hugepage.  Each eviction purges a
+	 * Those stats which are copied directly into the CTL-centric hpa shard
-	 * single hugepage.
+	 * stats.
-	 *
-	 * Guarded by the grow mutex.
 	 */
-	uint64_t nevictions;
+	hpa_shard_nonderived_stats_t stats;
 };
 
 /*

src/ctl.c

@@ -3506,7 +3506,7 @@ stats_arenas_i_extents_j_index(tsdn_t *tsdn, const size_t *mib,
 }
 
 CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_nevictions,
-    arenas_i(mib[2])->astats->hpastats.nevictions, uint64_t);
+    arenas_i(mib[2])->astats->hpastats.nonderived_stats.nevictions, uint64_t);
 
 /* Full, huge */
 CTL_RO_CGEN(config_stats, stats_arenas_i_hpa_shard_full_slabs_npageslabs_huge,

src/hpa.c

@@ -74,7 +74,7 @@ hpa_shard_init(hpa_shard_t *shard, emap_t *emap, base_t *base,
 	shard->eden_len = 0;
 	shard->ind = ind;
 	shard->emap = emap;
-	shard->nevictions = 0;
+	shard->stats.nevictions = 0;
 
 	/*
 	 * Fill these in last, so that if an hpa_shard gets used despite
@@ -95,10 +95,17 @@ hpa_shard_init(hpa_shard_t *shard, emap_t *emap, base_t *base,
  * only combines the stats from one stats objet to another.  Hence the lack of
  * locking here.
  */
+static void
+hpa_shard_nonderived_stats_accum(hpa_shard_nonderived_stats_t *dst,
+    hpa_shard_nonderived_stats_t *src) {
+	dst->nevictions += src->nevictions;
+}
+
 void
 hpa_shard_stats_accum(hpa_shard_stats_t *dst, hpa_shard_stats_t *src) {
 	psset_stats_accum(&dst->psset_stats, &src->psset_stats);
-	dst->nevictions += src->nevictions;
+	hpa_shard_nonderived_stats_accum(&dst->nonderived_stats,
+	    &src->nonderived_stats);
 }
 
 void
@@ -107,7 +114,7 @@ hpa_shard_stats_merge(tsdn_t *tsdn, hpa_shard_t *shard,
 	malloc_mutex_lock(tsdn, &shard->grow_mtx);
 	malloc_mutex_lock(tsdn, &shard->mtx);
 	psset_stats_accum(&dst->psset_stats, &shard->psset.stats);
-	dst->nevictions += shard->nevictions;
+	hpa_shard_nonderived_stats_accum(&dst->nonderived_stats, &shard->stats);
 	malloc_mutex_unlock(tsdn, &shard->mtx);
 	malloc_mutex_unlock(tsdn, &shard->grow_mtx);
 }
@@ -290,7 +297,7 @@ hpa_handle_ps_eviction(tsdn_t *tsdn, hpa_shard_t *shard, hpdata_t *ps) {
 	malloc_mutex_assert_not_owner(tsdn, &shard->grow_mtx);
 
 	malloc_mutex_lock(tsdn, &shard->grow_mtx);
-	shard->nevictions++;
+	shard->stats.nevictions++;
 	hpdata_list_prepend(&shard->unused_slabs, ps);
 	malloc_mutex_unlock(tsdn, &shard->grow_mtx);
 }
@@ -431,7 +438,7 @@ hpa_alloc_psset(tsdn_t *tsdn, hpa_shard_t *shard, size_t size) {
 	malloc_mutex_lock(tsdn, &shard->mtx);
 	edata = edata_cache_small_get(tsdn, &shard->ecs);
 	if (edata == NULL) {
-		shard->nevictions++;
+		shard->stats.nevictions++;
 		malloc_mutex_unlock(tsdn, &shard->mtx);
 		malloc_mutex_unlock(tsdn, &shard->grow_mtx);
 		hpa_handle_ps_eviction(tsdn, shard, ps);
@@ -452,7 +459,7 @@ hpa_alloc_psset(tsdn_t *tsdn, hpa_shard_t *shard, size_t size) {
 
 		edata_cache_small_put(tsdn, &shard->ecs, edata);
 
-		shard->nevictions++;
+		shard->stats.nevictions++;
 		malloc_mutex_unlock(tsdn, &shard->mtx);
 		malloc_mutex_unlock(tsdn, &shard->grow_mtx);