Use exponential series to size extents.

If virtual memory is retained, allocate extents such that their sizes
form an exponentially growing series.  This limits the number of
disjoint virtual memory ranges so that extent merging can be effective
even if multiple arenas' extent allocation requests are highly
interleaved.

This resolves #462.

include/jemalloc/internal/arena.h

@@ -232,6 +232,15 @@ struct arena_s {
 		void		*extent_hooks_pun;
 	};
 
+	/*
+	 * Next extent size class in a growing series to use when satisfying a
+	 * request via the extent hooks (only if !config_munmap).  This limits
+	 * the number of disjoint virtual memory ranges so that extent merging
+	 * can be effective even if multiple arenas' extent allocation requests
+	 * are highly interleaved.
+	 */
+	pszind_t		extent_grow_next;
+
 	/* Cache of extent structures that were allocated via base_alloc(). */
 	ql_head(extent_t)	extent_cache;
 	malloc_mutex_t		extent_cache_mtx;

src/arena.c

@@ -1686,6 +1686,9 @@ arena_new(tsdn_t *tsdn, unsigned ind)
 
 	arena->extent_hooks = (extent_hooks_t *)&extent_hooks_default;
 
+	if (!config_munmap)
+		arena->extent_grow_next = psz2ind(HUGEPAGE);
+
 	ql_new(&arena->extent_cache);
 	if (malloc_mutex_init(&arena->extent_cache_mtx, "arena_extent_cache",
 	    WITNESS_RANK_ARENA_EXTENT_CACHE))

src/extent.c

@@ -265,6 +265,41 @@ extent_interior_register(tsdn_t *tsdn, rtree_ctx_t *rtree_ctx,
 	}
 }
 
+static void
+extent_gprof_add(tsdn_t *tsdn, const extent_t *extent)
+{
+
+	cassert(config_prof);
+
+	if (opt_prof && extent_active_get(extent)) {
+		size_t nadd = extent_size_get(extent) >> LG_PAGE;
+		size_t cur = atomic_add_zu(&curpages, nadd);
+		size_t high = atomic_read_zu(&highpages);
+		while (cur > high && atomic_cas_zu(&highpages, high, cur)) {
+			/*
+			 * Don't refresh cur, because it may have decreased
+			 * since this thread lost the highpages update race.
+			 */
+			high = atomic_read_zu(&highpages);
+		}
+		if (cur > high && prof_gdump_get_unlocked())
+			prof_gdump(tsdn);
+	}
+}
+
+static void
+extent_gprof_sub(tsdn_t *tsdn, const extent_t *extent)
+{
+
+	cassert(config_prof);
+
+	if (opt_prof && extent_active_get(extent)) {
+		size_t nsub = extent_size_get(extent) >> LG_PAGE;
+		assert(atomic_read_zu(&curpages) >= nsub);
+		atomic_sub_zu(&curpages, nsub);
+	}
+}
+
 static bool
 extent_register(tsdn_t *tsdn, const extent_t *extent)
 {
@@ -280,20 +315,8 @@ extent_register(tsdn_t *tsdn, const extent_t *extent)
 		extent_interior_register(tsdn, rtree_ctx, extent);
 	extent_rtree_release(tsdn, elm_a, elm_b);
 
-	if (config_prof && opt_prof && extent_active_get(extent)) {
+	if (config_prof)
-		size_t nadd = extent_size_get(extent) >> LG_PAGE;
+		extent_gprof_add(tsdn, extent);
-		size_t cur = atomic_add_zu(&curpages, nadd);
-		size_t high = atomic_read_zu(&highpages);
-		while (cur > high && atomic_cas_zu(&highpages, high, cur)) {
-			/*
-			 * Don't refresh cur, because it may have decreased
-			 * since this thread lost the highpages update race.
-			 */
-			high = atomic_read_zu(&highpages);
-		}
-		if (cur > high && prof_gdump_get_unlocked())
-			prof_gdump(tsdn);
-	}
 
 	return (false);
 }
@@ -336,11 +359,8 @@ extent_deregister(tsdn_t *tsdn, extent_t *extent)
 	}
 	extent_rtree_release(tsdn, elm_a, elm_b);
 
-	if (config_prof && opt_prof && extent_active_get(extent)) {
+	if (config_prof)
-		size_t nsub = extent_size_get(extent) >> LG_PAGE;
+		extent_gprof_sub(tsdn, extent);
-		assert(atomic_read_zu(&curpages) >= nsub);
-		atomic_sub_zu(&curpages, nsub);
-	}
 }
 
 /*
@@ -507,15 +527,17 @@ extent_recycle(tsdn_t *tsdn, arena_t *arena, extent_hooks_t **r_extent_hooks,
 		extent_usize_set(extent, usize);
 	}
 
-	if (commit && !extent_committed_get(extent) &&
+	if (*commit && !extent_committed_get(extent)) {
-	    extent_commit_wrapper(tsdn, arena, r_extent_hooks, extent, 0,
+		if (extent_commit_wrapper(tsdn, arena, r_extent_hooks, extent,
-	    extent_size_get(extent))) {
+		    0, extent_size_get(extent))) {
 			if (!locked)
 				malloc_mutex_unlock(tsdn, &arena->extents_mtx);
-		extent_record(tsdn, arena, r_extent_hooks, extent_heaps, cache,
+			extent_record(tsdn, arena, r_extent_hooks, extent_heaps,
-		    extent);
+			    cache, extent);
 			return (NULL);
 		}
+		extent_zeroed_set(extent, true);
+	}
 
 	if (pad != 0)
 		extent_addr_randomize(tsdn, extent, alignment);
@@ -591,8 +613,6 @@ extent_alloc_cache_impl(tsdn_t *tsdn, arena_t *arena,
 	extent = extent_recycle(tsdn, arena, r_extent_hooks,
 	    arena->extents_cached, locked, true, new_addr, usize, pad,
 	    alignment, zero, commit, slab);
-	if (extent == NULL)
-		return (NULL);
 	return (extent);
 }
 
@@ -626,9 +646,6 @@ extent_alloc_default_impl(tsdn_t *tsdn, arena_t *arena, void *new_addr,
 
 	ret = extent_alloc_core(tsdn, arena, new_addr, size, alignment, zero,
 	    commit, arena->dss_prec);
-	if (ret == NULL)
-		return (NULL);
-
 	return (ret);
 }
 
@@ -653,6 +670,136 @@ extent_alloc_default(extent_hooks_t *extent_hooks, void *new_addr, size_t size,
 	    alignment, zero, commit));
 }
 
+static void
+extent_retain(tsdn_t *tsdn, arena_t *arena, extent_hooks_t **r_extent_hooks,
+    extent_t *extent)
+{
+
+	if (config_stats)
+		arena->stats.retained += extent_size_get(extent);
+	extent_record(tsdn, arena, r_extent_hooks, arena->extents_retained,
+	    false, extent);
+}
+
+/*
+ * If virtual memory is retained, create increasingly larger extents from which
+ * to split requested extents in order to limit the total number of disjoint
+ * virtual memory ranges retained by each arena.
+ */
+static extent_t *
+extent_grow_retained(tsdn_t *tsdn, arena_t *arena,
+    extent_hooks_t **r_extent_hooks, void *new_addr, size_t usize, size_t pad,
+    size_t alignment, bool *zero, bool *commit, bool slab)
+{
+	extent_t *extent;
+	void *ptr;
+	size_t size, alloc_size, alloc_size_min, leadsize, trailsize;
+	bool zeroed, committed;
+
+	/*
+	 * Check whether the next extent size in the series would be large
+	 * enough to satisfy this request.  If no, just bail, so that e.g. a
+	 * series of unsatisfiable allocation requests doesn't cause unused
+	 * extent creation as a side effect.
+	 */
+	size = usize + pad;
+	alloc_size = pind2sz(arena->extent_grow_next);
+	alloc_size_min = size + PAGE_CEILING(alignment) - PAGE;
+	/* Beware size_t wrap-around. */
+	if (alloc_size_min < usize)
+		return (NULL);
+	if (alloc_size < alloc_size_min)
+		return (NULL);
+	extent = extent_alloc(tsdn, arena);
+	if (extent == NULL)
+		return (NULL);
+	zeroed = false;
+	committed = false;
+	ptr = extent_alloc_core(tsdn, arena, new_addr, alloc_size, PAGE,
+	    &zeroed, &committed, arena->dss_prec);
+	extent_init(extent, arena, ptr, alloc_size, alloc_size,
+	    arena_extent_sn_next(arena), false, zeroed, committed, false);
+	if (ptr == NULL || extent_register(tsdn, extent)) {
+		extent_dalloc(tsdn, arena, extent);
+		return (NULL);
+	}
+	/*
+	 * Set the extent as active *after registration so that no gprof-related
+	 * accounting occurs during registration.
+	 */
+	extent_active_set(extent, true);
+
+	leadsize = ALIGNMENT_CEILING((uintptr_t)ptr, PAGE_CEILING(alignment)) -
+	    (uintptr_t)ptr;
+	assert(new_addr == NULL || leadsize == 0);
+	assert(alloc_size >= leadsize + size);
+	trailsize = alloc_size - leadsize - size;
+	if (extent_zeroed_get(extent))
+		*zero = true;
+	if (extent_committed_get(extent))
+		*commit = true;
+
+	/* Split the lead. */
+	if (leadsize != 0) {
+		extent_t *lead = extent;
+		extent = extent_split_wrapper(tsdn, arena, r_extent_hooks, lead,
+		    leadsize, leadsize, size + trailsize, usize + trailsize);
+		if (extent == NULL) {
+			extent_deregister(tsdn, lead);
+			extent_leak(tsdn, arena, r_extent_hooks, false, lead);
+			return (NULL);
+		}
+		extent_retain(tsdn, arena, r_extent_hooks, lead);
+	}
+
+	/* Split the trail. */
+	if (trailsize != 0) {
+		extent_t *trail = extent_split_wrapper(tsdn, arena,
+		    r_extent_hooks, extent, size, usize, trailsize, trailsize);
+		if (trail == NULL) {
+			extent_deregister(tsdn, extent);
+			extent_leak(tsdn, arena, r_extent_hooks, false, extent);
+			return (NULL);
+		}
+		extent_retain(tsdn, arena, r_extent_hooks, trail);
+	} else if (leadsize == 0) {
+		/*
+		 * Splitting causes usize to be set as a side effect, but no
+		 * splitting occurred.
+		 */
+		extent_usize_set(extent, usize);
+	}
+
+	if (*commit && !extent_committed_get(extent)) {
+		if (extent_commit_wrapper(tsdn, arena, r_extent_hooks, extent,
+		    0, extent_size_get(extent))) {
+			extent_retain(tsdn, arena, r_extent_hooks, extent);
+			return (NULL);
+		}
+		extent_zeroed_set(extent, true);
+	}
+
+	if (config_prof) {
+		/* Adjust gprof stats now that extent is final size. */
+		extent_gprof_add(tsdn, extent);
+	}
+	if (pad != 0)
+		extent_addr_randomize(tsdn, extent, alignment);
+	if (slab) {
+		rtree_ctx_t rtree_ctx_fallback;
+		rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn,
+		    &rtree_ctx_fallback);
+
+		extent_slab_set(extent, true);
+		extent_interior_register(tsdn, rtree_ctx, extent);
+	}
+	if (*zero && !extent_zeroed_get(extent))
+		memset(extent_addr_get(extent), 0, extent_usize_get(extent));
+	if (arena->extent_grow_next + 1 < NPSIZES)
+		arena->extent_grow_next++;
+	return (extent);
+}
+
 static extent_t *
 extent_alloc_retained(tsdn_t *tsdn, arena_t *arena,
     extent_hooks_t **r_extent_hooks, void *new_addr, size_t usize, size_t pad,
@@ -669,6 +816,12 @@ extent_alloc_retained(tsdn_t *tsdn, arena_t *arena,
 	if (extent != NULL && config_stats) {
 		size_t size = usize + pad;
 		arena->stats.retained -= size;
+		if (config_prof)
+			extent_gprof_add(tsdn, extent);
+	}
+	if (!config_munmap && extent == NULL) {
+		extent = extent_grow_retained(tsdn, arena, r_extent_hooks,
+		    new_addr, usize, pad, alignment, zero, commit, slab);
 	}
 
 	return (extent);
@@ -909,6 +1062,8 @@ extent_dalloc_wrapper(tsdn_t *tsdn, arena_t *arena,
 
 	if (config_stats)
 		arena->stats.retained += extent_size_get(extent);
+	if (config_prof)
+		extent_gprof_sub(tsdn, extent);
 
 	extent_record(tsdn, arena, r_extent_hooks, arena->extents_retained,
 	    false, extent);