Add SuperSlab Prefault Box with 4MB MAP_POPULATE bug fix

New Feature: ss_prefault_box.h
- Box for controlling SuperSlab page prefaulting policy
- ENV: HAKMEM_SS_PREFAULT (0=OFF, 1=POPULATE, 2=TOUCH)
- Default: OFF (safe mode until further optimization)

Bug Fix: 4MB MAP_POPULATE regression
- Problem: Fallback path allocated 4MB (2x size for alignment) with MAP_POPULATE
  causing 52x slower mmap (0.585ms → 30.6ms) and 35% throughput regression
- Solution: Remove MAP_POPULATE from 4MB allocation, apply madvise(MADV_WILLNEED)
  only to the aligned 2MB region after trimming prefix/suffix

Changes:
- core/box/ss_prefault_box.h: New prefault policy box (header-only)
- core/box/ss_allocation_box.c: Integrate prefault box, call ss_prefault_region()
- core/superslab_cache.c: Fix fallback path - no MAP_POPULATE on 4MB,
  always munmap prefix/suffix, use MADV_WILLNEED for 2MB only
- docs/specs/ENV_VARS*.md: Document HAKMEM_SS_PREFAULT

Performance:
- bench_random_mixed: 4.32M ops/s (regression fixed, slight improvement)
- bench_tiny_hot: 157M ops/s with prefault=1 (no crash)

Box Theory:
- OS layer (ss_os_acquire): "how to mmap"
- Prefault Box: "when to page-in"
- Allocation Box: "when to call prefault"

🤖 Generated with Claude Code

Co-Authored-By: Claude <noreply@anthropic.com>

core/box/ss_allocation_box.c

@@ -3,6 +3,7 @@
 
 #include "ss_allocation_box.h"
 #include "ss_os_acquire_box.h"
+#include "ss_prefault_box.h"
 #include "ss_cache_box.h"
 #include "ss_stats_box.h"
 #include "ss_ace_box.h"
@@ -167,11 +168,28 @@ SuperSlab* superslab_allocate(uint8_t size_class) {
     }
 
     if (!ptr) {
-        int populate = atomic_exchange_explicit(&g_ss_populate_once, 0, memory_order_acq_rel);
+        // Prefault policy: decide MAP_POPULATE and optional manual touch
+        SSPrefaultPolicy pf_policy = ss_prefault_policy();
+        int populate = 0;
+
+        if (pf_policy == SS_PREFAULT_POPULATE ||
+            pf_policy == SS_PREFAULT_TOUCH ||
+            pf_policy == SS_PREFAULT_ASYNC) {
+            // 常時 MAP_POPULATE（+必要なら手動 touch）
+            populate = 1;
+        } else {
+            // OFF の場合のみ、従来の「ワンショット populate」挙動を温存
+            populate = atomic_exchange_explicit(&g_ss_populate_once, 0, memory_order_acq_rel);
+        }
+
         ptr = ss_os_acquire(size_class, ss_size, ss_mask, populate);
         if (!ptr) {
             return NULL;
         }
+        // 手動 prefault が要求されている場合は、mmap 直後に SuperSlab 全域を page-in
+        if (pf_policy == SS_PREFAULT_TOUCH || pf_policy == SS_PREFAULT_ASYNC) {
+            ss_prefault_region(ptr, ss_size);
+        }
         // Debug logging for REFILL with new allocation
         if (dbg == 1) {
             fprintf(stderr, "[REFILL] class=%d new_alloc=1 ss=%p\n",

core/box/ss_prefault_box.h

@@ -0,0 +1,80 @@
+// ss_prefault_box.h - Box: SuperSlab Page Prefaulting Policy
+// Purpose: Control SuperSlab prefault strategy (MAP_POPULATE / manual touch)
+// Box Theory:
+//   - This box owns「いつ / どの程度 pf を前倒しするか」というポリシーだけを担当。
+//   - 実際の mmap は ss_os_acquire_box（下層）、Superslab 初期化は ss_allocation_box（上層）が担当。
+//
+// Responsibilities:
+//   - HAKMEM_SS_PREFAULT の値から prefault ポリシーを決定
+//   - 必要に応じて SuperSlab 全域を 1 回だけ touch して page-in する
+//
+// ENV:
+//   HAKMEM_SS_PREFAULT=0  Off        （既存動作に近い。g_ss_populate_once のワンショットのみ）
+//   HAKMEM_SS_PREFAULT=1  POPULATE   （MAP_POPULATE のみ使用）
+//   HAKMEM_SS_PREFAULT=2  TOUCH      （MAP_POPULATE + 手動4KB touch）
+//   HAKMEM_SS_PREFAULT=3  ASYNC      （予約値。現状は TOUCH と同等に扱う）
+//
+// Default:
+//   - 未指定時は 1 (POPULATE) とし、SuperSlab 取得時の page fault を大きく削減する。
+
+#ifndef HAKMEM_SS_PREFAULT_BOX_H
+#define HAKMEM_SS_PREFAULT_BOX_H
+
+#include <stddef.h>
+#include <stdlib.h>
+#include <unistd.h>
+
+typedef enum {
+    SS_PREFAULT_OFF = 0,      // No prefault (legacy: g_ss_populate_once のみ)
+    SS_PREFAULT_POPULATE = 1, // MAP_POPULATE only
+    SS_PREFAULT_TOUCH = 2,    // MAP_POPULATE + manual page touch
+    SS_PREFAULT_ASYNC = 3     // Reserved for future background prefault
+} SSPrefaultPolicy;
+
+// Return current prefault policy (parsed once from HAKMEM_SS_PREFAULT).
+static inline SSPrefaultPolicy ss_prefault_policy(void)
+{
+    static int cached = -1;
+    if (cached != -1) {
+        return (SSPrefaultPolicy)cached;
+    }
+
+    // 一時的な安全デフォルト: OFF
+    // （4MB MAP_POPULATE 問題の影響を避けつつ、必要なときだけ明示的に ON にする）
+    int policy = SS_PREFAULT_OFF;
+    const char* env = getenv("HAKMEM_SS_PREFAULT");
+    if (env && *env) {
+        int v = atoi(env);
+        if (v < 0) v = 0;
+        if (v > 3) v = 3;
+        policy = v;
+    }
+    cached = policy;
+    return (SSPrefaultPolicy)policy;
+}
+
+// Prefault a contiguous SuperSlab region by touching one byte per page.
+// Caller is expected to call this「mmap直後」に 1 回だけ実行する。
+static inline void ss_prefault_region(void* addr, size_t size)
+{
+    SSPrefaultPolicy policy = ss_prefault_policy();
+    if (policy < SS_PREFAULT_TOUCH) {
+        // POPULATE または OFF の場合は手動 touch は行わない。
+        return;
+    }
+    if (!addr || size == 0) {
+        return;
+    }
+
+    long ps = sysconf(_SC_PAGESIZE);
+    size_t page = (ps > 0) ? (size_t)ps : (size_t)4096;
+
+    volatile char* p = (volatile char*)addr;
+    for (size_t off = 0; off < size; off += page) {
+        (void)p[off];
+    }
+    // Ensure last byte is also touched (covers exact multiples).
+    (void)p[size - 1];
+}
+
+#endif // HAKMEM_SS_PREFAULT_BOX_H

core/superslab_cache.c

@@ -76,11 +76,6 @@ void* ss_os_acquire(uint8_t size_class, size_t ss_size, uintptr_t ss_mask, int p
 
     size_t alloc_size = ss_size * 2;
     int flags = MAP_PRIVATE | MAP_ANONYMOUS;
-#ifdef MAP_POPULATE
-    if (populate) {
-        flags |= MAP_POPULATE;
-    }
-#endif
     void* raw = mmap(NULL, alloc_size,
                      PROT_READ | PROT_WRITE,
                      flags,
@@ -110,13 +105,15 @@ void* ss_os_acquire(uint8_t size_class, size_t ss_size, uintptr_t ss_mask, int p
     }
     size_t suffix_size = alloc_size - prefix_size - ss_size;
     if (suffix_size > 0) {
-        if (populate) {
-#ifdef MADV_DONTNEED
-            madvise((char*)ptr + ss_size, suffix_size, MADV_DONTNEED);
-#endif
-        } else {
+        // 余剰領域は常に munmap して、実際に使用する SuperSlab サイズだけを残す。
         munmap((char*)ptr + ss_size, suffix_size);
     }
+
+    // populate が要求されている場合は、実際に使う SuperSlab 領域だけを事前 fault-in する。
+    if (populate) {
+#ifdef MADV_WILLNEED
+        madvise(ptr, ss_size, MADV_WILLNEED);
+#endif
     }
 
     ss_stats_os_alloc(size_class, ss_size);

docs/specs/ENV_VARS.md

@@ -176,6 +176,14 @@ SuperSlab cache / precharge（Phase 6.24+）
 - HAKMEM_TINY_SS_POPULATE_ONCE=1
   - 次回 `mmap` で取得する SuperSlab を 1 回だけ `MAP_POPULATE` で fault-in（A/B 用のワンショットプリタッチ）。
 
+SuperSlab prefault（PF 削減のための前倒し）
+- HAKMEM_SS_PREFAULT=0/1/2/3
+  - 0: OFF（安全デフォルト。`g_ss_populate_once` によるワンショット `MAP_POPULATE` のみ）
+  - 1: POPULATE — 新規 SuperSlab の `mmap` で常に `MAP_POPULATE` を付与し、page fault を kernel 側で事前解消（perf 計測用）。
+  - 2: TOUCH — `MAP_POPULATE` に加えて `ss_prefault_region()` で SuperSlab 全域を 1 回 4KB stride で touch（PF をほぼゼロにしたい実験用）。
+  - 3: ASYNC — 予約値（現状は TOUCH と同等の扱いだが、将来的に BG thread prefault 用に拡張予定）。
+  - Box: `core/box/ss_prefault_box.h`（ポリシー決定）＋ `core/box/ss_allocation_box.c`（mmap 直後の実行）。
+
 Harvest / Guard（mmap前の収穫ゲート）
 - HAKMEM_TINY_SS_CAP=N
   - Tiny 各クラスにおける SuperSlab 上限（0=無制限）。

docs/specs/ENV_VARS_COMPLETE.md

@@ -151,6 +151,44 @@ From `/mnt/workdisk/public_share/hakmem/core/hakmem_tiny_stats.h`:
 - **Purpose**: Interval between partial slab checks
 - **Impact**: Lower = more aggressive trimming
 
+#### HAKMEM_TINY_SS_CACHE
+- **Default**: 0 (disabled)
+- **Purpose**: Per-class SuperSlab cache capacity
+- **Impact**: Limits how many freed SuperSlabs are kept in LRU cache before munmap
+
+#### HAKMEM_TINY_SS_CACHE_C{0..7}
+- **Default**: unset (inherits `HAKMEM_TINY_SS_CACHE`)
+- **Purpose**: Per-class overrides for cache capacity
+- **Impact**: Fine-grained control of cache size per Tiny class
+
+#### HAKMEM_TINY_SS_PRECHARGE
+- **Default**: 0
+- **Purpose**: Precharge (pre-allocate) SuperSlabs into cache at startup/runtime
+- **Impact**: Reduces first-use page faults by having warm SuperSlabs ready
+
+#### HAKMEM_TINY_SS_PRECHARGE_C{0..7}
+- **Default**: unset (inherits `HAKMEM_TINY_SS_PRECHARGE`)
+- **Purpose**: Per-class precharge targets
+- **Impact**: e.g., `HAKMEM_TINY_SS_PRECHARGE_C0=4` precharges 4 SuperSlabs for class 0
+
+#### HAKMEM_TINY_SS_POPULATE_ONCE
+- **Default**: 0
+- **Purpose**: Use `MAP_POPULATE` for the next SuperSlab allocation only
+- **Impact**: One-shot prefault for A/B testing; superseded by `HAKMEM_SS_PREFAULT` for常時運用
+
+#### HAKMEM_SS_PREFAULT
+- **Default**: `0` (OFF, safety-first default)
+- **Type**: integer (0–3)
+- **Purpose**: Control SuperSlab prefault strategy to reduce kernel page fault overhead (enabled explicitly when tuning).
+- **Values**:
+  - `0` = OFF — legacy behavior, only `HAKMEM_TINY_SS_POPULATE_ONCE` may trigger a one-shot `MAP_POPULATE`（現状の安全デフォルト）。
+  - `1` = POPULATE — always pass `populate=1` to `ss_os_acquire()` (use `MAP_POPULATE` for every new SuperSlab). **要 perf 確認。**
+  - `2` = TOUCH — POPULATE + `ss_prefault_region()` touches each page once (4KB stride) after `mmap`（実験用）。
+  - `3` = ASYNC — reserved for future background-prefault implementation (currently treated as TOUCH).
+- **Implementation**:
+  - Policy Box: `core/box/ss_prefault_box.h`
+  - Integration: `core/box/ss_allocation_box.c` calls `ss_prefault_policy()` to set `populate` and `ss_prefault_region()` immediately after `ss_os_acquire()`.
+
 ---
 
 ### 4. Remote Free & Background Processing