Files
hakmem/core/box/pool_api.inc.h

304 lines
18 KiB
C
Raw Normal View History

CRITICAL FIX: TLS 未初期化による 4T SEGV を完全解消 **問題:** - Larson 4T で 100% SEGV (1T は 2.09M ops/s で完走) - System/mimalloc は 4T で 33.52M ops/s 正常動作 - SS OFF + Remote OFF でも 4T で SEGV **根本原因: (Task agent ultrathink 調査結果)** ``` CRASH: mov (%r15),%r13 R15 = 0x6261 ← ASCII "ba" (ゴミ値、未初期化TLS) ``` Worker スレッドの TLS 変数が未初期化: - `__thread void* g_tls_sll_head[TINY_NUM_CLASSES];` ← 初期化なし - pthread_create() で生成されたスレッドでゼロ初期化されない - NULL チェックが通過 (0x6261 != NULL) → dereference → SEGV **修正内容:** 全 TLS 配列に明示的初期化子 `= {0}` を追加: 1. **core/hakmem_tiny.c:** - `g_tls_sll_head[TINY_NUM_CLASSES] = {0}` - `g_tls_sll_count[TINY_NUM_CLASSES] = {0}` - `g_tls_live_ss[TINY_NUM_CLASSES] = {0}` - `g_tls_bcur[TINY_NUM_CLASSES] = {0}` - `g_tls_bend[TINY_NUM_CLASSES] = {0}` 2. **core/tiny_fastcache.c:** - `g_tiny_fast_cache[TINY_FAST_CLASS_COUNT] = {0}` - `g_tiny_fast_count[TINY_FAST_CLASS_COUNT] = {0}` - `g_tiny_fast_free_head[TINY_FAST_CLASS_COUNT] = {0}` - `g_tiny_fast_free_count[TINY_FAST_CLASS_COUNT] = {0}` 3. **core/hakmem_tiny_magazine.c:** - `g_tls_mags[TINY_NUM_CLASSES] = {0}` 4. **core/tiny_sticky.c:** - `g_tls_sticky_ss[TINY_NUM_CLASSES][TINY_STICKY_RING] = {0}` - `g_tls_sticky_idx[TINY_NUM_CLASSES][TINY_STICKY_RING] = {0}` - `g_tls_sticky_pos[TINY_NUM_CLASSES] = {0}` **効果:** ``` Before: 1T: 2.09M ✅ | 4T: SEGV 💀 After: 1T: 2.41M ✅ | 4T: 4.19M ✅ (+15% 1T, SEGV解消) ``` **テスト:** ```bash # 1 thread: 完走 ./larson_hakmem 2 8 128 1024 1 12345 1 → Throughput = 2,407,597 ops/s ✅ # 4 threads: 完走(以前は SEGV) ./larson_hakmem 2 8 128 1024 1 12345 4 → Throughput = 4,192,155 ops/s ✅ ``` **調査協力:** Task agent (ultrathink mode) による完璧な根本原因特定 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-07 01:27:04 +09:00
// pool_api.inc.h — Box: L2 Pool public API (alloc/free/lookup)
#ifndef POOL_API_INC_H
#define POOL_API_INC_H
void* hak_pool_try_alloc(size_t size, uintptr_t site_id) {
hak_pool_init(); // pthread_once() ensures thread-safe init (no data race!)
// P1.7 approach: Avoid using pool during ALL wrapper calls (conservative but safe)
extern int hak_in_wrapper(void);
if (hak_in_wrapper() && !g_wrap_l2_enabled) return NULL;
if (!hak_pool_is_poolable(size)) return NULL;
// Get class and shard indices
int class_idx = hak_pool_get_class_index(size);
if (class_idx < 0) return NULL;
// MF2: Per-Page Sharding path
if (g_mf2_enabled) {
return mf2_alloc_fast(class_idx, size, site_id);
}
// OLD PATH: TLS fast path (ring then local LIFO); drain TC only when needed
PoolTLSRing* ring = &g_tls_bin[class_idx].ring;
if (g_tc_enabled && ring->top < g_tc_drain_trigger && mid_tc_has_items(class_idx)) {
HKM_TIME_START(t_tc_drain);
if (mid_tc_drain_into_tls(class_idx, ring, &g_tls_bin[class_idx])) {
HKM_TIME_END(HKM_CAT_TC_DRAIN, t_tc_drain);
if (ring->top > 0) {
HKM_TIME_START(t_ring_pop0);
PoolBlock* tlsb = ring->items[--ring->top];
HKM_TIME_END(HKM_CAT_POOL_TLS_RING_POP, t_ring_pop0);
void* raw = (void*)tlsb;
AllocHeader* hdr = (AllocHeader*)raw;
mid_set_header(hdr, g_class_sizes[class_idx], site_id);
mid_page_inuse_inc(raw);
t_pool_rng ^= t_pool_rng << 13; t_pool_rng ^= t_pool_rng >> 17; t_pool_rng ^= t_pool_rng << 5;
if ((t_pool_rng & ((1u<<g_count_sample_exp)-1u)) == 0u) g_pool.hits[class_idx]++;
return (char*)raw + HEADER_SIZE;
}
} else { HKM_TIME_END(HKM_CAT_TC_DRAIN, t_tc_drain); }
}
if (g_tls_ring_enabled) {
if (ring->top == 0) {
atomic_fetch_add_explicit(&g_pool.ring_underflow, 1, memory_order_relaxed);
}
if (ring->top > 0) {
HKM_TIME_START(t_ring_pop1);
PoolBlock* tlsb = ring->items[--ring->top];
HKM_TIME_END(HKM_CAT_POOL_TLS_RING_POP, t_ring_pop1);
void* raw = (void*)tlsb;
AllocHeader* hdr = (AllocHeader*)raw;
mid_set_header(hdr, g_class_sizes[class_idx], site_id);
t_pool_rng ^= t_pool_rng << 13; t_pool_rng ^= t_pool_rng >> 17; t_pool_rng ^= t_pool_rng << 5;
if ((t_pool_rng & ((1u<<g_count_sample_exp)-1u)) == 0u) g_pool.hits[class_idx]++;
return (char*)raw + HEADER_SIZE;
}
}
if (g_tls_bin[class_idx].lo_head) {
HKM_TIME_START(t_lifo_pop0);
PoolBlock* b = g_tls_bin[class_idx].lo_head;
g_tls_bin[class_idx].lo_head = b->next;
if (g_tls_bin[class_idx].lo_count) g_tls_bin[class_idx].lo_count--;
HKM_TIME_END(HKM_CAT_POOL_TLS_LIFO_POP, t_lifo_pop0);
void* raw = (void*)b; AllocHeader* hdr = (AllocHeader*)raw;
mid_set_header(hdr, g_class_sizes[class_idx], site_id);
mid_page_inuse_inc(raw);
t_pool_rng ^= t_pool_rng << 13; t_pool_rng ^= t_pool_rng >> 17; t_pool_rng ^= t_pool_rng << 5;
if ((t_pool_rng & ((1u<<g_count_sample_exp)-1u)) == 0u) g_pool.hits[class_idx]++;
return (char*)raw + HEADER_SIZE;
}
// Compute shard only when we need to access shared structures
int shard_idx = hak_pool_get_shard_index(site_id);
// Try to batch-pop from a non-empty shard using trylock to fill TLS ring
if (g_tls_ring_enabled) {
int s0 = choose_nonempty_shard(class_idx, shard_idx);
for (int probe = 0; probe < g_trylock_probes; ++probe) {
int s = (s0 + probe) & (POOL_NUM_SHARDS - 1);
pthread_mutex_t* l = &g_pool.freelist_locks[class_idx][s].m;
atomic_fetch_add_explicit(&g_pool.trylock_attempts, 1, memory_order_relaxed);
if (pthread_mutex_trylock(l) == 0) {
atomic_fetch_add_explicit(&g_pool.trylock_success, 1, memory_order_relaxed);
// First, drain any remote frees into freelist
if (atomic_load_explicit(&g_pool.remote_count[class_idx][s], memory_order_relaxed) != 0) {
drain_remote_locked(class_idx, s);
}
PoolBlock* head = g_pool.freelist[class_idx][s];
int to_ring = POOL_L2_RING_CAP - ring->top; if (to_ring < 0) to_ring = 0;
while (head && to_ring-- > 0) { PoolBlock* nxt = head->next; ring->items[ring->top++] = head; head = nxt; }
while (head) { PoolBlock* nxt = head->next; head->next = g_tls_bin[class_idx].lo_head; g_tls_bin[class_idx].lo_head = head; g_tls_bin[class_idx].lo_count++; head = nxt; }
g_pool.freelist[class_idx][s] = head;
if (!head) clear_nonempty_bit(class_idx, s);
pthread_mutex_unlock(l);
if (ring->top > 0) {
PoolBlock* tlsb = ring->items[--ring->top];
void* raw = (void*)tlsb;
AllocHeader* hdr = (AllocHeader*)raw;
mid_set_header(hdr, g_class_sizes[class_idx], site_id);
mid_page_inuse_inc(raw);
t_pool_rng ^= t_pool_rng << 13; t_pool_rng ^= t_pool_rng >> 17; t_pool_rng ^= t_pool_rng << 5;
if ((t_pool_rng & ((1u<<g_count_sample_exp)-1u)) == 0u) g_pool.hits[class_idx]++;
return (char*)raw + HEADER_SIZE;
}
}
}
}
// Try TLS active pages (owner-only local bump-run, up to 3)
PoolTLSPage* ap = NULL;
if (g_tls_active_page_a[class_idx].page && g_tls_active_page_a[class_idx].count > 0 && g_tls_active_page_a[class_idx].bump < g_tls_active_page_a[class_idx].end) ap = &g_tls_active_page_a[class_idx];
else if (g_tls_active_page_b[class_idx].page && g_tls_active_page_b[class_idx].count > 0 && g_tls_active_page_b[class_idx].bump < g_tls_active_page_b[class_idx].end) ap = &g_tls_active_page_b[class_idx];
else if (g_tls_active_page_c[class_idx].page && g_tls_active_page_c[class_idx].count > 0 && g_tls_active_page_c[class_idx].bump < g_tls_active_page_c[class_idx].end) ap = &g_tls_active_page_c[class_idx];
if (ap) {
if (g_tls_ring_enabled && ring->top < POOL_L2_RING_CAP) {
int need = POOL_L2_RING_CAP - ring->top;
(void)refill_tls_from_active_page(class_idx, ring, &g_tls_bin[class_idx], ap, need);
}
PoolBlock* b = NULL;
if (ring->top > 0) { b = ring->items[--ring->top]; }
else if (ap->page && ap->count > 0 && ap->bump < ap->end) {
b = (PoolBlock*)(void*)ap->bump; ap->bump += (HEADER_SIZE + g_class_sizes[class_idx]); ap->count--; if (ap->bump >= ap->end || ap->count<=0){ ap->page=NULL; ap->count=0; }
}
if (b) {
void* raw = (void*)b; AllocHeader* hdr = (AllocHeader*)raw;
mid_set_header(hdr, g_class_sizes[class_idx], site_id);
mid_page_inuse_inc(raw);
g_pool.hits[class_idx]++;
return (char*)raw + HEADER_SIZE;
}
}
// Lock the shard freelist for this (class, shard)
pthread_mutex_t* lock = &g_pool.freelist_locks[class_idx][shard_idx].m;
HKM_TIME_START(t_lock);
struct timespec ts_lk1; int lk1 = hkm_prof_begin(&ts_lk1);
(void)ts_lk1; (void)lk1; // Unused profiling variables
pthread_mutex_lock(lock);
HKM_TIME_END(HKM_CAT_POOL_LOCK, t_lock);
hkm_prof_end(lk1, HKP_POOL_LOCK, &ts_lk1);
// Try to pop from freelist
PoolBlock* block = g_pool.freelist[class_idx][shard_idx];
if (!block) {
// Before refilling, try draining remote stack and simple shard steal
int stole = 0;
const FrozenPolicy* pol = hkm_policy_get();
if (pol) {
uint16_t cap = 0;
if (class_idx < 5) cap = pol->mid_cap[class_idx];
else if (class_idx == 5 && pol->mid_dyn1_bytes != 0) cap = pol->mid_cap_dyn1;
else if (class_idx == 6 && pol->mid_dyn2_bytes != 0) cap = pol->mid_cap_dyn2;
// Drain remotes
if (atomic_load_explicit(&g_pool.remote_count[class_idx][shard_idx], memory_order_relaxed) != 0) {
drain_remote_locked(class_idx, shard_idx);
block = g_pool.freelist[class_idx][shard_idx];
}
// Light shard steal when over cap
if (!block && cap > 0 && g_pool.pages_by_class[class_idx] >= cap) {
HKM_TIME_START(t_steal);
for (int d = 1; d <= 4 && !stole; d++) {
int s1 = (shard_idx + d) & (POOL_NUM_SHARDS - 1);
int s2 = (shard_idx - d) & (POOL_NUM_SHARDS - 1);
if (is_shard_nonempty(class_idx, s1)) {
pthread_mutex_t* l2 = &g_pool.freelist_locks[class_idx][s1].m;
pthread_mutex_lock(l2);
PoolBlock* b2 = g_pool.freelist[class_idx][s1];
if (b2) {
g_pool.freelist[class_idx][s1] = b2->next;
if (!g_pool.freelist[class_idx][s1]) clear_nonempty_bit(class_idx, s1);
block = b2; stole = 1;
}
pthread_mutex_unlock(l2);
}
if (!stole && is_shard_nonempty(class_idx, s2)) {
pthread_mutex_t* l3 = &g_pool.freelist_locks[class_idx][s2].m;
pthread_mutex_lock(l3);
PoolBlock* b3 = g_pool.freelist[class_idx][s2];
if (b3) {
g_pool.freelist[class_idx][s2] = b3->next;
if (!g_pool.freelist[class_idx][s2]) clear_nonempty_bit(class_idx, s2);
block = b3; stole = 1;
}
pthread_mutex_unlock(l3);
}
}
HKM_TIME_END(HKM_CAT_SHARD_STEAL, t_steal);
}
}
if (!stole && !block) {
// Freelist empty, refill page
PoolTLSPage* tap = NULL;
if (g_tls_active_page_a[class_idx].page == NULL || g_tls_active_page_a[class_idx].count == 0) tap = &g_tls_active_page_a[class_idx];
else if (g_tls_active_page_b[class_idx].page == NULL || g_tls_active_page_b[class_idx].count == 0) tap = &g_tls_active_page_b[class_idx];
else if (g_tls_active_page_c[class_idx].page == NULL || g_tls_active_page_c[class_idx].count == 0) tap = &g_tls_active_page_c[class_idx];
else tap = &g_tls_active_page_a[class_idx];
HKM_TIME_START(t_alloc_page);
if (alloc_tls_page(class_idx, tap)) {
HKM_TIME_END(HKM_CAT_POOL_ALLOC_TLS_PAGE, t_alloc_page);
pthread_mutex_unlock(lock);
// Top-up ring and return
ap = tap;
if (g_tls_ring_enabled && ring->top < POOL_L2_RING_CAP) {
int need = POOL_L2_RING_CAP - ring->top;
(void)refill_tls_from_active_page(class_idx, ring, &g_tls_bin[class_idx], ap, need);
}
PoolBlock* takeb = NULL;
if (ring->top > 0) { HKM_TIME_START(t_ring_pop2); takeb = ring->items[--ring->top]; HKM_TIME_END(HKM_CAT_POOL_TLS_RING_POP, t_ring_pop2);}
else if (ap->page && ap->count > 0 && ap->bump < ap->end) { takeb = (PoolBlock*)(void*)ap->bump; ap->bump += (HEADER_SIZE + g_class_sizes[class_idx]); ap->count--; if (ap->bump >= ap->end || ap->count==0){ ap->page=NULL; ap->count=0; } }
void* raw2 = (void*)takeb; AllocHeader* hdr2 = (AllocHeader*)raw2;
mid_set_header(hdr2, g_class_sizes[class_idx], site_id);
mid_page_inuse_inc(raw2);
g_pool.hits[class_idx]++;
return (char*)raw2 + HEADER_SIZE;
}
HKM_TIME_START(t_refill);
struct timespec ts_rf; int rf = hkm_prof_begin(&ts_rf);
int ok = refill_freelist(class_idx, shard_idx);
HKM_TIME_END(HKM_CAT_POOL_REFILL, t_refill);
hkm_prof_end(rf, HKP_POOL_REFILL, &ts_rf);
if (!ok) {
t_pool_rng ^= t_pool_rng << 13; t_pool_rng ^= t_pool_rng >> 17; t_pool_rng ^= t_pool_rng << 5;
if ((t_pool_rng & ((1u<<g_count_sample_exp)-1u)) == 0u) g_pool.misses[class_idx]++;
pthread_mutex_unlock(lock);
return NULL;
}
}
}
// Pop block and adopt page
g_pool.freelist[class_idx][shard_idx] = block->next;
mid_desc_adopt(block, class_idx, (uint64_t)(uintptr_t)pthread_self());
t_pool_rng ^= t_pool_rng << 13; t_pool_rng ^= t_pool_rng >> 17; t_pool_rng ^= t_pool_rng << 5;
if ((t_pool_rng & ((1u<<g_count_sample_exp)-1u)) == 0u) g_pool.hits[class_idx]++;
if (g_pool.freelist[class_idx][shard_idx] == NULL) clear_nonempty_bit(class_idx, shard_idx);
pthread_mutex_unlock(lock);
// Store to TLS then pop
PoolBlock* take;
if (g_tls_ring_enabled && ring->top < POOL_L2_RING_CAP) { ring->items[ring->top++] = block; take = ring->items[--ring->top]; }
else { block->next = g_tls_bin[class_idx].lo_head; g_tls_bin[class_idx].lo_head = block; g_tls_bin[class_idx].lo_count++;
if (g_tls_ring_enabled && ring->top > 0) { take = ring->items[--ring->top]; }
else { take = g_tls_bin[class_idx].lo_head; g_tls_bin[class_idx].lo_head = take->next; if (g_tls_bin[class_idx].lo_count) g_tls_bin[class_idx].lo_count--; } }
void* raw = (void*)take; AllocHeader* hdr = (AllocHeader*)raw;
mid_set_header(hdr, g_class_sizes[class_idx], site_id);
mid_page_inuse_inc(raw);
return (char*)raw + HEADER_SIZE;
}
void hak_pool_free(void* ptr, size_t size, uintptr_t site_id) {
if (!ptr) return;
hak_pool_init();
if (!hak_pool_is_poolable(size)) return;
if (g_mf2_enabled) { mf2_free(ptr); return; }
void* raw = (char*)ptr - HEADER_SIZE;
AllocHeader* hdr = (AllocHeader*)raw;
int mid_by_desc = 0; MidPageDesc* d_desc = mid_desc_lookup(ptr);
if (d_desc) mid_by_desc = 1;
if (!mid_by_desc && g_hdr_light_enabled < 2) {
if (hdr->magic != HAKMEM_MAGIC) { MF2_ERROR_LOG("Invalid magic 0x%X in pool_free, expected 0x%X", hdr->magic, HAKMEM_MAGIC); return; }
if (hdr->method != ALLOC_METHOD_POOL) { MF2_ERROR_LOG("Wrong method %d in pool_free, expected POOL (%d)", hdr->method, ALLOC_METHOD_POOL); return; }
}
int class_idx = mid_by_desc ? (int)d_desc->class_idx : hak_pool_get_class_index(size);
if (class_idx < 0) return;
PoolBlock* block = (PoolBlock*)raw;
if (g_pool.tls_free_enabled) {
int same_thread = 0;
if (g_hdr_light_enabled >= 1) { MidPageDesc* d = mid_desc_lookup(raw); if (d && d->owner_tid != 0 && d->owner_tid == (uint64_t)(uintptr_t)pthread_self()) { same_thread = 1; } }
else if (hdr->owner_tid != 0 && hdr->owner_tid == (uintptr_t)(uintptr_t)pthread_self()) { same_thread = 1; }
if (same_thread) {
PoolTLSRing* ring = &g_tls_bin[class_idx].ring;
if (g_tls_ring_enabled && ring->top < POOL_L2_RING_CAP) { ring->items[ring->top++] = block; }
else { block->next = g_tls_bin[class_idx].lo_head; g_tls_bin[class_idx].lo_head = block; g_tls_bin[class_idx].lo_count++; if ((int)g_tls_bin[class_idx].lo_count > g_tls_lo_max) {
size_t spill = g_tls_bin[class_idx].lo_count / 2; int shard = hak_pool_get_shard_index(site_id);
while (spill-- && g_tls_bin[class_idx].lo_head) { PoolBlock* b = g_tls_bin[class_idx].lo_head; g_tls_bin[class_idx].lo_head = b->next; g_tls_bin[class_idx].lo_count--; HKM_TIME_START(t_remote_push1); uintptr_t old_head; do { old_head = atomic_load_explicit(&g_pool.remote_head[class_idx][shard], memory_order_acquire); b->next = (PoolBlock*)old_head; } while (!atomic_compare_exchange_weak_explicit(&g_pool.remote_head[class_idx][shard], &old_head, (uintptr_t)b, memory_order_release, memory_order_relaxed)); atomic_fetch_add_explicit(&g_pool.remote_count[class_idx][shard], 1, memory_order_relaxed); HKM_TIME_END(HKM_CAT_POOL_REMOTE_PUSH, t_remote_push1); } set_nonempty_bit(class_idx, shard); } }
} else {
if (g_tc_enabled) { uint64_t owner_tid = 0; if (g_hdr_light_enabled < 2) owner_tid = hdr->owner_tid; if (owner_tid == 0) { MidPageDesc* d = mid_desc_lookup(raw); if (d) owner_tid = d->owner_tid; } if (owner_tid != 0) { MidTC* otc = mid_tc_lookup_by_tid(owner_tid); if (otc) { mid_tc_push(otc, class_idx, block); return; } } }
int shard = hak_pool_get_shard_index(site_id); uintptr_t old_head; HKM_TIME_START(t_remote_push2);
do { old_head = atomic_load_explicit(&g_pool.remote_head[class_idx][shard], memory_order_acquire); block->next = (PoolBlock*)old_head; } while (!atomic_compare_exchange_weak_explicit(&g_pool.remote_head[class_idx][shard], &old_head, (uintptr_t)block, memory_order_release, memory_order_relaxed));
atomic_fetch_add_explicit(&g_pool.remote_count[class_idx][shard], 1, memory_order_relaxed); HKM_TIME_END(HKM_CAT_POOL_REMOTE_PUSH, t_remote_push2); set_nonempty_bit(class_idx, shard);
}
} else {
int shard_idx2 = hak_pool_get_shard_index(site_id); pthread_mutex_t* lock = &g_pool.freelist_locks[class_idx][shard_idx2].m; pthread_mutex_lock(lock); block->next = g_pool.freelist[class_idx][shard_idx2]; g_pool.freelist[class_idx][shard_idx2] = block; set_nonempty_bit(class_idx, shard_idx2); pthread_mutex_unlock(lock);
}
t_pool_rng ^= t_pool_rng << 13; t_pool_rng ^= t_pool_rng >> 17; t_pool_rng ^= t_pool_rng << 5; if ((t_pool_rng & ((1u<<g_count_sample_exp)-1u)) == 0u) g_pool.frees[class_idx]++;
mid_page_inuse_dec_and_maybe_dn(raw);
}
int hak_pool_mid_lookup(void* ptr, size_t* out_size) {
if (g_mf2_enabled) { MidPage* page = mf2_addr_to_page(ptr); if (page) { int c = (int)page->class_idx; if (c < 0 || c >= POOL_NUM_CLASSES) return 0; size_t sz = g_class_sizes[c]; if (sz == 0) return 0; if (out_size) *out_size = sz; return 1; } }
MidPageDesc* d = mid_desc_lookup(ptr); if (!d) return 0; int c = (int)d->class_idx; if (c < 0 || c >= POOL_NUM_CLASSES) return 0; size_t sz = g_class_sizes[c]; if (sz == 0) return 0; if (out_size) *out_size = sz; return 1;
}
void hak_pool_free_fast(void* ptr, uintptr_t site_id) {
if (!ptr || !g_pool.initialized) return; if (g_mf2_enabled) { MidPage* page = mf2_addr_to_page(ptr); if (page) { mf2_free(ptr); return; } }
MidPageDesc* d = mid_desc_lookup(ptr); if (!d) return; size_t sz = g_class_sizes[(int)d->class_idx]; if (sz == 0) return; hak_pool_free(ptr, sz, site_id);
}
#endif // POOL_API_INC_H