branches/5.1: Port of r2267

This is a combination of changes that forward port the scalability fix applied to 5.0
through r1001.
It reverts changes r149 and r122 (these were 5.1 specific changes made in lieu of
scalability fix of 5.0)
Then it applies r1001 to 5.0 which is the original scalability fix.
Finally it applies r2082 which fixes an issue with the original fix.

Reviewed by: Heikki

include/os0sync.h

@@ -112,9 +112,13 @@ os_event_set(
 	os_event_t	event);	/* in: event to set */
 /**************************************************************
 Resets an event semaphore to the nonsignaled state. Waiting threads will
-stop to wait for the event. */
+stop to wait for the event.
+The return value should be passed to os_even_wait_low() if it is desired
+that this thread should not wait in case of an intervening call to
+os_event_set() between this os_event_reset() and the
+os_event_wait_low() call. See comments for os_event_wait_low(). */
 
-void
+ib_longlong
 os_event_reset(
 /*===========*/
 	os_event_t	event);	/* in: event to reset */
@@ -125,16 +129,38 @@ void
 os_event_free(
 /*==========*/
 	os_event_t	event);	/* in: event to free */
+
 /**************************************************************
 Waits for an event object until it is in the signaled state. If
 srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS this also exits the
 waiting thread when the event becomes signaled (or immediately if the
-event is already in the signaled state). */
+event is already in the signaled state).
+
+Typically, if the event has been signalled after the os_event_reset()
+we'll return immediately because event->is_set == TRUE.
+There are, however, situations (e.g.: sync_array code) where we may
+lose this information. For example:
+
+thread A calls os_event_reset()
+thread B calls os_event_set()   [event->is_set == TRUE]
+thread C calls os_event_reset() [event->is_set == FALSE]
+thread A calls os_event_wait()  [infinite wait!]
+thread C calls os_event_wait()  [infinite wait!]
+
+Where such a scenario is possible, to avoid infinite wait, the
+value returned by os_event_reset() should be passed in as
+reset_sig_count. */
+
+#define os_event_wait(event) os_event_wait_low((event), 0)
 
 void
-os_event_wait(
-/*==========*/
-	os_event_t	event);	/* in: event to wait */
+os_event_wait_low(
+/*==============*/
+	os_event_t	event,		/* in: event to wait */
+	ib_longlong	reset_sig_count);/* in: zero or the value
+					returned by previous call of
+					os_event_reset(). */
+
 /**************************************************************
 Waits for an event object until it is in the signaled state or
 a timeout is exceeded. In Unix the timeout is always infinite. */

include/sync0arr.h

@@ -66,26 +66,21 @@ sync_array_wait_event(
 	sync_array_t*	arr,	/* in: wait array */
 	ulint		index);	 /* in: index of the reserved cell */
 /**********************************************************************
-Frees the cell safely by reserving the sync array mutex and decrementing
-n_reserved if necessary. Should only be called from mutex_spin_wait. */
+Frees the cell. NOTE! sync_array_wait_event frees the cell
+automatically! */
 
 void
-sync_array_free_cell_protected(
-/*===========================*/
+sync_array_free_cell(
+/*=================*/
 	sync_array_t*	arr,	/* in: wait array */
 	ulint		index);	/* in: index of the cell in array */
 /**************************************************************************
-Looks for the cells in the wait array which refer
-to the wait object specified,
-and sets their corresponding events to the signaled state. In this
-way releases the threads waiting for the object to contend for the object.
-It is possible that no such cell is found, in which case does nothing. */
+Note that one of the wait objects was signalled. */
 
 void
-sync_array_signal_object(
-/*=====================*/
-	sync_array_t*	arr,	/* in: wait array */
-	void*		object);/* in: wait object */
+sync_array_object_signalled(
+/*========================*/
+	sync_array_t*	arr);	/* in: wait array */
 /**************************************************************************
 If the wakeup algorithm does not work perfectly at semaphore relases,
 this function will do the waking (see the comment in mutex_exit). This

include/sync0rw.h

@@ -421,6 +421,18 @@ blocked by readers, a writer may queue for the lock by setting the writer
 field. Then no new readers are allowed in. */
 
 struct rw_lock_struct {
+	os_event_t	event;	/* Used by sync0arr.c for thread queueing */
+
+#ifdef __WIN__
+	os_event_t	wait_ex_event;	/* This windows specific event is
+				used by the thread which has set the
+				lock state to RW_LOCK_WAIT_EX. The
+				rw_lock design guarantees that this
+				thread will be the next one to proceed
+				once the current the event gets
+				signalled. See LEMMA 2 in sync0sync.c */
+#endif
+
 	ulint	reader_count;	/* Number of readers who have locked this
 				lock in the shared mode */
 	ulint	writer;		/* This field is set to RW_LOCK_EX if there

include/sync0rw.ic

@@ -381,7 +381,11 @@ rw_lock_s_unlock_func(
 	mutex_exit(mutex);
 
 	if (UNIV_UNLIKELY(sg)) {
-		sync_array_signal_object(sync_primary_wait_array, lock);
+#ifdef __WIN__
+		os_event_set(lock->wait_ex_event);
+#endif
+		os_event_set(lock->event);
+		sync_array_object_signalled(sync_primary_wait_array);
 	}
 
 	ut_ad(rw_lock_validate(lock));
@@ -461,7 +465,11 @@ rw_lock_x_unlock_func(
 	mutex_exit(&(lock->mutex));
 
 	if (UNIV_UNLIKELY(sg)) {
-		sync_array_signal_object(sync_primary_wait_array, lock);
+#ifdef __WIN__
+		os_event_set(lock->wait_ex_event);
+#endif
+		os_event_set(lock->event);
+		sync_array_object_signalled(sync_primary_wait_array);
 	}
 
 	ut_ad(rw_lock_validate(lock));

include/sync0sync.h

@@ -470,6 +470,7 @@ Do not use its fields directly! The structure used in the spin lock
 implementation of a mutual exclusion semaphore. */
 
 struct mutex_struct {
+	os_event_t	event;	/* Used by sync0arr.c for the wait queue */
 	ulint	lock_word;	/* This ulint is the target of the atomic
 				test-and-set instruction in Win32 */
 #if !defined(_WIN32) || !defined(UNIV_CAN_USE_X86_ASSEMBLER)

include/sync0sync.ic

@@ -211,7 +211,7 @@ mutex_exit(
 	perform the read first, which could leave a waiting
 	thread hanging indefinitely.
 
-	Our current solution call every 10 seconds
+	Our current solution call every second
 	sync_arr_wake_threads_if_sema_free()
 	to wake up possible hanging threads if
 	they are missed in mutex_signal_object. */

os/os0sync.c

@@ -21,6 +21,7 @@ Created 9/6/1995 Heikki Tuuri
 
 /* Type definition for an operating system mutex struct */
 struct os_mutex_struct{
+	os_event_t	event;	/* Used by sync0arr.c for queing threads */
 	void*		handle;	/* OS handle to mutex */
 	ulint		count;	/* we use this counter to check
 				that the same thread does not
@@ -35,6 +36,7 @@ struct os_mutex_struct{
 /* Mutex protecting counts and the lists of OS mutexes and events */
 os_mutex_t	os_sync_mutex;
 ibool		os_sync_mutex_inited	= FALSE;
+ibool		os_sync_free_called	= FALSE;
 
 /* This is incremented by 1 in os_thread_create and decremented by 1 in
 os_thread_exit */
@@ -50,6 +52,10 @@ ulint	os_event_count		= 0;
 ulint	os_mutex_count		= 0;
 ulint	os_fast_mutex_count	= 0;
 
+/* Because a mutex is embedded inside an event and there is an
+event embedded inside a mutex, on free, this generates a recursive call.
+This version of the free event function doesn't acquire the global lock */
+static void os_event_free_internal(os_event_t	event);
 
 /*************************************************************
 Initializes global event and OS 'slow' mutex lists. */
@@ -76,6 +82,7 @@ os_sync_free(void)
 	os_event_t	event;
 	os_mutex_t	mutex;
 
+	os_sync_free_called = TRUE;
 	event = UT_LIST_GET_FIRST(os_event_list);
 
 	while (event) {
@@ -99,6 +106,7 @@ os_sync_free(void)
 
 		mutex = UT_LIST_GET_FIRST(os_mutex_list);
 	}
+	os_sync_free_called = FALSE;
 }
 
 /*************************************************************
@@ -144,17 +152,31 @@ os_event_create(
 	ut_a(0 == pthread_cond_init(&(event->cond_var), NULL));
 #endif
 	event->is_set = FALSE;
-	event->signal_count = 0;
+
+	/* We return this value in os_event_reset(), which can then be
+	be used to pass to the os_event_wait_low(). The value of zero
+	is reserved in os_event_wait_low() for the case when the
+	caller does not want to pass any signal_count value. To
+	distinguish between the two cases we initialize signal_count
+	to 1 here. */
+	event->signal_count = 1;
 #endif /* __WIN__ */
 
-	/* Put to the list of events */
+	/* The os_sync_mutex can be NULL because during startup an event
+	can be created [ because it's embedded in the mutex/rwlock ] before
+	this module has been initialized */
+	if (os_sync_mutex != NULL) {
 		os_mutex_enter(os_sync_mutex);
+	}
 
+	/* Put to the list of events */
 	UT_LIST_ADD_FIRST(os_event_list, os_event_list, event);
 
 	os_event_count++;
 
+	if (os_sync_mutex != NULL) {
 		os_mutex_exit(os_sync_mutex);
+	}
 
 	return(event);
 }
@@ -231,13 +253,20 @@ os_event_set(
 
 /**************************************************************
 Resets an event semaphore to the nonsignaled state. Waiting threads will
-stop to wait for the event. */
+stop to wait for the event.
+The return value should be passed to os_even_wait_low() if it is desired
+that this thread should not wait in case of an intervening call to
+os_event_set() between this os_event_reset() and the
+os_event_wait_low() call. See comments for os_event_wait_low(). */
 
-void
+ib_longlong
 os_event_reset(
 /*===========*/
+				/* out: current signal_count. */
 	os_event_t	event)	/* in: event to reset */
 {
+	ib_longlong	ret = 0;
+
 #ifdef __WIN__
 	ut_a(event);
 
@@ -252,9 +281,40 @@ os_event_reset(
 	} else {
 		event->is_set = FALSE;
 	}
+	ret = event->signal_count;
 
 	os_fast_mutex_unlock(&(event->os_mutex));
 #endif
+	return(ret);
+}
+
+/**************************************************************
+Frees an event object, without acquiring the global lock. */
+static
+void
+os_event_free_internal(
+/*===================*/
+	os_event_t	event)	/* in: event to free */
+{
+#ifdef __WIN__
+	ut_a(event);
+
+	ut_a(CloseHandle(event->handle));
+#else
+	ut_a(event);
+
+	/* This is to avoid freeing the mutex twice */
+	os_fast_mutex_free(&(event->os_mutex));
+
+	ut_a(0 == pthread_cond_destroy(&(event->cond_var)));
+#endif
+	/* Remove from the list of events */
+
+	UT_LIST_REMOVE(os_event_list, os_event_list, event);
+
+	os_event_count--;
+
+	ut_free(event);
 }
 
 /**************************************************************
@@ -293,18 +353,38 @@ os_event_free(
 Waits for an event object until it is in the signaled state. If
 srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS this also exits the
 waiting thread when the event becomes signaled (or immediately if the
-event is already in the signaled state). */
+event is already in the signaled state).
+
+Typically, if the event has been signalled after the os_event_reset()
+we'll return immediately because event->is_set == TRUE.
+There are, however, situations (e.g.: sync_array code) where we may
+lose this information. For example:
+
+thread A calls os_event_reset()
+thread B calls os_event_set()   [event->is_set == TRUE]
+thread C calls os_event_reset() [event->is_set == FALSE]
+thread A calls os_event_wait()  [infinite wait!]
+thread C calls os_event_wait()  [infinite wait!]
+
+Where such a scenario is possible, to avoid infinite wait, the
+value returned by os_event_reset() should be passed in as
+reset_sig_count. */
 
 void
-os_event_wait(
-/*==========*/
-	os_event_t	event)	/* in: event to wait */
+os_event_wait_low(
+/*==============*/
+	os_event_t	event,		/* in: event to wait */
+	ib_longlong	reset_sig_count)/* in: zero or the value
+					returned by previous call of
+					os_event_reset(). */
 {
 #ifdef __WIN__
 	DWORD	err;
 
 	ut_a(event);
 
+	UT_NOT_USED(reset_sig_count);
+
 	/* Specify an infinite time limit for waiting */
 	err = WaitForSingleObject(event->handle, INFINITE);
 
@@ -318,7 +398,11 @@ os_event_wait(
 
 	os_fast_mutex_lock(&(event->os_mutex));
 
+	if (reset_sig_count) {
+		old_signal_count = reset_sig_count;
+	} else {
 		old_signal_count = event->signal_count;
+	}
 
 	for (;;) {
 		if (event->is_set == TRUE
@@ -458,6 +542,7 @@ os_mutex_create(
 
 	mutex_str->handle = mutex;
 	mutex_str->count = 0;
+	mutex_str->event = os_event_create(NULL);
 
 	if (os_sync_mutex_inited) {
 		/* When creating os_sync_mutex itself we cannot reserve it */
@@ -534,6 +619,10 @@ os_mutex_free(
 {
 	ut_a(mutex);
 
+	if (!os_sync_free_called) {
+		os_event_free_internal(mutex->event);
+	}
+
 	if (os_sync_mutex_inited) {
 		os_mutex_enter(os_sync_mutex);
 	}

srv/srv0srv.c

@@ -1904,12 +1904,6 @@ loop:
 
 	os_thread_sleep(1000000);
 
-	/* In case mutex_exit is not a memory barrier, it is
-	theoretically possible some threads are left waiting though
-	the semaphore is already released. Wake up those threads: */
-
-	sync_arr_wake_threads_if_sema_free();
-
 	current_time = time(NULL);
 
 	time_elapsed = difftime(current_time, last_monitor_time);
@@ -2106,9 +2100,15 @@ loop:
 		srv_refresh_innodb_monitor_stats();
 	}
 
+	/* In case mutex_exit is not a memory barrier, it is
+	theoretically possible some threads are left waiting though
+	the semaphore is already released. Wake up those threads: */
+	
+	sync_arr_wake_threads_if_sema_free();
+
 	if (sync_array_print_long_waits()) {
 		fatal_cnt++;
-		if (fatal_cnt > 5) {
+		if (fatal_cnt > 10) {
 
 			fprintf(stderr,
 				"InnoDB: Error: semaphore wait has lasted"
@@ -2128,7 +2128,7 @@ loop:
 
 	fflush(stderr);
 
-	os_thread_sleep(2000000);
+	os_thread_sleep(1000000);
 
 	if (srv_shutdown_state < SRV_SHUTDOWN_CLEANUP) {
 

sync/sync0rw.c

@@ -151,6 +151,11 @@ rw_lock_create_func(
 	lock->last_x_file_name = "not yet reserved";
 	lock->last_s_line = 0;
 	lock->last_x_line = 0;
+	lock->event = os_event_create(NULL);
+
+#ifdef __WIN__
+	lock->wait_ex_event = os_event_create(NULL);
+#endif
 
 	mutex_enter(&rw_lock_list_mutex);
 
@@ -184,6 +189,11 @@ rw_lock_free(
 	mutex_free(rw_lock_get_mutex(lock));
 
 	mutex_enter(&rw_lock_list_mutex);
+	os_event_free(lock->event);
+
+#ifdef __WIN__
+	os_event_free(lock->wait_ex_event);
+#endif
 
 	if (UT_LIST_GET_PREV(list, lock)) {
 		ut_a(UT_LIST_GET_PREV(list, lock)->magic_n == RW_LOCK_MAGIC_N);
@@ -544,7 +554,15 @@ lock_loop:
 	rw_x_system_call_count++;
 
 	sync_array_reserve_cell(sync_primary_wait_array,
-				lock, RW_LOCK_EX,
+				lock,
+#ifdef __WIN__
+				/* On windows RW_LOCK_WAIT_EX signifies
+				that this thread should wait on the
+				special wait_ex_event. */
+				(state == RW_LOCK_WAIT_EX)
+				 ? RW_LOCK_WAIT_EX :
+#endif
+				RW_LOCK_EX,
 				file_name, line,
 				&index);
 

sync/sync0sync.c

@@ -95,17 +95,47 @@ have happened that the thread which was holding the mutex has just released
 it and did not see the waiters byte set to 1, a case which would lead the
 other thread to an infinite wait.
 
-LEMMA 1: After a thread resets the event of the cell it reserves for waiting
-========
-for a mutex, some thread will eventually call sync_array_signal_object with
-the mutex as an argument. Thus no infinite wait is possible.
+LEMMA 1: After a thread resets the event of a mutex (or rw_lock), some
+=======
+thread will eventually call os_event_set() on that particular event.
+Thus no infinite wait is possible in this case.
 
 Proof:	After making the reservation the thread sets the waiters field in the
 mutex to 1. Then it checks that the mutex is still reserved by some thread,
 or it reserves the mutex for itself. In any case, some thread (which may be
 also some earlier thread, not necessarily the one currently holding the mutex)
 will set the waiters field to 0 in mutex_exit, and then call
-sync_array_signal_object with the mutex as an argument.
+os_event_set() with the mutex as an argument.
+Q.E.D.
+
+LEMMA 2: If an os_event_set() call is made after some thread has called
+=======
+the os_event_reset() and before it starts wait on that event, the call
+will not be lost to the second thread. This is true even if there is an
+intervening call to os_event_reset() by another thread.
+Thus no infinite wait is possible in this case.
+
+Proof (non-windows platforms): os_event_reset() returns a monotonically
+increasing value of signal_count. This value is increased at every
+call of os_event_set() If thread A has called os_event_reset() followed
+by thread B calling os_event_set() and then some other thread C calling
+os_event_reset(), the is_set flag of the event will be set to FALSE;
+but now if thread A calls os_event_wait_low() with the signal_count
+value returned from the earlier call of os_event_reset(), it will
+return immediately without waiting.
+Q.E.D.
+
+Proof (windows): If there is a writer thread which is forced to wait for
+the lock, it may be able to set the state of rw_lock to RW_LOCK_WAIT_EX
+The design of rw_lock ensures that there is one and only one thread
+that is able to change the state to RW_LOCK_WAIT_EX and this thread is
+guaranteed to acquire the lock after it is released by the current
+holders and before any other waiter gets the lock.
+On windows this thread waits on a separate event i.e.: wait_ex_event.
+Since only one thread can wait on this event there is no chance
+of this event getting reset before the writer starts wait on it.
+Therefore, this thread is guaranteed to catch the os_set_event()
+signalled unconditionally at the release of the lock.
 Q.E.D. */
 
 /* The number of system calls made in this module. Intended for performance
@@ -217,6 +247,7 @@ mutex_create_func(
 	os_fast_mutex_init(&(mutex->os_fast_mutex));
 	mutex->lock_word = 0;
 #endif
+	mutex->event = os_event_create(NULL);
 	mutex_set_waiters(mutex, 0);
 #ifdef UNIV_DEBUG
 	mutex->magic_n = MUTEX_MAGIC_N;
@@ -300,6 +331,8 @@ mutex_free(
 		mutex_exit(&mutex_list_mutex);
 	}
 
+	os_event_free(mutex->event);
+
 #if !defined(_WIN32) || !defined(UNIV_CAN_USE_X86_ASSEMBLER)
 	os_fast_mutex_free(&(mutex->os_fast_mutex));
 #endif
@@ -509,8 +542,7 @@ spin_loop:
 		if (mutex_test_and_set(mutex) == 0) {
 			/* Succeeded! Free the reserved wait cell */
 
-			sync_array_free_cell_protected(sync_primary_wait_array,
-						       index);
+			sync_array_free_cell(sync_primary_wait_array, index);
 
 			ut_d(mutex->thread_id = os_thread_get_curr_id());
 #ifdef UNIV_SYNC_DEBUG
@@ -591,8 +623,8 @@ mutex_signal_object(
 
 	/* The memory order of resetting the waiters field and
 	signaling the object is important. See LEMMA 1 above. */
-
-	sync_array_signal_object(sync_primary_wait_array, mutex);
+	os_event_set(mutex->event);
+	sync_array_object_signalled(sync_primary_wait_array);
 }
 
 #ifdef UNIV_SYNC_DEBUG
@@ -1130,6 +1162,7 @@ sync_thread_add_level(
 		break;
 	case SYNC_TREE_NODE:
 		ut_a(sync_thread_levels_contain(array, SYNC_INDEX_TREE)
+		     || sync_thread_levels_contain(array, SYNC_DICT_OPERATION)
 		     || sync_thread_levels_g(array, SYNC_TREE_NODE - 1));
 		break;
 	case SYNC_TREE_NODE_NEW: