hakmem/core/hakmem_shared_pool.c

#include "hakmem_shared_pool.h"
#include "hakmem_tiny_superslab.h"
#include "hakmem_tiny_superslab_constants.h"

#include <stdlib.h>
#include <string.h>
#include <stdatomic.h>
#include <stdio.h>

// ============================================================================
// P0 Lock Contention Instrumentation
// ============================================================================
static _Atomic uint64_t g_lock_acquire_count = 0;      // Total lock acquisitions
static _Atomic uint64_t g_lock_release_count = 0;      // Total lock releases
static _Atomic uint64_t g_lock_acquire_slab_count = 0; // Locks from acquire_slab path
static _Atomic uint64_t g_lock_release_slab_count = 0; // Locks from release_slab path
static int g_lock_stats_enabled = -1;                  // -1=uninitialized, 0=off, 1=on

// Initialize lock stats from environment variable
static inline void lock_stats_init(void) {
    if (__builtin_expect(g_lock_stats_enabled == -1, 0)) {
        const char* env = getenv("HAKMEM_SHARED_POOL_LOCK_STATS");
        g_lock_stats_enabled = (env && *env && *env != '0') ? 1 : 0;
    }
}

// Report lock statistics at shutdown
static void __attribute__((destructor)) lock_stats_report(void) {
    if (g_lock_stats_enabled != 1) {
        return;
    }

    uint64_t acquires = atomic_load(&g_lock_acquire_count);
    uint64_t releases = atomic_load(&g_lock_release_count);
    uint64_t acquire_path = atomic_load(&g_lock_acquire_slab_count);
    uint64_t release_path = atomic_load(&g_lock_release_slab_count);

    fprintf(stderr, "\n=== SHARED POOL LOCK STATISTICS ===\n");
    fprintf(stderr, "Total lock ops:    %lu (acquire) + %lu (release) = %lu\n",
            acquires, releases, acquires + releases);
    fprintf(stderr, "Balance:           %ld (should be 0)\n",
            (int64_t)acquires - (int64_t)releases);
    fprintf(stderr, "\n--- Breakdown by Code Path ---\n");
    fprintf(stderr, "acquire_slab():    %lu (%.1f%%)\n",
            acquire_path, 100.0 * acquire_path / (acquires ? acquires : 1));
    fprintf(stderr, "release_slab():    %lu (%.1f%%)\n",
            release_path, 100.0 * release_path / (acquires ? acquires : 1));
    fprintf(stderr, "===================================\n");
}

// ============================================================================
// P0-4: Lock-Free Free Slot List - Node Pool
// ============================================================================

// Pre-allocated node pools (one per class, to avoid malloc/free)
FreeSlotNode g_free_node_pool[TINY_NUM_CLASSES_SS][MAX_FREE_NODES_PER_CLASS];
_Atomic uint32_t g_node_alloc_index[TINY_NUM_CLASSES_SS] = {0};

// Allocate a node from pool (lock-free fast path, may fall back to legacy path)
static inline FreeSlotNode* node_alloc(int class_idx) {
    if (class_idx < 0 || class_idx >= TINY_NUM_CLASSES_SS) {
        return NULL;
    }

    uint32_t idx = atomic_fetch_add(&g_node_alloc_index[class_idx], 1);
    if (idx >= MAX_FREE_NODES_PER_CLASS) {
        // Pool exhausted - should be rare. Caller must fall back to legacy
        // mutex-protected free list to preserve correctness.
        static _Atomic int warn_once = 0;
        if (atomic_exchange(&warn_once, 1) == 0) {
            fprintf(stderr, "[P0-4 WARN] Node pool exhausted for class %d\n", class_idx);
        }
        return NULL;
    }

    return &g_free_node_pool[class_idx][idx];
}

// ============================================================================
// Phase 12-2: SharedSuperSlabPool skeleton implementation
// Goal:
//   - Centralize SuperSlab allocation/registration
//   - Provide acquire_slab/release_slab APIs for later refill/free integration
//   - Keep logic simple & conservative; correctness and observability first.
//
// Notes:
//   - Concurrency: protected by g_shared_pool.alloc_lock for now.
//   - class_hints is best-effort: read lock-free, written under lock.
//   - LRU hooks left as no-op placeholders.

SharedSuperSlabPool g_shared_pool = {
    .slabs        = NULL,
    .capacity     = 0,
    .total_count  = 0,
    .active_count = 0,
    .alloc_lock   = PTHREAD_MUTEX_INITIALIZER,
    .class_hints  = { NULL },
    .lru_head     = NULL,
    .lru_tail     = NULL,
    .lru_count    = 0,
    // P0-4: Lock-free free slot lists (zero-initialized atomic pointers)
    .free_slots_lockfree = {{.head = ATOMIC_VAR_INIT(NULL)}},
    // Legacy: mutex-protected free lists
    .free_slots   = {{.entries = {{0}}, .count = 0}},
    // Phase 12: SP-SLOT fields (ss_metadata is fixed-size array, auto-zeroed)
    .ss_meta_count = 0
};

static void
shared_pool_ensure_capacity_unlocked(uint32_t min_capacity)
{
    if (g_shared_pool.capacity >= min_capacity) {
        return;
    }

    uint32_t new_cap = g_shared_pool.capacity ? g_shared_pool.capacity : 16;
    while (new_cap < min_capacity) {
        new_cap *= 2;
    }

    SuperSlab** new_slabs = (SuperSlab**)realloc(g_shared_pool.slabs,
                                                 new_cap * sizeof(SuperSlab*));
    if (!new_slabs) {
        // Allocation failure: keep old state; caller must handle NULL later.
        return;
    }

    // Zero new entries to keep scanning logic simple.
    memset(new_slabs + g_shared_pool.capacity, 0,
           (new_cap - g_shared_pool.capacity) * sizeof(SuperSlab*));

    g_shared_pool.slabs    = new_slabs;
    g_shared_pool.capacity = new_cap;
}

void
shared_pool_init(void)
{
    // Idempotent init; safe to call from multiple early paths.
    // pthread_mutex_t with static initializer is already valid.
    pthread_mutex_lock(&g_shared_pool.alloc_lock);
    if (g_shared_pool.capacity == 0 && g_shared_pool.slabs == NULL) {
        shared_pool_ensure_capacity_unlocked(16);
    }
    pthread_mutex_unlock(&g_shared_pool.alloc_lock);
}

// ============================================================================
// Phase 12: SP-SLOT Box - Modular Helper Functions
// ============================================================================

// ---------- Layer 1: Slot Operations (Low-level) ----------

// Find first unused slot in SharedSSMeta
// P0-5: Uses atomic load for state check
// Returns: slot_idx on success, -1 if no unused slots
static int sp_slot_find_unused(SharedSSMeta* meta) {
    if (!meta) return -1;

    for (int i = 0; i < meta->total_slots; i++) {
        SlotState state = atomic_load_explicit(&meta->slots[i].state, memory_order_acquire);
        if (state == SLOT_UNUSED) {
            return i;
        }
    }
    return -1;
}

// Mark slot as ACTIVE (UNUSED→ACTIVE or EMPTY→ACTIVE)
// P0-5: Uses atomic store for state transition (caller must hold mutex!)
// Returns: 0 on success, -1 on error
static int sp_slot_mark_active(SharedSSMeta* meta, int slot_idx, int class_idx) {
    if (!meta || slot_idx < 0 || slot_idx >= meta->total_slots) return -1;
    if (class_idx < 0 || class_idx >= TINY_NUM_CLASSES_SS) return -1;

    SharedSlot* slot = &meta->slots[slot_idx];

    // Load state atomically
    SlotState state = atomic_load_explicit(&slot->state, memory_order_acquire);

    // Transition: UNUSED→ACTIVE or EMPTY→ACTIVE
    if (state == SLOT_UNUSED || state == SLOT_EMPTY) {
        atomic_store_explicit(&slot->state, SLOT_ACTIVE, memory_order_release);
        slot->class_idx = (uint8_t)class_idx;
        slot->slab_idx = (uint8_t)slot_idx;
        meta->active_slots++;
        return 0;
    }

    return -1;  // Already ACTIVE or invalid state
}

// Mark slot as EMPTY (ACTIVE→EMPTY)
// P0-5: Uses atomic store for state transition (caller must hold mutex!)
// Returns: 0 on success, -1 on error
static int sp_slot_mark_empty(SharedSSMeta* meta, int slot_idx) {
    if (!meta || slot_idx < 0 || slot_idx >= meta->total_slots) return -1;

    SharedSlot* slot = &meta->slots[slot_idx];

    // Load state atomically
    SlotState state = atomic_load_explicit(&slot->state, memory_order_acquire);

    if (state == SLOT_ACTIVE) {
        atomic_store_explicit(&slot->state, SLOT_EMPTY, memory_order_release);
        if (meta->active_slots > 0) {
            meta->active_slots--;
        }
        return 0;
    }

    return -1;  // Not ACTIVE
}

// ---------- Layer 2: Metadata Management (Mid-level) ----------

// Ensure ss_metadata array has capacity for at least min_count entries
// Caller must hold alloc_lock
// Returns: 0 on success, -1 if capacity exceeded
// RACE FIX: No realloc! Fixed-size array prevents race with lock-free Stage 2
static int sp_meta_ensure_capacity(uint32_t min_count) {
    if (min_count > MAX_SS_METADATA_ENTRIES) {
        static int warn_once = 0;
        if (warn_once == 0) {
            fprintf(stderr, "[SP_META_CAPACITY_ERROR] Exceeded MAX_SS_METADATA_ENTRIES=%d\n",
                    MAX_SS_METADATA_ENTRIES);
            warn_once = 1;
        }
        return -1;
    }
    return 0;
}

// Find SharedSSMeta for given SuperSlab, or create if not exists
// Caller must hold alloc_lock
// Returns: SharedSSMeta* on success, NULL on error
static SharedSSMeta* sp_meta_find_or_create(SuperSlab* ss) {
    if (!ss) return NULL;

    // RACE FIX: Load count atomically for consistency (even under mutex)
    uint32_t count = atomic_load_explicit(&g_shared_pool.ss_meta_count, memory_order_relaxed);

    // Search existing metadata
    for (uint32_t i = 0; i < count; i++) {
        // RACE FIX: Load pointer atomically for consistency
        SuperSlab* meta_ss = atomic_load_explicit(&g_shared_pool.ss_metadata[i].ss, memory_order_relaxed);
        if (meta_ss == ss) {
            return &g_shared_pool.ss_metadata[i];
        }
    }

    // Create new metadata entry
    if (sp_meta_ensure_capacity(count + 1) != 0) {
        return NULL;
    }

    // RACE FIX: Read current count atomically (even under mutex for consistency)
    uint32_t current_count = atomic_load_explicit(&g_shared_pool.ss_meta_count, memory_order_relaxed);
    SharedSSMeta* meta = &g_shared_pool.ss_metadata[current_count];

    // RACE FIX: Store SuperSlab pointer atomically (visible to lock-free Stage 2)
    atomic_store_explicit(&meta->ss, ss, memory_order_relaxed);
    meta->total_slots = (uint8_t)ss_slabs_capacity(ss);
    meta->active_slots = 0;

    // Initialize all slots as UNUSED
    // P0-5: Use atomic store for state initialization
    for (int i = 0; i < meta->total_slots; i++) {
        atomic_store_explicit(&meta->slots[i].state, SLOT_UNUSED, memory_order_relaxed);
        meta->slots[i].class_idx = 0;
        meta->slots[i].slab_idx = (uint8_t)i;
    }

    // RACE FIX: Atomic increment with release semantics
    // This ensures all writes to metadata[current_count] (lines 268-278) are visible
    // before the count increment is visible to lock-free Stage 2 readers
    atomic_fetch_add_explicit(&g_shared_pool.ss_meta_count, 1, memory_order_release);
    return meta;
}

// ---------- Layer 3: Free List Management ----------

// Push empty slot to per-class free list
// Caller must hold alloc_lock
// Returns: 0 on success, -1 if list is full
static int sp_freelist_push(int class_idx, SharedSSMeta* meta, int slot_idx) {
    if (class_idx < 0 || class_idx >= TINY_NUM_CLASSES_SS) return -1;
    if (!meta || slot_idx < 0 || slot_idx >= meta->total_slots) return -1;

    FreeSlotList* list = &g_shared_pool.free_slots[class_idx];

    if (list->count >= MAX_FREE_SLOTS_PER_CLASS) {
        return -1;  // List full
    }

    list->entries[list->count].meta = meta;
    list->entries[list->count].slot_idx = (uint8_t)slot_idx;
    list->count++;
    return 0;
}

// Pop empty slot from per-class free list
// Caller must hold alloc_lock
// Returns: 1 if popped (out params filled), 0 if list empty
static int sp_freelist_pop(int class_idx, SharedSSMeta** out_meta, int* out_slot_idx) {
    if (class_idx < 0 || class_idx >= TINY_NUM_CLASSES_SS) return 0;
    if (!out_meta || !out_slot_idx) return 0;

    FreeSlotList* list = &g_shared_pool.free_slots[class_idx];

    if (list->count == 0) {
        return 0;  // List empty
    }

    // Pop from end (LIFO for cache locality)
    list->count--;
    *out_meta = list->entries[list->count].meta;
    *out_slot_idx = list->entries[list->count].slot_idx;
    return 1;
}

// ============================================================================
// P0-5: Lock-Free Slot Claiming (Stage 2 Optimization)
// ============================================================================

// Try to claim an UNUSED slot via lock-free CAS
// Returns: slot_idx on success, -1 if no UNUSED slots available
// LOCK-FREE: Can be called from any thread without mutex
static int sp_slot_claim_lockfree(SharedSSMeta* meta, int class_idx) {
    if (!meta) return -1;
    if (class_idx < 0 || class_idx >= TINY_NUM_CLASSES_SS) return -1;

    // Scan all slots for UNUSED state
    for (int i = 0; i < meta->total_slots; i++) {
        SlotState expected = SLOT_UNUSED;

        // Try to claim this slot atomically (UNUSED → ACTIVE)
        if (atomic_compare_exchange_strong_explicit(
            &meta->slots[i].state,
            &expected,
            SLOT_ACTIVE,
            memory_order_acq_rel,  // Success: acquire+release semantics
            memory_order_relaxed   // Failure: just retry next slot
        )) {
            // Successfully claimed! Update non-atomic fields
            // (Safe because we now own this slot)
            meta->slots[i].class_idx = (uint8_t)class_idx;
            meta->slots[i].slab_idx = (uint8_t)i;

            // Increment active_slots counter atomically
            // (Multiple threads may claim slots concurrently)
            atomic_fetch_add_explicit(
                (_Atomic uint8_t*)&meta->active_slots, 1,
                memory_order_relaxed
            );

            return i;  // Return claimed slot index
        }

        // CAS failed (slot was not UNUSED) - continue to next slot
    }

    return -1;  // No UNUSED slots available
}

// ============================================================================
// P0-4: Lock-Free Free Slot List Operations
// ============================================================================

// Push empty slot to lock-free per-class free list (LIFO)
// LOCK-FREE: Can be called from any thread without mutex
// Returns: 0 on success, -1 on failure (node pool exhausted)
static int sp_freelist_push_lockfree(int class_idx, SharedSSMeta* meta, int slot_idx) {
    if (class_idx < 0 || class_idx >= TINY_NUM_CLASSES_SS) return -1;
    if (!meta || slot_idx < 0 || slot_idx >= meta->total_slots) return -1;

    // Allocate node from pool
    FreeSlotNode* node = node_alloc(class_idx);
    if (!node) {
        // Fallback: push into legacy per-class free list
        // ASSUME: Caller already holds alloc_lock (e.g., shared_pool_release_slab:772)
        // Do NOT lock again to avoid deadlock on non-recursive mutex!
        (void)sp_freelist_push(class_idx, meta, slot_idx);
        return 0;
    }

    // Fill node data
    node->meta = meta;
    node->slot_idx = (uint8_t)slot_idx;

    // Lock-free LIFO push using CAS loop
    LockFreeFreeList* list = &g_shared_pool.free_slots_lockfree[class_idx];
    FreeSlotNode* old_head = atomic_load_explicit(&list->head, memory_order_relaxed);

    do {
        node->next = old_head;
    } while (!atomic_compare_exchange_weak_explicit(
        &list->head, &old_head, node,
        memory_order_release,   // Success: publish node to other threads
        memory_order_relaxed    // Failure: retry with updated old_head
    ));

    return 0;  // Success
}

// Pop empty slot from lock-free per-class free list (LIFO)
// LOCK-FREE: Can be called from any thread without mutex
// Returns: 1 if popped (out params filled), 0 if list empty
static int sp_freelist_pop_lockfree(int class_idx, SharedSSMeta** out_meta, int* out_slot_idx) {
    if (class_idx < 0 || class_idx >= TINY_NUM_CLASSES_SS) return 0;
    if (!out_meta || !out_slot_idx) return 0;

    LockFreeFreeList* list = &g_shared_pool.free_slots_lockfree[class_idx];
    FreeSlotNode* old_head = atomic_load_explicit(&list->head, memory_order_acquire);

    // Lock-free LIFO pop using CAS loop
    do {
        if (old_head == NULL) {
            return 0;  // List empty
        }
    } while (!atomic_compare_exchange_weak_explicit(
        &list->head, &old_head, old_head->next,
        memory_order_acquire,   // Success: acquire node data
        memory_order_acquire    // Failure: retry with updated old_head
    ));

    // Extract data from popped node
    *out_meta = old_head->meta;
    *out_slot_idx = old_head->slot_idx;

    // NOTE: We do NOT free the node back to pool (no node recycling yet)
    // This is acceptable because MAX_FREE_NODES_PER_CLASS (512) is generous
    // and workloads typically don't push/pop the same slot repeatedly

    return 1;  // Success
}

/*
 * Internal: allocate and register a new SuperSlab for the shared pool.
 *
 * Phase 12 NOTE:
 *  - We MUST use the real superslab_allocate() path so that:
 *      - backing memory is a full SuperSlab region (1–2MB),
 *      - header/layout are initialized correctly,
 *      - registry integration stays consistent.
 *  - shared_pool is responsible only for:
 *      - tracking pointers,
 *      - marking per-slab class_idx as UNASSIGNED initially.
 *    It does NOT bypass registry/LRU.
 *
 * Caller must hold alloc_lock.
 */
static SuperSlab*
shared_pool_allocate_superslab_unlocked(void)
{
    // Use size_class 0 as a neutral hint; Phase 12 per-slab class_idx is authoritative.
    extern SuperSlab* superslab_allocate(uint8_t size_class);
    SuperSlab* ss = superslab_allocate(0);
    if (!ss) {
        return NULL;
    }

    // superslab_allocate() already:
    //  - zeroes slab metadata / remote queues,
    //  - sets magic/lg_size/etc,
    //  - registers in global registry.
    // For shared-pool semantics we normalize all slab class_idx to UNASSIGNED.
    int max_slabs = ss_slabs_capacity(ss);
    for (int i = 0; i < max_slabs; i++) {
        ss->slabs[i].class_idx = 255; // UNASSIGNED
    }

    if (g_shared_pool.total_count >= g_shared_pool.capacity) {
        shared_pool_ensure_capacity_unlocked(g_shared_pool.total_count + 1);
        if (g_shared_pool.total_count >= g_shared_pool.capacity) {
            // Pool table expansion failed; leave ss alive (registry-owned),
            // but do not treat it as part of shared_pool.
            return NULL;
        }
    }

    g_shared_pool.slabs[g_shared_pool.total_count] = ss;
    g_shared_pool.total_count++;
    // Not counted as active until at least one slab is assigned.
    return ss;
}

SuperSlab*
shared_pool_acquire_superslab(void)
{
    // Phase 12 debug safety:
    //   If shared backend is disabled at Box API level, this function SHOULD NOT be called.
    //   But since bench currently SEGVs here even with legacy forced, treat this as a hard guard:
    //   we early-return error instead of touching potentially-bad state.
    //
    //   This isolates shared_pool from the current crash so we can validate legacy path first.
    // FIXED: Remove the return -1; that was preventing operation

    shared_pool_init();

    pthread_mutex_lock(&g_shared_pool.alloc_lock);

    // For now, always allocate a fresh SuperSlab and register it.
    // More advanced reuse/GC comes later.
    SuperSlab* ss = shared_pool_allocate_superslab_unlocked();

    pthread_mutex_unlock(&g_shared_pool.alloc_lock);
    return ss;
}

// ---------- Layer 4: Public API (High-level) ----------

int
shared_pool_acquire_slab(int class_idx, SuperSlab** ss_out, int* slab_idx_out)
{
    // Phase 12: SP-SLOT Box - 3-Stage Acquire Logic
    //
    // Stage 1: Reuse EMPTY slots from per-class free list (EMPTY→ACTIVE)
    // Stage 2: Find UNUSED slots in existing SuperSlabs
    // Stage 3: Get new SuperSlab (LRU pop or mmap)
    //
    // Invariants:
    //  - On success: *ss_out != NULL, 0 <= *slab_idx_out < total_slots
    //  - The chosen slab has meta->class_idx == class_idx

    if (!ss_out || !slab_idx_out) {
        return -1;
    }
    if (class_idx < 0 || class_idx >= TINY_NUM_CLASSES_SS) {
        return -1;
    }

    shared_pool_init();

    // Debug logging
    static int dbg_acquire = -1;
    if (__builtin_expect(dbg_acquire == -1, 0)) {
        const char* e = getenv("HAKMEM_SS_ACQUIRE_DEBUG");
        dbg_acquire = (e && *e && *e != '0') ? 1 : 0;
    }

    // ========== Stage 1 (Lock-Free): Try to reuse EMPTY slots ==========
    // P0-4: Lock-free pop from per-class free list (no mutex needed!)
    // Best case: Same class freed a slot, reuse immediately (cache-hot)
    SharedSSMeta* reuse_meta = NULL;
    int reuse_slot_idx = -1;

    if (sp_freelist_pop_lockfree(class_idx, &reuse_meta, &reuse_slot_idx)) {
        // Found EMPTY slot from lock-free list!
        // Now acquire mutex ONLY for slot activation and metadata update

        // P0 instrumentation: count lock acquisitions
        lock_stats_init();
        if (g_lock_stats_enabled == 1) {
            atomic_fetch_add(&g_lock_acquire_count, 1);
            atomic_fetch_add(&g_lock_acquire_slab_count, 1);
        }

        pthread_mutex_lock(&g_shared_pool.alloc_lock);

        // Activate slot under mutex (slot state transition requires protection)
        if (sp_slot_mark_active(reuse_meta, reuse_slot_idx, class_idx) == 0) {
            // RACE FIX: Load SuperSlab pointer atomically (consistency)
            SuperSlab* ss = atomic_load_explicit(&reuse_meta->ss, memory_order_relaxed);

            if (dbg_acquire == 1) {
                fprintf(stderr, "[SP_ACQUIRE_STAGE1_LOCKFREE] class=%d reusing EMPTY slot (ss=%p slab=%d)\n",
                        class_idx, (void*)ss, reuse_slot_idx);
            }

            // Update SuperSlab metadata
            ss->slab_bitmap |= (1u << reuse_slot_idx);
            ss->slabs[reuse_slot_idx].class_idx = (uint8_t)class_idx;

            if (ss->active_slabs == 0) {
                // Was empty, now active again
                ss->active_slabs = 1;
                g_shared_pool.active_count++;
            }

            // Update hint
            g_shared_pool.class_hints[class_idx] = ss;

            *ss_out = ss;
            *slab_idx_out = reuse_slot_idx;

            if (g_lock_stats_enabled == 1) {
                atomic_fetch_add(&g_lock_release_count, 1);
            }
            pthread_mutex_unlock(&g_shared_pool.alloc_lock);
            return 0;  // ✅ Stage 1 (lock-free) success
        }

        // Slot activation failed (race condition?) - release lock and fall through
        if (g_lock_stats_enabled == 1) {
            atomic_fetch_add(&g_lock_release_count, 1);
        }
        pthread_mutex_unlock(&g_shared_pool.alloc_lock);
    }

    // ========== Stage 2 (Lock-Free): Try to claim UNUSED slots ==========
    // P0-5: Lock-free atomic CAS claiming (no mutex needed for slot state transition!)
    // RACE FIX: Read ss_meta_count atomically (now properly declared as _Atomic)
    // No cast needed! memory_order_acquire synchronizes with release in sp_meta_find_or_create
    uint32_t meta_count = atomic_load_explicit(
        &g_shared_pool.ss_meta_count,
        memory_order_acquire
    );

    for (uint32_t i = 0; i < meta_count; i++) {
        SharedSSMeta* meta = &g_shared_pool.ss_metadata[i];

        // Try lock-free claiming (UNUSED → ACTIVE via CAS)
        int claimed_idx = sp_slot_claim_lockfree(meta, class_idx);
        if (claimed_idx >= 0) {
            // RACE FIX: Load SuperSlab pointer atomically (critical for lock-free Stage 2)
            // Use memory_order_acquire to synchronize with release in sp_meta_find_or_create
            SuperSlab* ss = atomic_load_explicit(&meta->ss, memory_order_acquire);
            if (!ss) {
                // SuperSlab was freed between claiming and loading - skip this entry
                continue;
            }

            if (dbg_acquire == 1) {
                fprintf(stderr, "[SP_ACQUIRE_STAGE2_LOCKFREE] class=%d claimed UNUSED slot (ss=%p slab=%d)\n",
                        class_idx, (void*)ss, claimed_idx);
            }

            // P0 instrumentation: count lock acquisitions
            lock_stats_init();
            if (g_lock_stats_enabled == 1) {
                atomic_fetch_add(&g_lock_acquire_count, 1);
                atomic_fetch_add(&g_lock_acquire_slab_count, 1);
            }

            pthread_mutex_lock(&g_shared_pool.alloc_lock);

            // Update SuperSlab metadata under mutex
            ss->slab_bitmap |= (1u << claimed_idx);
            ss->slabs[claimed_idx].class_idx = (uint8_t)class_idx;

            if (ss->active_slabs == 0) {
                ss->active_slabs = 1;
                g_shared_pool.active_count++;
            }

            // Update hint
            g_shared_pool.class_hints[class_idx] = ss;

            *ss_out = ss;
            *slab_idx_out = claimed_idx;

            if (g_lock_stats_enabled == 1) {
                atomic_fetch_add(&g_lock_release_count, 1);
            }
            pthread_mutex_unlock(&g_shared_pool.alloc_lock);
            return 0;  // ✅ Stage 2 (lock-free) success
        }

        // Claim failed (no UNUSED slots in this meta) - continue to next SuperSlab
    }

    // ========== Stage 3: Mutex-protected fallback (new SuperSlab allocation) ==========
    // All existing SuperSlabs have no UNUSED slots → need new SuperSlab
    // P0 instrumentation: count lock acquisitions
    lock_stats_init();
    if (g_lock_stats_enabled == 1) {
        atomic_fetch_add(&g_lock_acquire_count, 1);
        atomic_fetch_add(&g_lock_acquire_slab_count, 1);
    }

    pthread_mutex_lock(&g_shared_pool.alloc_lock);

    // ========== Stage 3: Get new SuperSlab ==========
    // Try LRU cache first, then mmap
    SuperSlab* new_ss = NULL;

    // Stage 3a: Try LRU cache
    extern SuperSlab* hak_ss_lru_pop(uint8_t size_class);
    new_ss = hak_ss_lru_pop((uint8_t)class_idx);

    int from_lru = (new_ss != NULL);

    // Stage 3b: If LRU miss, allocate new SuperSlab
    if (!new_ss) {
        new_ss = shared_pool_allocate_superslab_unlocked();
    }

    if (dbg_acquire == 1 && new_ss) {
        fprintf(stderr, "[SP_ACQUIRE_STAGE3] class=%d new SuperSlab (ss=%p from_lru=%d)\n",
                class_idx, (void*)new_ss, from_lru);
    }

    if (!new_ss) {
        if (g_lock_stats_enabled == 1) {
            atomic_fetch_add(&g_lock_release_count, 1);
        }
        pthread_mutex_unlock(&g_shared_pool.alloc_lock);
        return -1;  // ❌ Out of memory
    }

    // Create metadata for this new SuperSlab
    SharedSSMeta* new_meta = sp_meta_find_or_create(new_ss);
    if (!new_meta) {
        if (g_lock_stats_enabled == 1) {
            atomic_fetch_add(&g_lock_release_count, 1);
        }
        pthread_mutex_unlock(&g_shared_pool.alloc_lock);
        return -1;  // ❌ Metadata allocation failed
    }

    // Assign first slot to this class
    int first_slot = 0;
    if (sp_slot_mark_active(new_meta, first_slot, class_idx) != 0) {
        if (g_lock_stats_enabled == 1) {
            atomic_fetch_add(&g_lock_release_count, 1);
        }
        pthread_mutex_unlock(&g_shared_pool.alloc_lock);
        return -1;  // ❌ Should not happen
    }

    // Update SuperSlab metadata
    new_ss->slab_bitmap |= (1u << first_slot);
    new_ss->slabs[first_slot].class_idx = (uint8_t)class_idx;
    new_ss->active_slabs = 1;
    g_shared_pool.active_count++;

    // Update hint
    g_shared_pool.class_hints[class_idx] = new_ss;

    *ss_out = new_ss;
    *slab_idx_out = first_slot;

    if (g_lock_stats_enabled == 1) {
        atomic_fetch_add(&g_lock_release_count, 1);
    }
    pthread_mutex_unlock(&g_shared_pool.alloc_lock);
    return 0;  // ✅ Stage 3 success
}

void
shared_pool_release_slab(SuperSlab* ss, int slab_idx)
{
    // Phase 12: SP-SLOT Box - Slot-based Release
    //
    // Flow:
    //   1. Validate inputs and check meta->used == 0
    //   2. Find SharedSSMeta for this SuperSlab
    //   3. Mark slot ACTIVE → EMPTY
    //   4. Push to per-class free list (enables same-class reuse)
    //   5. If all slots EMPTY → superslab_free() → LRU cache

    if (!ss) {
        return;
    }
    if (slab_idx < 0 || slab_idx >= SLABS_PER_SUPERSLAB_MAX) {
        return;
    }

    // Debug logging
    static int dbg = -1;
    if (__builtin_expect(dbg == -1, 0)) {
        const char* e = getenv("HAKMEM_SS_FREE_DEBUG");
        dbg = (e && *e && *e != '0') ? 1 : 0;
    }

    // P0 instrumentation: count lock acquisitions
    lock_stats_init();
    if (g_lock_stats_enabled == 1) {
        atomic_fetch_add(&g_lock_acquire_count, 1);
        atomic_fetch_add(&g_lock_release_slab_count, 1);
    }

    pthread_mutex_lock(&g_shared_pool.alloc_lock);

    TinySlabMeta* slab_meta = &ss->slabs[slab_idx];
    if (slab_meta->used != 0) {
        // Not actually empty; nothing to do
        if (g_lock_stats_enabled == 1) {
            atomic_fetch_add(&g_lock_release_count, 1);
        }
        pthread_mutex_unlock(&g_shared_pool.alloc_lock);
        return;
    }

    uint8_t class_idx = slab_meta->class_idx;

    if (dbg == 1) {
        fprintf(stderr, "[SP_SLOT_RELEASE] ss=%p slab_idx=%d class=%d used=0 (marking EMPTY)\n",
                (void*)ss, slab_idx, class_idx);
    }

    // Find SharedSSMeta for this SuperSlab
    SharedSSMeta* sp_meta = NULL;
    uint32_t count = atomic_load_explicit(&g_shared_pool.ss_meta_count, memory_order_relaxed);
    for (uint32_t i = 0; i < count; i++) {
        // RACE FIX: Load pointer atomically
        SuperSlab* meta_ss = atomic_load_explicit(&g_shared_pool.ss_metadata[i].ss, memory_order_relaxed);
        if (meta_ss == ss) {
            sp_meta = &g_shared_pool.ss_metadata[i];
            break;
        }
    }

    if (!sp_meta) {
        // SuperSlab not in SP-SLOT system yet - create metadata
        sp_meta = sp_meta_find_or_create(ss);
        if (!sp_meta) {
            pthread_mutex_unlock(&g_shared_pool.alloc_lock);
            return;  // Failed to create metadata
        }
    }

    // Mark slot as EMPTY (ACTIVE → EMPTY)
    if (sp_slot_mark_empty(sp_meta, slab_idx) != 0) {
        if (g_lock_stats_enabled == 1) {
            atomic_fetch_add(&g_lock_release_count, 1);
        }
        pthread_mutex_unlock(&g_shared_pool.alloc_lock);
        return;  // Slot wasn't ACTIVE
    }

    // Update SuperSlab metadata
    uint32_t bit = (1u << slab_idx);
    if (ss->slab_bitmap & bit) {
        ss->slab_bitmap &= ~bit;
        slab_meta->class_idx = 255;  // UNASSIGNED

        if (ss->active_slabs > 0) {
            ss->active_slabs--;
            if (ss->active_slabs == 0 && g_shared_pool.active_count > 0) {
                g_shared_pool.active_count--;
            }
        }
    }

    // P0-4: Push to lock-free per-class free list (enables reuse by same class)
    // Note: push BEFORE releasing mutex (slot state already updated under lock)
    if (class_idx < TINY_NUM_CLASSES_SS) {
        sp_freelist_push_lockfree(class_idx, sp_meta, slab_idx);

        if (dbg == 1) {
            fprintf(stderr, "[SP_SLOT_FREELIST_LOCKFREE] class=%d pushed slot (ss=%p slab=%d) active_slots=%u/%u\n",
                    class_idx, (void*)ss, slab_idx,
                    sp_meta->active_slots, sp_meta->total_slots);
        }
    }

    // Check if SuperSlab is now completely empty (all slots EMPTY or UNUSED)
    if (sp_meta->active_slots == 0) {
        if (dbg == 1) {
            fprintf(stderr, "[SP_SLOT_COMPLETELY_EMPTY] ss=%p active_slots=0 (calling superslab_free)\n",
                    (void*)ss);
        }

        if (g_lock_stats_enabled == 1) {
            atomic_fetch_add(&g_lock_release_count, 1);
        }

        // RACE FIX: Set meta->ss to NULL BEFORE unlocking mutex
        // This prevents Stage 2 from accessing freed SuperSlab
        atomic_store_explicit(&sp_meta->ss, NULL, memory_order_release);

        pthread_mutex_unlock(&g_shared_pool.alloc_lock);

        // Free SuperSlab:
        // 1. Try LRU cache (hak_ss_lru_push) - lazy deallocation
        // 2. Or munmap if LRU is full - eager deallocation
        extern void superslab_free(SuperSlab* ss);
        superslab_free(ss);
        return;
    }

    if (g_lock_stats_enabled == 1) {
        atomic_fetch_add(&g_lock_release_count, 1);
    }
    pthread_mutex_unlock(&g_shared_pool.alloc_lock);
}