Merge with 4.0.13

15 files changed, 430 insertions, 133 deletions

diff --git a/innobase/include/os0file.h b/innobase/include/os0file.h
index a7624a90d5e..86f27a2d3eb 100644
--- a/innobase/include/os0file.h
+++ b/innobase/include/os0file.h
@@ -301,6 +301,13 @@ os_aio(
 				are ignored */
 	void*		message2);
 /****************************************************************************
+Wakes up all async i/o threads so that they know to exit themselves in
+shutdown. */
+
+void
+os_aio_wake_all_threads_at_shutdown(void);
+/*=====================================*/
+/****************************************************************************
 Waits until there are no pending writes in os_aio_write_array. There can
 be other, synchronous, pending writes. */
 
diff --git a/innobase/include/os0sync.h b/innobase/include/os0sync.h
index b2d613c4619..bad8e6e120a 100644
--- a/innobase/include/os0sync.h
+++ b/innobase/include/os0sync.h
@@ -10,15 +10,16 @@ Created 9/6/1995 Heikki Tuuri
 #define os0sync_h
 
 #include "univ.i"
+#include "ut0lst.h"
 
 #ifdef __WIN__
-
 #define os_fast_mutex_t CRITICAL_SECTION
-typedef void*		os_event_t;
-
+typedef HANDLE          os_event_t;
 #else
-
 typedef pthread_mutex_t	os_fast_mutex_t;
+
+typedef struct os_event_struct os_event_struct_t;
+typedef os_event_struct_t*     os_event_t;
 struct os_event_struct {
 	os_fast_mutex_t	os_mutex;	/* this mutex protects the next
 					fields */
@@ -26,9 +27,9 @@ struct os_event_struct {
 					not reserved */
 	pthread_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 */
 };
-typedef struct os_event_struct os_event_struct_t;
-typedef os_event_struct_t*     os_event_t;
 #endif
 
 typedef struct os_mutex_struct	os_mutex_str_t;
@@ -38,6 +39,29 @@ typedef os_mutex_str_t*		os_mutex_t;
 
 #define OS_SYNC_TIME_EXCEEDED	1
 
+/* Mutex protecting the thread count and event and OS 'slow' mutex lists */
+extern os_mutex_t	os_sync_mutex;
+
+/* This is incremented by 1 in os_thread_create and decremented by 1 in
+os_thread_exit */
+extern ulint		os_thread_count;
+
+/* The following are approximate counters for debugging in Unix */
+extern ulint		os_event_count;
+extern ulint		os_mutex_count;
+
+/*************************************************************
+Initializes global event and OS 'slow' mutex lists. */
+
+void
+os_sync_init(void);
+/*==============*/
+/*************************************************************
+Frees created events (not in Windows) and OS 'slow' mutexes. */
+
+void
+os_sync_free(void);
+/*==============*/
 /*************************************************************
 Creates an event semaphore, i.e., a semaphore which may
 just have two states: signaled and nonsignaled.
@@ -85,7 +109,10 @@ os_event_free(
 /*==========*/
 	os_event_t	event);	/* in: event to free */
 /**************************************************************
-Waits for an event object until it is in the signaled state. */
+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). */
 
 void
 os_event_wait(
diff --git a/innobase/include/os0thread.h b/innobase/include/os0thread.h
index 629cfef23a8..92187f315c2 100644
--- a/innobase/include/os0thread.h
+++ b/innobase/include/os0thread.h
@@ -41,7 +41,6 @@ typedef os_thread_t          	os_thread_id_t;	/* In Unix we use the thread
 						the thread */
 #endif
 
-
 /* Define a function pointer type to use in a typecast */
 typedef void* (*os_posix_f_t) (void*);
 
@@ -83,12 +82,13 @@ os_thread_create(
 	os_thread_id_t*		thread_id);	/* out: id of the created
 						thread */
 /*********************************************************************
-A thread calling this function ends its execution. */
+Exits the current thread. */
 
 void
 os_thread_exit(
 /*===========*/
-	ulint	code);	/* in: exit code */
+	void*	exit_value);	/* in: exit value; in Windows this void*
+				is cast as a DWORD */
 /*********************************************************************
 Returns the thread identifier of current thread. */
 
@@ -144,7 +144,6 @@ ulint
 os_thread_get_last_error(void);
 /*==========================*/
 
-
 #ifndef UNIV_NONINL
 #include "os0thread.ic"
 #endif
diff --git a/innobase/include/srv0srv.h b/innobase/include/srv0srv.h
index 121e9c44a24..8fd0fc2dd6d 100644
--- a/innobase/include/srv0srv.h
+++ b/innobase/include/srv0srv.h
@@ -212,6 +212,12 @@ void
 srv_init(void);
 /*==========*/
 /*************************************************************************
+Frees the OS fast mutex created in srv_init(). */
+
+void
+srv_free(void);
+/*==========*/
+/*************************************************************************
 Initializes the synchronization primitives, memory system, and the thread
 local storage. */
 
diff --git a/innobase/include/srv0start.h b/innobase/include/srv0start.h
index aec3ebfeea9..8d2c3fa12c5 100644
--- a/innobase/include/srv0start.h
+++ b/innobase/include/srv0start.h
@@ -86,11 +86,12 @@ extern	ibool	srv_startup_is_before_trx_rollback_phase;
 extern	ibool	srv_is_being_shut_down;
 
 /* At a shutdown the value first climbs from 0 to SRV_SHUTDOWN_CLEANUP
-and then to SRV_SHUTDOWN_LAST_PHASE */
+and then to SRV_SHUTDOWN_LAST_PHASE, and so on */
 
 extern  ulint   srv_shutdown_state;
 
-#define SRV_SHUTDOWN_CLEANUP	1
-#define SRV_SHUTDOWN_LAST_PHASE	2
+#define SRV_SHUTDOWN_CLEANUP	   1
+#define SRV_SHUTDOWN_LAST_PHASE	   2
+#define SRV_SHUTDOWN_EXIT_THREADS  3
 
 #endif
diff --git a/innobase/include/univ.i b/innobase/include/univ.i
index e29f3ec92e1..4854e5a7b78 100644
--- a/innobase/include/univ.i
+++ b/innobase/include/univ.i
@@ -187,7 +187,11 @@ management to ensure correct alignment for doubles etc. */
 /* Another basic type we use is unsigned long integer which is intended to be
 equal to the word size of the machine. */
 
+#ifdef _WIN64
+typedef unsigned __int64	ulint;
+#else
 typedef unsigned long int	ulint;
+#endif
 
 typedef long int		lint;
 
diff --git a/innobase/log/log0log.c b/innobase/log/log0log.c
index e717d897e41..6892e6fc6c1 100644
--- a/innobase/log/log0log.c
+++ b/innobase/log/log0log.c
@@ -375,7 +375,7 @@ log_pad_current_log_block(void)
 	log_close();
 	log_release();
 
-	ut_a((ut_dulint_get_low(lsn) % OS_FILE_LOG_BLOCK_SIZE)
+	ut_ad((ut_dulint_get_low(lsn) % OS_FILE_LOG_BLOCK_SIZE)
 						== LOG_BLOCK_HDR_SIZE);
 }
 
@@ -998,6 +998,8 @@ log_group_file_header_flush(
 {
 	byte*	buf;
 	ulint	dest_offset;
+	
+	UT_NOT_USED(type);
 
 	ut_ad(mutex_own(&(log_sys->mutex)));
 
@@ -1068,8 +1070,8 @@ log_group_write_buf(
 	ulint	i;
 	
 	ut_ad(mutex_own(&(log_sys->mutex)));
-	ut_a(len % OS_FILE_LOG_BLOCK_SIZE == 0);
-	ut_a(ut_dulint_get_low(start_lsn) % OS_FILE_LOG_BLOCK_SIZE == 0);
+	ut_ad(len % OS_FILE_LOG_BLOCK_SIZE == 0);
+	ut_ad(ut_dulint_get_low(start_lsn) % OS_FILE_LOG_BLOCK_SIZE == 0);
 
 	if (new_data_offset == 0) {
 		write_header = TRUE;
@@ -2901,10 +2903,9 @@ logs_empty_and_mark_files_at_shutdown(void)
 	dulint	lsn;
 	ulint	arch_log_no;
 
-	if (srv_print_verbose_log)
-	{
-	  ut_print_timestamp(stderr);
-	  fprintf(stderr, "  InnoDB: Starting shutdown...\n");
+	if (srv_print_verbose_log) {
+	        ut_print_timestamp(stderr);
+	        fprintf(stderr, "  InnoDB: Starting shutdown...\n");
 	}
 	/* Wait until the master thread and all other operations are idle: our
 	algorithm only works if the server is idle at shutdown */
@@ -3006,15 +3007,17 @@ loop:
 		goto loop;
 	}
 
+	/* Make some checks that the server really is quiet */
+	ut_a(buf_all_freed());
+	ut_a(0 == ut_dulint_cmp(lsn, log_sys->lsn));
+
 	fil_write_flushed_lsn_to_data_files(lsn, arch_log_no);	
 
 	fil_flush_file_spaces(FIL_TABLESPACE);
 
-	if (srv_print_verbose_log)
-	{
-	  ut_print_timestamp(stderr);
-	  fprintf(stderr, "  InnoDB: Shutdown completed\n");
-	}
+	/* Make some checks that the server really is quiet */
+	ut_a(buf_all_freed());
+	ut_a(0 == ut_dulint_cmp(lsn, log_sys->lsn));
 }
 
 /**********************************************************
diff --git a/innobase/os/os0file.c b/innobase/os/os0file.c
index 53224eb59c5..9da5c6a47a5 100644
--- a/innobase/os/os0file.c
+++ b/innobase/os/os0file.c
@@ -295,7 +295,8 @@ os_file_handle_error(
 				/* out: TRUE if we should retry the
 				operation */
 	os_file_t	file,	/* in: file pointer */
-	char*		name)	/* in: name of a file or NULL */
+	char*		name,	/* in: name of a file or NULL */
+	const char*	operation) /* in: type of operation */
 {
 	ulint	err;
 
@@ -337,7 +338,8 @@ os_file_handle_error(
 	        if (name) {
 	                fprintf(stderr, "InnoDB: File name %s\n", name);
 	        }
-	  
+		fprintf(stderr, "InnoDB: system call %s\n", operation);
+
 		fprintf(stderr, "InnoDB: Cannot continue operation.\n");
 
 		fflush(stderr);
@@ -419,7 +421,9 @@ try_again:
 	if (file == INVALID_HANDLE_VALUE) {
 		*success = FALSE;
 
-		retry = os_file_handle_error(file, name);
+		retry = os_file_handle_error(file, name,
+					     create_mode == OS_FILE_OPEN ?
+					     "open" : "create");
 
 		if (retry) {
 			goto try_again;
@@ -460,7 +464,10 @@ try_again:
 	if (file == -1) {
 		*success = FALSE;
 
-		retry = os_file_handle_error(file, name);
+		retry = os_file_handle_error(file, name,
+					     create_mode == OS_FILE_OPEN ?
+					     "open" : "create");
+
 
 		if (retry) {
 			goto try_again;
@@ -568,7 +575,9 @@ try_again:
 	if (file == INVALID_HANDLE_VALUE) {
 		*success = FALSE;
 
-		retry = os_file_handle_error(file, name);
+		retry = os_file_handle_error(file, name,
+					     create_mode == OS_FILE_OPEN ?
+					     "open" : "create");
 
 		if (retry) {
 			goto try_again;
@@ -615,7 +624,9 @@ try_again:
 	if (file == -1) {
 		*success = FALSE;
 
-		retry = os_file_handle_error(file, name);
+		retry = os_file_handle_error(file, name,
+					     create_mode == OS_FILE_OPEN ?
+					     "open" : "create");
 
 		if (retry) {
 			goto try_again;
@@ -649,7 +660,7 @@ os_file_close(
 		return(TRUE);
 	}
 
-	os_file_handle_error(file, NULL);
+	os_file_handle_error(file, NULL, "close");
 	return(FALSE);
 #else
 	int	ret;
@@ -657,7 +668,7 @@ os_file_close(
 	ret = close(file);
 
 	if (ret == -1) {
-		os_file_handle_error(file, NULL);
+		os_file_handle_error(file, NULL, "close");
 		return(FALSE);
 	}
 
@@ -825,7 +836,7 @@ os_file_flush(
 		return(TRUE);
 	}
 
-	os_file_handle_error(file, NULL);
+	os_file_handle_error(file, NULL, "flush");
 
 	/* It is a fatal error if a file flush does not succeed, because then
 	the database can get corrupt on disk */
@@ -858,7 +869,7 @@ os_file_flush(
 	fprintf(stderr,
 		"  InnoDB: Error: the OS said file flush did not succeed\n");
 
-	os_file_handle_error(file, NULL);
+	os_file_handle_error(file, NULL, "flush");
 
 	/* It is a fatal error if a file flush does not succeed, because then
 	the database can get corrupt on disk */
@@ -1099,7 +1110,7 @@ try_again:
 #ifdef __WIN__
 error_handling:
 #endif
-	retry = os_file_handle_error(file, NULL); 
+	retry = os_file_handle_error(file, NULL, "read"); 
 
 	if (retry) {
 		goto try_again;
@@ -1295,7 +1306,6 @@ os_aio_array_create(
 #endif	
 	ut_a(n > 0);
 	ut_a(n_segments > 0);
-	ut_a(n % n_segments == 0);
 
 	array = ut_malloc(sizeof(os_aio_array_t));
 
@@ -1404,6 +1414,50 @@ os_aio_init(
 	pthread_sigmask(SIG_BLOCK, &sigset, NULL); */
 #endif
 }
+
+#ifdef WIN_ASYNC_IO
+/****************************************************************************
+Wakes up all async i/o threads in the array in Windows async i/o at
+shutdown. */
+static
+void
+os_aio_array_wake_win_aio_at_shutdown(
+/*==================================*/
+	os_aio_array_t*	array)	/* in: aio array */
+{
+	ulint	i;
+
+	for (i = 0; i < array->n_slots; i++) {
+
+	        os_event_set(*(array->events + i));
+	}
+}
+#endif
+
+/****************************************************************************
+Wakes up all async i/o threads so that they know to exit themselves in
+shutdown. */
+
+void
+os_aio_wake_all_threads_at_shutdown(void)
+/*=====================================*/
+{
+	ulint	i;
+
+#ifdef WIN_ASYNC_IO
+        /* This code wakes up all ai/o threads in Windows native aio */
+	os_aio_array_wake_win_aio_at_shutdown(os_aio_read_array);
+	os_aio_array_wake_win_aio_at_shutdown(os_aio_write_array);
+	os_aio_array_wake_win_aio_at_shutdown(os_aio_ibuf_array);
+	os_aio_array_wake_win_aio_at_shutdown(os_aio_log_array);
+#endif
+	/* This loop wakes up all simulated ai/o threads */
+
+	for (i = 0; i < os_aio_n_segments; i++) {
+	    	
+		os_event_set(os_aio_segment_wait_events[i]);
+	}	
+}
 				
 /****************************************************************************
 Waits until there are no pending writes in os_aio_write_array. There can
@@ -1971,7 +2025,7 @@ try_again:
 
 	os_aio_array_free_slot(array, slot);
 
-	retry = os_file_handle_error(file, name);
+	retry = os_file_handle_error(file, name, "aio");
 
 	if (retry) {
 
@@ -2070,7 +2124,7 @@ os_aio_windows_handle(
 		         ut_a(TRUE == os_file_flush(slot->file));
 		}
 	} else {
-		os_file_handle_error(slot->file, slot->name);
+		os_file_handle_error(slot->file, slot->name, "aio");
 		
 		ret_val = FALSE;
 	}		  
diff --git a/innobase/os/os0sync.c b/innobase/os/os0sync.c
index 407b280f805..4f322ee82b2 100644
--- a/innobase/os/os0sync.c
+++ b/innobase/os/os0sync.c
@@ -17,6 +17,7 @@ Created 9/6/1995 Heikki Tuuri
 #endif
 
 #include "ut0mem.h"
+#include "srv0start.h"
 
 /* Type definition for an operating system mutex struct */
 struct os_mutex_struct{ 
@@ -26,9 +27,75 @@ struct os_mutex_struct{
 				recursively lock the mutex: we
 				do not assume that the OS mutex
 				supports recursive locking, though
-				NT seems to do that */				
+				NT seems to do that */
+	UT_LIST_NODE_T(os_mutex_str_t) os_mutex_list;
+				/* list of all 'slow' OS mutexes created */
 };
 
+/* Mutex protecting the thread count and the lists of OS mutexes
+and events */
+os_mutex_t	os_sync_mutex;
+ibool		os_sync_mutex_inited	= FALSE;
+
+/* This is incremented by 1 in os_thread_create and decremented by 1 in
+os_thread_exit */
+ulint		os_thread_count		= 0;
+
+/* The list of all events created (not in Windows) */
+UT_LIST_BASE_NODE_T(os_event_struct_t)	os_event_list;
+
+/* The list of all OS 'slow' mutexes */
+UT_LIST_BASE_NODE_T(os_mutex_str_t)	os_mutex_list;
+
+/* The following are approximate counters for debugging in Unix */
+ulint		os_event_count		= 0;
+ulint		os_mutex_count		= 0;
+
+
+/*************************************************************
+Initializes global event and OS 'slow' mutex lists. */
+
+void
+os_sync_init(void)
+/*==============*/
+{
+	UT_LIST_INIT(os_event_list);
+	UT_LIST_INIT(os_mutex_list);
+
+	os_sync_mutex = os_mutex_create(NULL);
+
+	os_sync_mutex_inited = TRUE;
+}
+
+/*************************************************************
+Frees created events (not in Windows) and OS 'slow' mutexes. */
+
+void
+os_sync_free(void)
+/*==============*/
+{
+	os_event_t	event;
+	os_mutex_t	mutex;
+
+	event = UT_LIST_GET_FIRST(os_event_list);
+
+	while (event) {
+
+	      os_event_free(event);
+
+	      event = UT_LIST_GET_FIRST(os_event_list);
+	}
+
+	mutex = UT_LIST_GET_FIRST(os_mutex_list);
+
+	while (mutex) {
+
+	      os_mutex_free(mutex);
+
+	      mutex = UT_LIST_GET_FIRST(os_mutex_list);
+	}
+}
+
 /*************************************************************
 Creates an event semaphore, i.e., a semaphore which may
 just have two states: signaled and nonsignaled.
@@ -43,8 +110,8 @@ os_event_create(
 			the event is created without a name */
 {
 #ifdef __WIN__
-	HANDLE	event;
-	
+        os_event_t event;
+
 	event = CreateEvent(NULL,	/* No security attributes */
 			TRUE,		/* Manual reset */
 			FALSE,		/* Initial state nonsignaled */
@@ -75,6 +142,14 @@ os_event_create(
 #endif
 	event->is_set = FALSE;
 
+	os_mutex_enter(os_sync_mutex);
+
+	UT_LIST_ADD_FIRST(os_event_list, os_event_list, event);
+
+	os_event_count++;
+
+	os_mutex_exit(os_sync_mutex);
+
 	return(event);
 #endif
 }
@@ -92,7 +167,7 @@ os_event_create_auto(
 			the event is created without a name */
 {
 #ifdef __WIN__
-	HANDLE	event;
+        os_event_t event;
 
 	event = CreateEvent(NULL,	/* No security attributes */
 			FALSE,		/* Auto-reset */
@@ -106,6 +181,8 @@ os_event_create_auto(
 
 	UT_NOT_USED(name);
 
+	ut_a(0);
+
 	return(NULL);
 #endif
 }
@@ -185,12 +262,23 @@ os_event_free(
 	os_fast_mutex_free(&(event->os_mutex));
 	ut_a(0 == pthread_cond_destroy(&(event->cond_var)));
 
+	os_mutex_enter(os_sync_mutex);
+
+	UT_LIST_REMOVE(os_event_list, os_event_list, event);
+
+	os_event_count--;
+
+	os_mutex_exit(os_sync_mutex);
+
 	ut_free(event);
 #endif
 }
 
 /**************************************************************
-Waits for an event object until it is in the signaled state. */
+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). */
 
 void
 os_event_wait(
@@ -206,12 +294,20 @@ os_event_wait(
 	err = WaitForSingleObject(event, INFINITE);
 
 	ut_a(err == WAIT_OBJECT_0);
+
+	if (srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS) {
+	        os_thread_exit(NULL);
+	}
 #else
 	os_fast_mutex_lock(&(event->os_mutex));
 loop:
 	if (event->is_set == TRUE) {
 		os_fast_mutex_unlock(&(event->os_mutex));
 
+		if (srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS) {
+
+		        os_thread_exit(NULL);
+		}
 		/* Ok, we may return */
 
 		return;
@@ -291,14 +387,17 @@ os_event_wait_multiple(
 	ut_a(event_array);
 	ut_a(n > 0);
 
-	index = WaitForMultipleObjects(n,
-					event_array,
+	index = WaitForMultipleObjects(n, event_array,
 					FALSE,	   /* Wait for any 1 event */
 					INFINITE); /* Infinite wait time
 						   limit */
 	ut_a(index >= WAIT_OBJECT_0);
 	ut_a(index < WAIT_OBJECT_0 + n);
 
+	if (srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS) {
+	        os_thread_exit(NULL);
+	}
+
 	return(index - WAIT_OBJECT_0);
 #else
 	ut_a(n == 0);
@@ -337,6 +436,16 @@ os_mutex_create(
 	mutex_str->handle = mutex;
 	mutex_str->count = 0;
 
+	if (os_sync_mutex_inited) {
+	        os_mutex_enter(os_sync_mutex);
+	}
+
+	UT_LIST_ADD_FIRST(os_mutex_list, os_mutex_list, mutex_str);
+
+	if (os_sync_mutex_inited) {
+		os_mutex_exit(os_sync_mutex);
+	}
+
 	return(mutex_str);
 #else
 	os_fast_mutex_t*	os_mutex;
@@ -353,6 +462,16 @@ os_mutex_create(
 	mutex_str->handle = os_mutex;
 	mutex_str->count = 0;
 
+	if (os_sync_mutex_inited) {
+	        os_mutex_enter(os_sync_mutex);
+	}
+
+	UT_LIST_ADD_FIRST(os_mutex_list, os_mutex_list, mutex_str);
+
+	if (os_sync_mutex_inited) {
+		os_mutex_exit(os_sync_mutex);
+	}
+
 	return(mutex_str);
 #endif	
 }
@@ -424,9 +543,22 @@ os_mutex_free(
 #ifdef __WIN__
 	ut_a(mutex);
 
+	os_mutex_enter(os_sync_mutex);
+
+	UT_LIST_REMOVE(os_mutex_list, os_mutex_list, mutex);
+
+	os_mutex_exit(os_sync_mutex);
+
 	ut_a(CloseHandle(mutex->handle));
+
 	ut_free(mutex);
 #else
+	os_mutex_enter(os_sync_mutex);
+
+	UT_LIST_REMOVE(os_mutex_list, os_mutex_list, mutex);
+
+	os_mutex_exit(os_sync_mutex);
+
 	os_fast_mutex_free(mutex->handle);
 	ut_free(mutex->handle);
 	ut_free(mutex);
@@ -451,6 +583,7 @@ os_fast_mutex_init(
 #else
 	ut_a(0 == pthread_mutex_init(fast_mutex, MY_MUTEX_INIT_FAST));
 #endif
+	os_mutex_count++;
 #endif
 }
 
@@ -498,5 +631,6 @@ os_fast_mutex_free(
 	DeleteCriticalSection((LPCRITICAL_SECTION) fast_mutex);
 #else
 	ut_a(0 == pthread_mutex_destroy(fast_mutex));
+	os_mutex_count--;
 #endif
 }
diff --git a/innobase/os/os0thread.c b/innobase/os/os0thread.c
index b0076921e43..02ea2c227a7 100644
--- a/innobase/os/os0thread.c
+++ b/innobase/os/os0thread.c
@@ -1,6 +1,5 @@
 /******************************************************
-The interface to the operating system
-process and thread control primitives
+The interface to the operating system thread control primitives
 
 (c) 1995 Innobase Oy
 
@@ -17,6 +16,7 @@ Created 9/8/1995 Heikki Tuuri
 #endif
 
 #include "srv0srv.h"
+#include "os0sync.h"
 
 /*******************************************************************
 Compares two thread ids for equality. */
@@ -102,6 +102,10 @@ os_thread_create(
 	os_thread_t	thread;
 	ulint           win_thread_id;
 
+	os_mutex_enter(os_sync_mutex);
+	os_thread_count++;
+	os_mutex_exit(os_sync_mutex);
+
 	thread = CreateThread(NULL,	/* no security attributes */
 				0,	/* default size stack */
 				(LPTHREAD_START_ROUTINE)start_f,
@@ -144,6 +148,9 @@ os_thread_create(
 		 exit(1);
 	}
 #endif
+	os_mutex_enter(os_sync_mutex);
+	os_thread_count++;
+	os_mutex_exit(os_sync_mutex);
 
 #if defined(UNIV_HOTBACKUP) && defined(UNIV_HPUX10)
 	ret = pthread_create(&pthread, pthread_attr_default, start_f, arg);
@@ -171,6 +178,26 @@ os_thread_create(
 }
 
 /*********************************************************************
+Exits the current thread. */
+
+void
+os_thread_exit(
+/*===========*/
+	void*	exit_value)	/* in: exit value; in Windows this void*
+				is cast as a DWORD */
+{
+	os_mutex_enter(os_sync_mutex);
+	os_thread_count--;
+	os_mutex_exit(os_sync_mutex);
+
+#ifdef __WIN__
+        ExitThread((DWORD)exit_value);
+#else
+	pthread_exit(exit_value);
+#endif
+}
+
+/*********************************************************************
 Returns handle to the current thread. */
 
 os_thread_t
diff --git a/innobase/row/row0ins.c b/innobase/row/row0ins.c
index 3af9e1b752b..e96c08a715b 100644
--- a/innobase/row/row0ins.c
+++ b/innobase/row/row0ins.c
@@ -1275,6 +1275,10 @@ row_ins_unique_report_err(
 	dtuple_t*	entry,	/* in: index entry to insert in the index */
 	dict_index_t*	index)	/* in: index */
 {
+#ifdef notdefined
+        /* Disable reporting to test if the slowdown of REPLACE in 4.0.13 was
+	caused by this! */
+
 	char*	buf	= dict_unique_err_buf;
 
 	/* The foreign err mutex protects also dict_unique_err_buf */
@@ -1303,6 +1307,7 @@ row_ins_unique_report_err(
 	ut_a(strlen(buf) < DICT_FOREIGN_ERR_BUF_LEN);
 
 	mutex_exit(&dict_foreign_err_mutex);
+#endif
 }
 
 /*******************************************************************
diff --git a/innobase/srv/srv0srv.c b/innobase/srv/srv0srv.c
index 7cd831fe239..9353766248d 100644
--- a/innobase/srv/srv0srv.c
+++ b/innobase/srv/srv0srv.c
@@ -1705,76 +1705,33 @@ srv_init(void)
 		ut_a(conc_slot->event);
 	}
 }	
-	
+
 /*************************************************************************
-Initializes the synchronization primitives, memory system, and the thread
-local storage. */
+Frees the OS fast mutex created in srv_init(). */
 
 void
-srv_general_init(void)
-/*==================*/
+srv_free(void)
+/*==========*/
 {
-	sync_init();
-	mem_init(srv_mem_pool_size);
-	thr_local_init();
+	os_fast_mutex_free(&srv_conc_mutex);
 }
 
-
-#if defined(__NETWARE__) || defined(SAFE_MUTEX_DETECT_DESTROY)
-/* NetWare requires some cleanup of mutexes */
-
 /*************************************************************************
-Deinitializes the synchronization primitives, memory system, and the thread
+Initializes the synchronization primitives, memory system, and the thread
 local storage. */
 
 void
-srv_general_free(void)
+srv_general_init(void)
 /*==================*/
 {
-  sync_close();
+	os_sync_init();
+	sync_init();
+	mem_init(srv_mem_pool_size);
+	thr_local_init();
 }
-#endif /* __NETWARE__ */
-
 
 /*======================= InnoDB Server FIFO queue =======================*/
 
-#if defined(__NETWARE__) || defined(SAFE_MUTEX_DETECT_DESTROY)
-/* NetWare requires some cleanup of mutexes */
-
-/*************************************************************************
-Deinitializes the server. */
-
-void
-srv_free(void)
-/*==========*/
-{
-  srv_conc_slot_t* conc_slot;
-  srv_slot_t* slot;
-  ulint i;
-
-  for (i = 0; i < OS_THREAD_MAX_N; i++)
-  {
-    slot = srv_table_get_nth_slot(i);
-    os_event_free(slot->event);
-  }
-
-  /* TODO: free(srv_sys->threads); */
-
-  for (i = 0; i < OS_THREAD_MAX_N; i++)
-  {
-    slot = srv_mysql_table + i;
-    os_event_free(slot->event);
-  }
-
-  /* TODO: free(srv_mysql_table); */
-
-  for (i = 0; i < OS_THREAD_MAX_N; i++)
-  {
-    conc_slot = srv_conc_slots + i;
-    os_event_free(conc_slot->event);
-  }
-}
-#endif /* __NETWARE__ */
 
 /*************************************************************************
 Puts an OS thread to wait if there are too many concurrent threads
@@ -2734,6 +2691,8 @@ loop:
 
 	srv_error_monitor_active = FALSE;
 
+	os_thread_exit(NULL);
+
 #ifndef __WIN__
         return(NULL);
 #else
@@ -3173,6 +3132,13 @@ suspend_thread:
 
 	os_event_wait(event);
 
+	if (srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS) {
+	        /* This is only extra safety, the thread should exit
+		already when the event wait ends */
+
+	        os_thread_exit(NULL);
+	}
+
 	/* When there is user activity, InnoDB will set the event and the main
 	thread goes back to loop: */
 
diff --git a/innobase/srv/srv0start.c b/innobase/srv/srv0start.c
index 536aa5b67e4..9a2bf72130b 100644
--- a/innobase/srv/srv0start.c
+++ b/innobase/srv/srv0start.c
@@ -1481,9 +1481,7 @@ innobase_start_or_create_for_mysql(void)
 
 	os_fast_mutex_unlock(&srv_os_test_mutex);
 
-#if defined(__NETWARE__) || defined(SAFE_MUTEX_DETECT_DESTROY)
-	os_fast_mutex_free(&srv_os_test_mutex);  /* all platforms? */
-#endif /* __NETWARE__ */
+	os_fast_mutex_free(&srv_os_test_mutex);
 
 	if (srv_print_verbose_log) {
 	  	ut_print_timestamp(stderr);
@@ -1509,6 +1507,8 @@ innobase_shutdown_for_mysql(void)
 /*=============================*/
 				/* out: DB_SUCCESS or error code */
 {
+	ulint   i;
+
         if (!srv_was_started) {
 	  	if (srv_is_being_started) {
 	    		ut_print_timestamp(stderr);
@@ -1521,7 +1521,7 @@ innobase_shutdown_for_mysql(void)
 	  	return(DB_SUCCESS);
 	}
 
-	/* Flush buffer pool to disk, write the current lsn to
+	/* 1. Flush buffer pool to disk, write the current lsn to
 	the tablespace header(s), and copy all log data to archive */
 
 	logs_empty_and_mark_files_at_shutdown();
@@ -1533,30 +1533,77 @@ innobase_shutdown_for_mysql(void)
 		srv_conc_n_threads);
 	}
 
-#if defined(__NETWARE__) || defined(SAFE_MUTEX_DETECT_DESTROY)
-	/*
-	  TODO: Fix this temporary solution
-	  We are having a race condition occure with io_handler_thread threads.
-	  When they yield in os_aio_simulated_handle during shutdown, this
-	  thread was able to free the memory early.
-	*/
-	os_thread_yield();
+	/* 2. Make all threads created by InnoDB to exit */
+
+	srv_shutdown_state = SRV_SHUTDOWN_EXIT_THREADS;
+
+	/* All threads end up waiting for certain events. Put those events
+	to the signaled state. Then the threads will exit themselves in
+	os_thread_event_wait(). */
+
+	for (i = 0; i < 1000; i++) {
+	        /* NOTE: IF YOU CREATE THREADS IN INNODB, YOU MUST EXIT THEM
+	        HERE OR EARLIER */
+		
+		/* 1. Let the lock timeout thread exit */
+		os_event_set(srv_lock_timeout_thread_event);		
+
+		/* 2. srv error monitor thread exits automatically, no need
+		to do anything here */
+
+		/* 3. We wake the master thread so that it exits */
+		srv_wake_master_thread();
+
+		/* 4. Exit the i/o threads */
+
+		os_aio_wake_all_threads_at_shutdown();
+
+		os_mutex_enter(os_sync_mutex);
+
+		if (os_thread_count == 0) {
+		        /* All the threads have exited or are just exiting;
+			NOTE that the threads may not have completed their
+			exit yet. Should we use pthread_join() to make sure
+			they have exited? Now we just sleep 0.1 seconds and
+			hope that is enough! */
+
+			os_mutex_exit(os_sync_mutex);
+
+			os_thread_sleep(100000);
+
+			break;
+		}
+
+		os_mutex_exit(os_sync_mutex);
+
+		os_thread_sleep(100000);
+	}
+
+	if (i == 1000) {
+	        fprintf(stderr,
+"InnoDB: Warning: %lu threads created by InnoDB had not exited at shutdown!\n",
+		      os_thread_count);
+	}
+
+	/* 3. Free all InnoDB's own mutexes */
+
+	sync_close();
+
+	/* 4. Free all OS synchronization primitives (in Windows currently
+	events are not freed) */
 
-	/* TODO: Where should this be called? */
 	srv_free();
+	os_sync_free();
+
+	/* 5. Free all allocated memory (and the os_fast_mutex created in
+	ut0mem.c */
 
-	/* TODO: Where should this be called? */
-	srv_general_free();
-#endif
-	/*
-	TODO: We should exit the i/o-handler and other utility threads
-	before freeing all memory. Now this can potentially cause a seg
-	fault!
-	*/
-#if defined(NOT_WORKING_YET) || defined(__NETWARE__) || defined(SAFE_MUTEX_DETECT_DESTROY)
-        /* NetWare requires this free */
         ut_free_all_mem();
-#endif 
+
+	if (srv_print_verbose_log) {
+	        ut_print_timestamp(stderr);
+	        fprintf(stderr, "  InnoDB: Shutdown completed\n");
+	}
 
 	return((int) DB_SUCCESS);
 }
diff --git a/innobase/sync/sync0sync.c b/innobase/sync/sync0sync.c
index 20d68ba5a9f..32615ce88ac 100644
--- a/innobase/sync/sync0sync.c
+++ b/innobase/sync/sync0sync.c
@@ -235,8 +235,7 @@ mutex_create_func(
 	mutex->cline = cline;
 	
 	/* Check that lock_word is aligned; this is important on Intel */
-
-	ut_a(((ulint)(&(mutex->lock_word))) % 4 == 0);
+	ut_ad(((ulint)(&(mutex->lock_word))) % 4 == 0);
 
 	/* NOTE! The very first mutexes are not put to the mutex list */
 
@@ -266,11 +265,14 @@ mutex_free(
 	ut_a(mutex_get_lock_word(mutex) == 0);
 	ut_a(mutex_get_waiters(mutex) == 0);
 	
-	mutex_enter(&mutex_list_mutex);
+	if (mutex != &mutex_list_mutex && mutex != &sync_thread_mutex) {
 
-	UT_LIST_REMOVE(list, mutex_list, mutex);
+	        mutex_enter(&mutex_list_mutex);
 
-	mutex_exit(&mutex_list_mutex);
+	        UT_LIST_REMOVE(list, mutex_list, mutex);
+
+		mutex_exit(&mutex_list_mutex);
+	}
 
 #if !defined(_WIN32) || !defined(UNIV_CAN_USE_X86_ASSEMBLER) 
 	os_fast_mutex_free(&(mutex->os_fast_mutex));
@@ -1230,13 +1232,26 @@ sync_init(void)
 }
 
 /**********************************************************************
-Frees the resources in synchronization data structures. */
+Frees the resources in InnoDB's own synchronization data structures. Use
+os_sync_free() after calling this. */
 
 void
 sync_close(void)
 /*===========*/
 {
+	mutex_t*	mutex;
+
 	sync_array_free(sync_primary_wait_array);
+
+	mutex = UT_LIST_GET_FIRST(mutex_list);
+
+	while (mutex) {
+	        mutex_free(mutex);
+		mutex = UT_LIST_GET_FIRST(mutex_list);
+	}
+
+	mutex_free(&mutex_list_mutex);
+	mutex_free(&sync_thread_mutex);	
 }
 
 /***********************************************************************
diff --git a/innobase/ut/ut0mem.c b/innobase/ut/ut0mem.c
index 23deed0fa55..2609b8f5241 100644
--- a/innobase/ut/ut0mem.c
+++ b/innobase/ut/ut0mem.c
@@ -225,6 +225,8 @@ ut_free_all_mem(void)
 	os_fast_mutex_unlock(&ut_list_mutex);
 
 	ut_a(ut_total_allocated_memory == 0);
+
+        os_fast_mutex_free(&ut_list_mutex);
 }
 
 /**************************************************************************