perf: move AG and SCNG into native __thread storage

Replay of GH-22231

Closes GH-22595

Author: henderkes

NEWS

@@ -12,6 +12,7 @@ PHP                                                                        NEWS
   . Lock unmodified readonly properties for modification after clone-with.
     (NickSdot)
   . abort() instead of exit() on hard OOM. (realFlowControl)
+  . perf: ZTS: move AG and SCNG into native __thread storage. (henderkes)
 
 - Calendar:
   . Fixed bug GH-22602 (gregoriantojd() and juliantojd() integer overflow with

TSRM/TSRM.c

@@ -41,6 +41,8 @@ typedef struct {
 	ts_allocate_ctor ctor;
 	ts_allocate_dtor dtor;
 	ptrdiff_t fast_offset;
+	/* When set, storage comes from __thread memory instead of being allocated by TSRM. */
+	void *(*tls_addr)(void);
 	int done;
 } tsrm_resource_type;
 
@@ -164,14 +166,20 @@ TSRM_API bool tsrm_startup(int expected_threads, int expected_resources, int deb
 
 static void ts_free_resources(tsrm_tls_entry *thread_resources)
 {
+	bool own_thread = thread_resources->thread_id == tsrm_thread_id();
+
 	/* Need to destroy in reverse order to respect dependencies. */
 	for (int i = thread_resources->count - 1; i >= 0; i--) {
 		if (!resource_types_table[i].done) {
+			/* A __thread block of a foreign thread is inaccessible. */
+			if (resource_types_table[i].tls_addr && !own_thread) {
+				continue;
+			}
 			if (resource_types_table[i].dtor) {
 				resource_types_table[i].dtor(thread_resources->storage[i]);
 			}
 
-			if (!resource_types_table[i].fast_offset) {
+			if (!resource_types_table[i].fast_offset && !resource_types_table[i].tls_addr) {
 				free(thread_resources->storage[i]);
 			}
 		}
@@ -180,7 +188,9 @@ static void ts_free_resources(tsrm_tls_entry *thread_resources)
 	free(thread_resources->storage);
 }
 
-/* Shutdown TSRM (call once for the entire process) */
+/* Shutdown TSRM (call once for the entire process). Tears down every thread left
+ * in the table. For resources allocated with ts_allocate_tls_id(), only the dtor
+ * of the calling thread is invoked. */
 TSRM_API void tsrm_shutdown(void)
 {/*{{{*/
 	if (is_thread_shutdown) {
@@ -258,7 +268,10 @@ static void tsrm_update_active_threads(void)
 
 				p->storage = (void *) realloc(p->storage, sizeof(void *)*id_count);
 				for (j=p->count; j<id_count; j++) {
-					if (resource_types_table[j].fast_offset) {
+					if (resource_types_table[j].tls_addr) {
+						TSRM_ASSERT(p->thread_id == tsrm_thread_id());
+						p->storage[j] = resource_types_table[j].tls_addr();
+					} else if (resource_types_table[j].fast_offset) {
 						p->storage[j] = (void *) (((char*)p) + resource_types_table[j].fast_offset);
 					} else {
 						p->storage[j] = (void *) malloc(resource_types_table[j].size);
@@ -303,6 +316,7 @@ TSRM_API ts_rsrc_id ts_allocate_id(ts_rsrc_id *rsrc_id, size_t size, ts_allocate
 	resource_types_table[TSRM_UNSHUFFLE_RSRC_ID(*rsrc_id)].ctor = ctor;
 	resource_types_table[TSRM_UNSHUFFLE_RSRC_ID(*rsrc_id)].dtor = dtor;
 	resource_types_table[TSRM_UNSHUFFLE_RSRC_ID(*rsrc_id)].fast_offset = 0;
+	resource_types_table[TSRM_UNSHUFFLE_RSRC_ID(*rsrc_id)].tls_addr = NULL;
 	resource_types_table[TSRM_UNSHUFFLE_RSRC_ID(*rsrc_id)].done = 0;
 
 	tsrm_update_active_threads();
@@ -381,6 +395,7 @@ TSRM_API ts_rsrc_id ts_allocate_fast_id_at(ts_rsrc_id *rsrc_id, size_t *offset,
 	resource_types_table[TSRM_UNSHUFFLE_RSRC_ID(*rsrc_id)].ctor = ctor;
 	resource_types_table[TSRM_UNSHUFFLE_RSRC_ID(*rsrc_id)].dtor = dtor;
 	resource_types_table[TSRM_UNSHUFFLE_RSRC_ID(*rsrc_id)].fast_offset = *offset;
+	resource_types_table[TSRM_UNSHUFFLE_RSRC_ID(*rsrc_id)].tls_addr = NULL;
 	resource_types_table[TSRM_UNSHUFFLE_RSRC_ID(*rsrc_id)].done = 0;
 
 	tsrm_update_active_threads();
@@ -390,6 +405,41 @@ TSRM_API ts_rsrc_id ts_allocate_fast_id_at(ts_rsrc_id *rsrc_id, size_t *offset,
 	return *rsrc_id;
 }
 
+/* allocates a resource id whose per-thread storage is a native __thread block */
+TSRM_API ts_rsrc_id ts_allocate_tls_id(ts_rsrc_id *rsrc_id, void *(*tls_addr)(void), size_t size, ts_allocate_ctor ctor, ts_allocate_dtor dtor)
+{
+	TSRM_ERROR((TSRM_ERROR_LEVEL_CORE, "Obtaining a new TLS resource id, %d bytes", size));
+
+	tsrm_mutex_lock(tsmm_mutex);
+
+	*rsrc_id = TSRM_SHUFFLE_RSRC_ID(id_count++);
+
+	if (resource_types_table_size < id_count) {
+		tsrm_resource_type *_tmp;
+		_tmp = (tsrm_resource_type *) realloc(resource_types_table, sizeof(tsrm_resource_type)*id_count);
+		if (!_tmp) {
+			TSRM_ERROR((TSRM_ERROR_LEVEL_ERROR, "Unable to allocate storage for resource"));
+			*rsrc_id = 0;
+			tsrm_mutex_unlock(tsmm_mutex);
+			return 0;
+		}
+		resource_types_table = _tmp;
+		resource_types_table_size = id_count;
+	}
+	resource_types_table[TSRM_UNSHUFFLE_RSRC_ID(*rsrc_id)].size = size;
+	resource_types_table[TSRM_UNSHUFFLE_RSRC_ID(*rsrc_id)].ctor = ctor;
+	resource_types_table[TSRM_UNSHUFFLE_RSRC_ID(*rsrc_id)].dtor = dtor;
+	resource_types_table[TSRM_UNSHUFFLE_RSRC_ID(*rsrc_id)].fast_offset = 0;
+	resource_types_table[TSRM_UNSHUFFLE_RSRC_ID(*rsrc_id)].tls_addr = tls_addr;
+	resource_types_table[TSRM_UNSHUFFLE_RSRC_ID(*rsrc_id)].done = 0;
+
+	tsrm_update_active_threads();
+	tsrm_mutex_unlock(tsmm_mutex);
+
+	TSRM_ERROR((TSRM_ERROR_LEVEL_CORE, "Successfully allocated new TLS resource id %d", *rsrc_id));
+	return *rsrc_id;
+}
+
 static void set_thread_local_storage_resource_to(tsrm_tls_entry *thread_resource)
 {
 	tsrm_tls_set(thread_resource);
@@ -422,7 +472,9 @@ static void allocate_new_resource(tsrm_tls_entry **thread_resources_ptr, THREAD_
 		if (resource_types_table[i].done) {
 			(*thread_resources_ptr)->storage[i] = NULL;
 		} else {
-			if (resource_types_table[i].fast_offset) {
+			if (resource_types_table[i].tls_addr) {
+				(*thread_resources_ptr)->storage[i] = resource_types_table[i].tls_addr();
+			} else if (resource_types_table[i].fast_offset) {
 				(*thread_resources_ptr)->storage[i] = (void *) (((char*)(*thread_resources_ptr)) + resource_types_table[i].fast_offset);
 			} else {
 				(*thread_resources_ptr)->storage[i] = (void *) malloc(resource_types_table[i].size);
@@ -510,7 +562,9 @@ TSRM_API void *ts_resource_ex(ts_rsrc_id id, THREAD_T *th_id)
 		/* In case that extensions don't use the pointer passed from the dtor, but incorrectly
 		 * use the global pointer, we need to setup the global pointer temporarily here. */
 		set_thread_local_storage_resource_to(thread_resources);
-		/* Free up the old resource from the old thread instance */
+		/* Dead thread with a recycled id: its __thread blocks are gone, and this
+		 * thread's blocks were never constructed, so keep tls dtors from running. */
+		thread_resources->thread_id = 0;
 		ts_free_resources(thread_resources);
 		free((char *) thread_resources - tsrm_reserved_front);
 		/* Allocate a new resource at the same point in the linked list, and relink the next pointer */
@@ -584,7 +638,7 @@ void ts_free_id(ts_rsrc_id id)
 						if (resource_types_table[rsrc_id].dtor) {
 							resource_types_table[rsrc_id].dtor(p->storage[rsrc_id]);
 						}
-						if (!resource_types_table[rsrc_id].fast_offset) {
+						if (!resource_types_table[rsrc_id].fast_offset && !resource_types_table[rsrc_id].tls_addr) {
 							free(p->storage[rsrc_id]);
 						}
 					}

TSRM/TSRM.h

@@ -100,6 +100,10 @@ TSRM_API ts_rsrc_id ts_allocate_fast_id(ts_rsrc_id *rsrc_id, size_t *offset, siz
 TSRM_API void tsrm_reserve_fast_front(size_t size);
 TSRM_API ts_rsrc_id ts_allocate_fast_id_at(ts_rsrc_id *rsrc_id, size_t *offset, ptrdiff_t fixed_offset, size_t size, ts_allocate_ctor ctor, ts_allocate_dtor dtor);
 
+/* Resource whose per-thread storage is a native __thread block.
+ * Must be called at startup before any other thread exists. */
+TSRM_API ts_rsrc_id ts_allocate_tls_id(ts_rsrc_id *rsrc_id, void *(*tls_addr)(void), size_t size, ts_allocate_ctor ctor, ts_allocate_dtor dtor);
+
 /* fetches the requested resource for the current thread */
 TSRM_API void *ts_resource_ex(ts_rsrc_id id, THREAD_T *th_id);
 #define ts_resource(id)			ts_resource_ex(id, NULL)

UPGRADING.INTERNALS

@@ -235,3 +235,11 @@ PHP 8.6 INTERNALS UPGRADE NOTES
 ========================
 5. SAPI changes
 ========================
+
+- SAPIs should explicitly release a thread's resources by calling
+  ts_free_thread() before terminating it. tsrm_shutdown() can only release the
+  resources of the calling thread, for resources allocated with
+  ts_allocate_tls_id().
+
+- AG and SCNG are now allocated with ts_allocate_tls_id() and live in native
+  __thread storage on ZTS builds.

Zend/zend.c

@@ -54,6 +54,9 @@ ZEND_API int compiler_globals_id;
 ZEND_API int executor_globals_id;
 ZEND_API size_t compiler_globals_offset;
 ZEND_API size_t executor_globals_offset;
+/* ts_allocate_tls_id takes a callback so each thread resolves its own block.
+ * A plain &language_scanner_globals would capture only the registering thread's address. */
+static void *language_scanner_globals_tls_addr(void) { return &language_scanner_globals; }
 static HashTable *global_function_table = NULL;
 static HashTable *global_class_table = NULL;
 static HashTable *global_constants_table = NULL;
@@ -1021,10 +1024,9 @@ void zend_startup(zend_utility_functions *utility_functions) /* {{{ */
 #ifdef ZTS
 	ts_allocate_fast_id_at(&compiler_globals_id, &compiler_globals_offset, ZEND_CG_OFFSET, sizeof(zend_compiler_globals), (ts_allocate_ctor) compiler_globals_ctor, (ts_allocate_dtor) compiler_globals_dtor);
 	ts_allocate_fast_id_at(&executor_globals_id, &executor_globals_offset, ZEND_EG_OFFSET, sizeof(zend_executor_globals), (ts_allocate_ctor) executor_globals_ctor, (ts_allocate_dtor) executor_globals_dtor);
-	ts_allocate_fast_id_at(&language_scanner_globals_id, &language_scanner_globals_offset, ZEND_SCNG_OFFSET, sizeof(zend_php_scanner_globals), (ts_allocate_ctor) php_scanner_globals_ctor, NULL);
+	ts_allocate_tls_id(&language_scanner_globals_id, language_scanner_globals_tls_addr, sizeof(zend_php_scanner_globals), (ts_allocate_ctor) php_scanner_globals_ctor, NULL);
 	ZEND_ASSERT(compiler_globals_offset == ZEND_CG_OFFSET);
 	ZEND_ASSERT(executor_globals_offset == ZEND_EG_OFFSET);
-	ZEND_ASSERT(language_scanner_globals_offset == ZEND_SCNG_OFFSET);
 	ts_allocate_fast_id(&ini_scanner_globals_id, &ini_scanner_globals_offset, sizeof(zend_ini_scanner_globals), (ts_allocate_ctor) ini_scanner_globals_ctor, NULL);
 	compiler_globals = ts_resource(compiler_globals_id);
 	executor_globals = ts_resource(executor_globals_id);

Zend/zend_alloc.c

@@ -2613,13 +2613,12 @@ typedef struct _zend_alloc_globals {
 
 #ifdef ZTS
 static int alloc_globals_id;
-static size_t alloc_globals_offset;
-# define ZEND_AG_OFFSET (ZEND_SCNG_OFFSET - (ptrdiff_t) TSRM_ALIGNED_SIZE(sizeof(zend_alloc_globals)))
-# define AG(v) ZEND_TSRMG_FAST(ZEND_AG_OFFSET, zend_alloc_globals *, v)
+static TSRM_TLS TSRM_TLS_MODEL_ATTR zend_alloc_globals alloc_globals;
+static void *alloc_globals_tls_addr(void) { return &alloc_globals; }
 #else
-# define AG(v) (alloc_globals.v)
 static zend_alloc_globals alloc_globals;
 #endif
+#define AG(v) (alloc_globals.v)
 
 ZEND_API bool is_zend_mm(void)
 {
@@ -3336,8 +3335,7 @@ ZEND_API void start_memory_manager(void)
 #  endif
 #endif
 #ifdef ZTS
-	ts_allocate_fast_id_at(&alloc_globals_id, &alloc_globals_offset, ZEND_AG_OFFSET, sizeof(zend_alloc_globals), (ts_allocate_ctor) alloc_globals_ctor, (ts_allocate_dtor) alloc_globals_dtor);
-	ZEND_ASSERT(alloc_globals_offset == ZEND_AG_OFFSET);
+	ts_allocate_tls_id(&alloc_globals_id, alloc_globals_tls_addr, sizeof(zend_alloc_globals), (ts_allocate_ctor) alloc_globals_ctor, (ts_allocate_dtor) alloc_globals_dtor);
 #else
 	alloc_globals_ctor(&alloc_globals);
 #endif
@@ -3583,9 +3581,3 @@ ZEND_API char * __zend_strdup(const char *s)
 	zend_out_of_memory();
 }
 
-#ifdef ZTS
-size_t zend_mm_globals_size(void)
-{
-	return sizeof(zend_alloc_globals);
-}
-#endif

Zend/zend_alloc.h

@@ -403,10 +403,6 @@ static void apc_init_heap(void)
 
 */
 
-#ifdef ZTS
-size_t zend_mm_globals_size(void);
-#endif
-
 END_EXTERN_C()
 
 #endif

Zend/zend_globals.h

@@ -330,11 +330,9 @@ struct _zend_executor_globals {
 };
 
 #ifdef ZTS
-/* Compile-time offsets of the hot globals, in a reserved region just before the
- * cache pointer. ZEND_AG_OFFSET is furthest, in zend_alloc.c. */
+/* Compile-time offsets of the hot globals, in a reserved region just before *_tsrm_ls_cache. */
 # define ZEND_CG_OFFSET   (-(ptrdiff_t) TSRM_ALIGNED_SIZE(sizeof(zend_compiler_globals)))
 # define ZEND_EG_OFFSET   (ZEND_CG_OFFSET - (ptrdiff_t) TSRM_ALIGNED_SIZE(sizeof(zend_executor_globals)))
-# define ZEND_SCNG_OFFSET (ZEND_EG_OFFSET - (ptrdiff_t) TSRM_ALIGNED_SIZE(sizeof(zend_php_scanner_globals)))
 #endif
 
 #define EG_FLAGS_INITIAL				(0)

Zend/zend_globals_macros.h

@@ -48,12 +48,20 @@ extern ZEND_API zend_executor_globals executor_globals;
 
 /* Language Scanner */
 #ifdef ZTS
-# define LANG_SCNG(v) ZEND_TSRMG_FAST(ZEND_SCNG_OFFSET, zend_php_scanner_globals *, v)
 extern ZEND_API ts_rsrc_id language_scanner_globals_id;
-extern ZEND_API size_t language_scanner_globals_offset;
+# if defined(ZEND_WIN32) && !defined(LIBZEND_EXPORTS)
+# define LANG_SCNG(v) TSRMG(language_scanner_globals_id, zend_php_scanner_globals *, v)
+# else
+#  ifdef ZEND_WIN32
+extern TSRM_TLS zend_php_scanner_globals language_scanner_globals;
+#  else
+extern ZEND_API TSRM_TLS TSRM_TLS_MODEL_ATTR zend_php_scanner_globals language_scanner_globals;
+#  endif
+#  define LANG_SCNG(v) (language_scanner_globals.v)
+# endif
 #else
-# define LANG_SCNG(v) (language_scanner_globals.v)
 extern ZEND_API zend_php_scanner_globals language_scanner_globals;
+# define LANG_SCNG(v) (language_scanner_globals.v)
 #endif
 
 

Zend/zend_language_scanner.l

@@ -84,7 +84,11 @@
 #define SCNG	LANG_SCNG
 #ifdef ZTS
 ZEND_API ts_rsrc_id language_scanner_globals_id;
-ZEND_API size_t language_scanner_globals_offset;
+# ifdef ZEND_WIN32
+TSRM_TLS zend_php_scanner_globals language_scanner_globals;
+# else
+ZEND_API TSRM_TLS TSRM_TLS_MODEL_ATTR zend_php_scanner_globals language_scanner_globals;
+# endif
 #else
 ZEND_API zend_php_scanner_globals language_scanner_globals;
 #endif