Improve InnoDB synchronization primitives on Windows

This patch was originally developed by Vladislav Vaintroub.
The main changes are:

 * Use TryEnterCriticalSection in os_fast_mutex_trylock().
 * Use lightweight condition variables on Vista or later Windows;
   but fall back to events on older Windows, such as XP.

This patch also fixes the following bugs:
  bug# 52102 InnoDB Plugin shows performance drop compared to InnoDB
             on Windows
  bug# 53204 os_fastmutex_trylock is implemented incorrectly on Windows

rb://363 approved by Inaam Rana

storage/innobase/CMakeLists.txt

@@ -188,11 +188,7 @@ IF(SIZEOF_PTHREAD_T)
 ENDIF()
 
 IF(MSVC)
-  # Windows atomics do not perform well. Disable Windows atomics by default.
-  # See bug#52102 for details.
-
-  #ADD_DEFINITIONS(-DHAVE_WINDOWS_ATOMICS -DINNODB_RW_LOCKS_USE_ATOMICS -DHAVE_IB_PAUSE_INSTRUCTION)
-  ADD_DEFINITIONS(-DHAVE_IB_PAUSE_INSTRUCTION)
+  ADD_DEFINITIONS(-DHAVE_WINDOWS_ATOMICS -DHAVE_IB_PAUSE_INSTRUCTION)
 ENDIF()
 
 

storage/innobase/include/os0file.h

@@ -177,6 +177,13 @@ log. */
 #define OS_WIN95	2	/*!< Microsoft Windows 95 */
 #define OS_WINNT	3	/*!< Microsoft Windows NT 3.x */
 #define OS_WIN2000	4	/*!< Microsoft Windows 2000 */
+#define OS_WINXP	5	/*!< Microsoft Windows XP
+				or Windows Server 2003 */
+#define OS_WINVISTA	6	/*!< Microsoft Windows Vista
+				or Windows Server 2008 */
+#define OS_WIN7		7	/*!< Microsoft Windows 7
+				or Windows Server 2008 R2 */
+
 
 extern ulint	os_n_file_reads;
 extern ulint	os_n_file_writes;
@@ -368,7 +375,8 @@ typedef DIR*	os_file_dir_t;	/*!< directory stream */
 
 /***********************************************************************//**
 Gets the operating system version. Currently works only on Windows.
-@return	OS_WIN95, OS_WIN31, OS_WINNT, or OS_WIN2000 */
+@return	OS_WIN95, OS_WIN31, OS_WINNT, OS_WIN2000, OS_WINXP, OS_WINVISTA,
+OS_WIN7. */
 UNIV_INTERN
 ulint
 os_get_os_version(void);

storage/innobase/include/os0sync.h

@@ -38,28 +38,18 @@ Created 9/6/1995 Heikki Tuuri
 #include "ut0lst.h"
 
 #ifdef __WIN__
-
+/** Native event (slow)*/
+typedef HANDLE			os_native_event_t;
 /** Native mutex */
-#define os_fast_mutex_t CRITICAL_SECTION
-
-/** Native event */
-typedef HANDLE		os_native_event_t;
-
-/** Operating system event */
-typedef struct os_event_struct	os_event_struct_t;
-/** Operating system event handle */
-typedef os_event_struct_t*	os_event_t;
-
-/** An asynchronous signal sent between threads */
-struct os_event_struct {
-	os_native_event_t		  handle;
-					/*!< Windows event */
-	UT_LIST_NODE_T(os_event_struct_t) os_event_list;
-					/*!< list of all created events */
-};
+typedef CRITICAL_SECTION	os_fast_mutex_t;
+/** Native condition variable. */
+typedef CONDITION_VARIABLE	os_cond_t;
 #else
 /** Native mutex */
-typedef pthread_mutex_t	os_fast_mutex_t;
+typedef pthread_mutex_t		os_fast_mutex_t;
+/** Native condition variable */
+typedef pthread_cond_t		os_cond_t;
+#endif
 
 /** Operating system event */
 typedef struct os_event_struct	os_event_struct_t;
@@ -68,6 +58,10 @@ typedef os_event_struct_t*	os_event_t;
 
 /** An asynchronous signal sent between threads */
 struct os_event_struct {
+#ifdef __WIN__
+	HANDLE		handle;		/*!< kernel event object, slow,
+					used on older Windows */
+#endif
 	os_fast_mutex_t	os_mutex;	/*!< this mutex protects the next
 					fields */
 	ibool		is_set;		/*!< this is TRUE when the event is
@@ -76,24 +70,17 @@ struct os_event_struct {
 					this event */
 	ib_int64_t	signal_count;	/*!< this is incremented each time
 					the event becomes signaled */
-	pthread_cond_t	cond_var;	/*!< condition variable is used in
+	os_cond_t	cond_var;	/*!< condition variable is used in
 					waiting for the event */
 	UT_LIST_NODE_T(os_event_struct_t) os_event_list;
 					/*!< list of all created events */
 };
-#endif
 
 /** Operating system mutex */
 typedef struct os_mutex_struct	os_mutex_str_t;
 /** Operating system mutex handle */
 typedef os_mutex_str_t*		os_mutex_t;
 
-/** Denotes an infinite delay for os_event_wait_time() */
-#define OS_SYNC_INFINITE_TIME	((ulint)(-1))
-
-/** Return value of os_event_wait_time() when the time is exceeded */
-#define OS_SYNC_TIME_EXCEEDED	1
-
 /** Mutex protecting counts and the event and OS 'slow' mutex lists */
 extern os_mutex_t	os_sync_mutex;
 
@@ -187,42 +174,14 @@ os_event_wait_low(
 
 #define os_event_wait(event) os_event_wait_low(event, 0)
 
-/**********************************************************//**
-Waits for an event object until it is in the signaled state or
-a timeout is exceeded. In Unix the timeout is always infinite.
-@return	0 if success, OS_SYNC_TIME_EXCEEDED if timeout was exceeded */
-UNIV_INTERN
-ulint
-os_event_wait_time(
-/*===============*/
-	os_event_t	event,	/*!< in: event to wait */
-	ulint		time);	/*!< in: timeout in microseconds, or
-				OS_SYNC_INFINITE_TIME */
-#ifdef __WIN__
-/**********************************************************//**
-Waits for any event in an OS native event array. Returns if even a single
-one is signaled or becomes signaled.
-@return	index of the event which was signaled */
-UNIV_INTERN
-ulint
-os_event_wait_multiple(
-/*===================*/
-	ulint			n,	/*!< in: number of events in the
-					array */
-	os_native_event_t*	native_event_array);
-					/*!< in: pointer to an array of event
-					handles */
-#endif
 /*********************************************************//**
 Creates an operating system mutex semaphore. Because these are slow, the
 mutex semaphore of InnoDB itself (mutex_t) should be used where possible.
 @return	the mutex handle */
 UNIV_INTERN
 os_mutex_t
-os_mutex_create(
-/*============*/
-	const char*	name);	/*!< in: the name of the mutex, if NULL
-				the mutex is created without a name */
+os_mutex_create(void);
+/*=================*/
 /**********************************************************//**
 Acquires ownership of a mutex semaphore. */
 UNIV_INTERN

storage/innobase/include/os0sync.ic

@@ -28,8 +28,7 @@ Created 9/6/1995 Heikki Tuuri
 #endif
 
 /**********************************************************//**
-Acquires ownership of a fast mutex. Currently in Windows this is the same
-as os_fast_mutex_lock!
+Acquires ownership of a fast mutex.
 @return	0 if success, != 0 if was reserved by another thread */
 UNIV_INLINE
 ulint
@@ -38,9 +37,13 @@ os_fast_mutex_trylock(
 	os_fast_mutex_t*	fast_mutex)	/*!< in: mutex to acquire */
 {
 #ifdef __WIN__
-	EnterCriticalSection(fast_mutex);
+	if (TryEnterCriticalSection(fast_mutex)) {
 
-	return(0);
+		return(0);
+	} else {
+
+		return(1);
+	}
 #else
 	/* NOTE that the MySQL my_pthread.h redefines pthread_mutex_trylock
 	so that it returns 0 on success. In the operating system

storage/innobase/include/srv0srv.h

@@ -112,6 +112,9 @@ OS (provided we compiled Innobase with it in), otherwise we will
 use simulated aio we build below with threads.
 Currently we support native aio on windows and linux */
 extern my_bool	srv_use_native_aio;
+#ifdef __WIN__
+extern ibool	srv_use_native_conditions;
+#endif
 extern ulint	srv_n_data_files;
 extern char**	srv_data_file_names;
 extern ulint*	srv_data_file_sizes;

storage/innobase/os/os0file.c

@@ -183,7 +183,7 @@ struct os_aio_slot_struct{
 					which pending aio operation was
 					completed */
 #ifdef WIN_ASYNC_IO
-	os_event_t	event;		/*!< event object we need in the
+	HANDLE		handle;		/*!< handle object we need in the
 					OVERLAPPED struct */
 	OVERLAPPED	control;	/*!< Windows control block for the
 					aio request */
@@ -225,7 +225,7 @@ struct os_aio_array_struct{
 				aio array outside the ibuf segment */
 	os_aio_slot_t*	slots;	/*!< Pointer to the slots in the array */
 #ifdef __WIN__
-	os_native_event_t* native_events;
+	HANDLE*		handles;
 				/*!< Pointer to an array of OS native
 				event handles where we copied the
 				handles from slots, in the same
@@ -304,7 +304,8 @@ UNIV_INTERN ulint	os_n_pending_reads = 0;
 
 /***********************************************************************//**
 Gets the operating system version. Currently works only on Windows.
-@return	OS_WIN95, OS_WIN31, OS_WINNT, OS_WIN2000 */
+@return	OS_WIN95, OS_WIN31, OS_WINNT, OS_WIN2000, OS_WINXP, OS_WINVISTA,
+OS_WIN7. */
 UNIV_INTERN
 ulint
 os_get_os_version(void)
@@ -322,10 +323,18 @@ os_get_os_version(void)
 	} else if (os_info.dwPlatformId == VER_PLATFORM_WIN32_WINDOWS) {
 		return(OS_WIN95);
 	} else if (os_info.dwPlatformId == VER_PLATFORM_WIN32_NT) {
-		if (os_info.dwMajorVersion <= 4) {
-			return(OS_WINNT);
-		} else {
-			return(OS_WIN2000);
+		switch (os_info.dwMajorVersion) {
+		case 3:
+		case 4:
+			return OS_WINNT;
+		case 5:
+			return (os_info.dwMinorVersion == 0) ? OS_WIN2000
+							     : OS_WINXP;
+		case 6:
+			return (os_info.dwMinorVersion == 0) ? OS_WINVISTA
+							     : OS_WIN7;
+		default:
+			return OS_WIN7;
 		}
 	} else {
 		ut_error;
@@ -673,10 +682,10 @@ os_io_init_simple(void)
 {
 	ulint	i;
 
-	os_file_count_mutex = os_mutex_create(NULL);
+	os_file_count_mutex = os_mutex_create();
 
 	for (i = 0; i < OS_FILE_N_SEEK_MUTEXES; i++) {
-		os_file_seek_mutexes[i] = os_mutex_create(NULL);
+		os_file_seek_mutexes[i] = os_mutex_create();
 	}
 }
 
@@ -3217,7 +3226,7 @@ os_aio_array_create(
 
 	array = ut_malloc(sizeof(os_aio_array_t));
 
-	array->mutex		= os_mutex_create(NULL);
+	array->mutex		= os_mutex_create();
 	array->not_full		= os_event_create(NULL);
 	array->is_empty		= os_event_create(NULL);
 
@@ -3229,7 +3238,7 @@ os_aio_array_create(
 	array->cur_seg		= 0;
 	array->slots		= ut_malloc(n * sizeof(os_aio_slot_t));
 #ifdef __WIN__
-	array->native_events	= ut_malloc(n * sizeof(os_native_event_t));
+	array->handles		= ut_malloc(n * sizeof(HANDLE));
 #endif
 
 #if defined(LINUX_NATIVE_AIO)
@@ -3273,13 +3282,13 @@ os_aio_array_create(
 		slot->pos = i;
 		slot->reserved = FALSE;
 #ifdef WIN_ASYNC_IO
-		slot->event = os_event_create(NULL);
+		slot->handle = CreateEvent(NULL,TRUE, FALSE, NULL);
 
 		over = &(slot->control);
 
-		over->hEvent = slot->event->handle;
+		over->hEvent = slot->handle;
 
-		*((array->native_events) + i) = over->hEvent;
+		*((array->handles) + i) = over->hEvent;
 
 #elif defined(LINUX_NATIVE_AIO)
 
@@ -3305,12 +3314,12 @@ os_aio_array_free(
 
 	for (i = 0; i < array->n_slots; i++) {
 		os_aio_slot_t*	slot = os_aio_array_get_nth_slot(array, i);
-		os_event_free(slot->event);
+		CloseHandle(slot->handle);
 	}
 #endif /* WIN_ASYNC_IO */
 
 #ifdef __WIN__
-	ut_free(array->native_events);
+	ut_free(array->handles);
 #endif /* __WIN__ */
 	os_mutex_free(array->mutex);
 	os_event_free(array->not_full);
@@ -3463,7 +3472,7 @@ os_aio_array_wake_win_aio_at_shutdown(
 
 	for (i = 0; i < array->n_slots; i++) {
 
-		os_event_set((array->slots + i)->event);
+		SetEvent((array->slots + i)->handle);
 	}
 }
 #endif
@@ -3702,7 +3711,7 @@ os_aio_array_reserve_slot(
 	control = &(slot->control);
 	control->Offset = (DWORD)offset;
 	control->OffsetHigh = (DWORD)offset_high;
-	os_event_reset(slot->event);
+	ResetEvent(slot->handle);
 
 #elif defined(LINUX_NATIVE_AIO)
 
@@ -3774,7 +3783,7 @@ os_aio_array_free_slot(
 
 #ifdef WIN_ASYNC_IO
 
-	os_event_reset(slot->event);
+	ResetEvent(slot->handle);
 
 #elif defined(LINUX_NATIVE_AIO)
 
@@ -4208,13 +4217,20 @@ os_aio_windows_handle(
 	n = array->n_slots / array->n_segments;
 
 	if (array == os_aio_sync_array) {
-		os_event_wait(os_aio_array_get_nth_slot(array, pos)->event);
+		WaitForSingleObject(
+			os_aio_array_get_nth_slot(array, pos)->handle,
+			INFINITE);
 		i = pos;
 	} else {
 		srv_set_io_thread_op_info(orig_seg, "wait Windows aio");
-		i = os_event_wait_multiple(n,
-					   (array->native_events)
-					   + segment * n);
+		i = WaitForMultipleObjects((DWORD) n,
+					   array->handles + segment * n,
+					   FALSE,
+					   INFINITE);
+	}
+
+	if (srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS) {
+		os_thread_exit(NULL);
 	}
 
 	os_mutex_enter(array->mutex);

storage/innobase/os/os0thread.c

@@ -252,7 +252,7 @@ os_thread_yield(void)
 /*=================*/
 {
 #if defined(__WIN__)
-	Sleep(0);
+	SwitchToThread();
 #elif (defined(HAVE_SCHED_YIELD) && defined(HAVE_SCHED_H))
 	sched_yield();
 #elif defined(HAVE_PTHREAD_YIELD_ZERO_ARG)

storage/innobase/srv/srv0srv.c

@@ -142,6 +142,21 @@ use simulated aio we build below with threads.
 Currently we support native aio on windows and linux */
 UNIV_INTERN my_bool	srv_use_native_aio = TRUE;
 
+#ifdef __WIN__
+/* Windows native condition variables. We use runtime loading / function
+pointers, because they are not available on Windows Server 2003 and
+Windows XP/2000.
+
+We use condition for events on Windows if possible, even if os_event
+resembles Windows kernel event object well API-wise. The reason is
+performance, kernel objects are heavyweights and WaitForSingleObject() is a
+performance killer causing calling thread to context switch. Besides, Innodb
+is preallocating large number (often millions) of os_events. With kernel event
+objects it takes a big chunk out of non-paged pool, which is better suited
+for tasks like IO than for storing idle event objects. */
+UNIV_INTERN ibool	srv_use_native_conditions = FALSE;
+#endif /* __WIN__ */
+
 UNIV_INTERN ulint	srv_n_data_files = 0;
 UNIV_INTERN char**	srv_data_file_names = NULL;
 /* size in database pages */

storage/innobase/srv/srv0start.c

@@ -1160,9 +1160,17 @@ innobase_start_or_create_for_mysql(void)
 
 		srv_use_native_aio = FALSE;
 		break;
+
+	case OS_WIN2000:
+	case OS_WINXP:
+		/* On 2000 and XP, async IO is available. */
+		srv_use_native_aio = TRUE;
+		break;
+
 	default:
-		/* On Win 2000 and XP use async i/o */
+		/* Vista and later have both async IO and condition variables */
 		srv_use_native_aio = TRUE;
+		srv_use_native_conditions = TRUE;
 		break;
 	}
 

storage/innobase/sync/sync0arr.c

@@ -250,7 +250,7 @@ sync_array_create(
 
 	/* Then create the mutex to protect the wait array complex */
 	if (protection == SYNC_ARRAY_OS_MUTEX) {
-		arr->os_mutex = os_mutex_create(NULL);
+		arr->os_mutex = os_mutex_create();
 	} else if (protection == SYNC_ARRAY_MUTEX) {
 		mutex_create(syn_arr_mutex_key,
 			     &arr->mutex, SYNC_NO_ORDER_CHECK);