Bug 12323643 - CLEAN UP THE INNODB THREAD SHUTDOWN AND ASSERTIONS (WL#5136)

On shutdown, do not exit threads in os_event_wait(). This method of
exiting was only used by the I/O handler threads. Exit them on a
higher level.

os_event_wait_low(), os_event_wait_time_low(): Do not exit on shutdown.

os_thread_exit(), ut_dbg_assertion_failed(), ut_print_timestamp(): Add
attribute cold, so that GCC knows that these functions are rarely
invoked and can be optimized for size.

os_aio_linux_collect(): Return on shutdown.

os_aio_linux_handle(), os_aio_simulated_handle(), os_aio_windows_handle():
Set *message1 = *message2 = NULL and return TRUE on shutdown.

fil_aio_wait(): Return on shutdown.

logs_empty_and_mark_files_at_shutdown(): Even in very fast shutdown
(innodb_fast_shutdown=2), allow the background threads to exit, but
skip the flushing and log checkpointing.

innobase_shutdown_for_mysql(): Always wait for all the threads to exit.

rb:633 approved by Sunny Bains

storage/innobase/fil/fil0fil.c

@@ -4527,8 +4527,8 @@ fil_aio_wait(
 		ret = os_aio_linux_handle(segment, &fil_node,
 					  &message, &type);
 #else
-		ret = 0; /* Eliminate compiler warning */
 		ut_error;
+		ret = 0; /* Eliminate compiler warning */
 #endif
 	} else {
 		srv_set_io_thread_op_info(segment, "simulated aio handle");
@@ -4538,6 +4538,10 @@ fil_aio_wait(
 	}
 
 	ut_a(ret);
+	if (UNIV_UNLIKELY(fil_node == NULL)) {
+		ut_ad(srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS);
+		return;
+	}
 
 	srv_set_io_thread_op_info(segment, "complete io for fil node");
 

storage/innobase/include/os0sync.h

@@ -150,10 +150,7 @@ 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).
+Waits for an event object until it is 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.

storage/innobase/include/os0thread.h

@@ -107,8 +107,9 @@ UNIV_INTERN
 void
 os_thread_exit(
 /*===========*/
-	void*	exit_value);	/*!< in: exit value; in Windows this void*
+	void*	exit_value)	/*!< in: exit value; in Windows this void*
 				is cast as a DWORD */
+	__attribute__((cold, noreturn));
 /*****************************************************************//**
 Returns the thread identifier of current thread.
 @return	current thread identifier */

storage/innobase/include/ut0dbg.h

@@ -50,9 +50,10 @@ UNIV_INTERN
 void
 ut_dbg_assertion_failed(
 /*====================*/
-	const char* expr,	/*!< in: the failed assertion */
-	const char* file,	/*!< in: source file containing the assertion */
-	ulint line);		/*!< in: line number of the assertion */
+	const char*	expr,	/*!< in: the failed assertion */
+	const char*	file,	/*!< in: source file containing the assertion */
+	ulint		line)	/*!< in: line number of the assertion */
+	__attribute__((nonnull(2), cold));
 
 #if defined(__WIN__) || defined(__INTEL_COMPILER)
 # undef UT_DBG_USE_ABORT

storage/innobase/include/ut0ut.h

@@ -275,7 +275,8 @@ UNIV_INTERN
 void
 ut_print_timestamp(
 /*===============*/
-	FILE*  file); /*!< in: file where to print */
+	FILE*	file)	/*!< in: file where to print */
+	__attribute__((nonnull, cold));
 /**********************************************************//**
 Sprintfs a timestamp to a buffer, 13..14 chars plus terminating NUL. */
 UNIV_INTERN

storage/innobase/log/log0log.c

@@ -3078,6 +3078,7 @@ logs_empty_and_mark_files_at_shutdown(void)
 {
 	ib_uint64_t	lsn;
 	ulint		arch_log_no;
+	ibool		server_busy;
 
 	if (srv_print_verbose_log) {
 		ut_print_timestamp(stderr);
@@ -3092,14 +3093,12 @@ loop:
 
 	mutex_enter(&kernel_mutex);
 
-	/* We need the monitor threads to stop before we proceed with a
-	normal shutdown. In case of very fast shutdown, however, we can
-	proceed without waiting for monitor threads. */
+	/* We need the monitor threads to stop before we proceed with
+	a shutdown. */
 
-	if (srv_fast_shutdown < 2
-	   && (srv_error_monitor_active
-	      || srv_lock_timeout_active
-	      || srv_monitor_active)) {
+	if (srv_error_monitor_active
+	    || srv_lock_timeout_active
+	    || srv_monitor_active) {
 
 		mutex_exit(&kernel_mutex);
 
@@ -3114,65 +3113,57 @@ loop:
 	for the 'very fast' shutdown, because the InnoDB layer may have
 	committed or prepared transactions and we don't want to lose them. */
 
-	if (trx_n_mysql_transactions > 0
-	    || UT_LIST_GET_LEN(trx_sys->trx_list) > 0) {
-
-		mutex_exit(&kernel_mutex);
-
-		goto loop;
-	}
-
-	if (srv_fast_shutdown == 2) {
-		/* In this fastest shutdown we do not flush the buffer pool:
-		it is essentially a 'crash' of the InnoDB server. Make sure
-		that the log is all flushed to disk, so that we can recover
-		all committed transactions in a crash recovery. We must not
-		write the lsn stamps to the data files, since at a startup
-		InnoDB deduces from the stamps if the previous shutdown was
-		clean. */
-
-		log_buffer_flush_to_disk();
-
-		mutex_exit(&kernel_mutex);
-
-		return; /* We SKIP ALL THE REST !! */
-	}
-
+	server_busy = trx_n_mysql_transactions > 0
+		|| UT_LIST_GET_LEN(trx_sys->trx_list) > 0;
 	mutex_exit(&kernel_mutex);
 
-	/* Check that the background threads are suspended */
-
-	if (srv_is_any_background_thread_active()) {
+	if (server_busy || srv_is_any_background_thread_active()) {
 		goto loop;
 	}
 
-	mutex_enter(&(log_sys->mutex));
-
-	if (log_sys->n_pending_checkpoint_writes
+	mutex_enter(&log_sys->mutex);
+	server_busy = log_sys->n_pending_checkpoint_writes
 #ifdef UNIV_LOG_ARCHIVE
-	    || log_sys->n_pending_archive_ios
+		|| log_sys->n_pending_archive_ios
 #endif /* UNIV_LOG_ARCHIVE */
-	    || log_sys->n_pending_writes) {
-
-		mutex_exit(&(log_sys->mutex));
-
-		goto loop;
-	}
-
-	mutex_exit(&(log_sys->mutex));
-
-	if (!buf_pool_check_no_pending_io()) {
+		|| log_sys->n_pending_writes;
+	mutex_exit(&log_sys->mutex);
 
+	if (server_busy || !buf_pool_check_no_pending_io()) {
 		goto loop;
 	}
 
 #ifdef UNIV_LOG_ARCHIVE
 	log_archive_all();
 #endif /* UNIV_LOG_ARCHIVE */
+	if (srv_fast_shutdown == 2) {
+		/* In this fastest shutdown we do not flush the buffer
+		pool: it is essentially a 'crash' of the InnoDB
+		server. Make sure that the log is all flushed to disk,
+		so that we can recover all committed transactions in a
+		crash recovery. We must not write the lsn stamps to
+		the data files, since at a startup InnoDB deduces from
+		the stamps if the previous shutdown was clean. */
+
+		log_buffer_flush_to_disk();
+
+		/* Check that the background threads stay suspended */
+		if (srv_is_any_background_thread_active()) {
+			fprintf(stderr,
+				"InnoDB: Warning: some background thread"
+				" woke up during shutdown\n");
+			goto loop;
+		}
+
+		srv_shutdown_state = SRV_SHUTDOWN_LAST_PHASE;
+		fil_close_all_files();
+		ut_a(!srv_is_any_background_thread_active());
+		return;
+	}
 
 	log_make_checkpoint_at(IB_ULONGLONG_MAX, TRUE);
 
-	mutex_enter(&(log_sys->mutex));
+	mutex_enter(&log_sys->mutex);
 
 	lsn = log_sys->lsn;
 
@@ -3183,7 +3174,7 @@ loop:
 #endif /* UNIV_LOG_ARCHIVE */
 	    ) {
 
-		mutex_exit(&(log_sys->mutex));
+		mutex_exit(&log_sys->mutex);
 
 		goto loop;
 	}
@@ -3201,7 +3192,7 @@ loop:
 	log_archive_close_groups(TRUE);
 #endif /* UNIV_LOG_ARCHIVE */
 
-	mutex_exit(&(log_sys->mutex));
+	mutex_exit(&log_sys->mutex);
 
 	/* Check that the background threads stay suspended */
 	if (srv_is_any_background_thread_active()) {

storage/innobase/os/os0file.c

@@ -4064,13 +4064,13 @@ os_aio_func(
 	}
 
 try_again:
-	if (mode == OS_AIO_NORMAL) {
-		if (type == OS_FILE_READ) {
-			array = os_aio_read_array;
-		} else {
-			array = os_aio_write_array;
-		}
-	} else if (mode == OS_AIO_IBUF) {
+	switch (mode) {
+	case OS_AIO_NORMAL:
+		array = (type == OS_FILE_READ)
+			? os_aio_read_array
+			: os_aio_write_array;
+		break;
+	case OS_AIO_IBUF:
 		ut_ad(type == OS_FILE_READ);
 		/* Reduce probability of deadlock bugs in connection with ibuf:
 		do not let the ibuf i/o handler sleep */
@@ -4078,19 +4078,21 @@ try_again:
 		wake_later = FALSE;
 
 		array = os_aio_ibuf_array;
-	} else if (mode == OS_AIO_LOG) {
-
+		break;
+	case OS_AIO_LOG:
 		array = os_aio_log_array;
-	} else if (mode == OS_AIO_SYNC) {
+		break;
+	case OS_AIO_SYNC:
 		array = os_aio_sync_array;
 
 #if defined(LINUX_NATIVE_AIO)
 		/* In Linux native AIO we don't use sync IO array. */
 		ut_a(!srv_use_native_aio);
 #endif /* LINUX_NATIVE_AIO */
-	} else {
-		array = NULL; /* Eliminate compiler warning */
+		break;
+	default:
 		ut_error;
+		array = NULL; /* Eliminate compiler warning */
 	}
 
 	slot = os_aio_array_reserve_slot(type, array, message1, message2, file,
@@ -4253,11 +4255,17 @@ os_aio_windows_handle(
 					   INFINITE);
 	}
 
-	if (srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS) {
-		os_thread_exit(NULL);
+	os_mutex_enter(array->mutex);
+
+	if (srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS
+	    && array->n_reserved == 0) {
+		*message1 = NULL;
+		*message2 = NULL;
+		os_mutex_exit(array->mutex);
+		return(TRUE);
 	}
 
-	os_mutex_enter(array->mutex);
+	ut_a(i >= WAIT_OBJECT_0 && i <= WAIT_OBJECT_0 + n);
 
 	slot = os_aio_array_get_nth_slot(array, i + segment * n);
 
@@ -4403,14 +4411,6 @@ os_aio_linux_collect(
 
 retry:
 
-	/* Go down if we are in shutdown mode.
-	In case of srv_fast_shutdown == 2, there may be pending
-	IO requests but that should be OK as we essentially treat
-	that as a crash of InnoDB. */
-	if (srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS) {
-		os_thread_exit(NULL);
-	}
-
 	/* Initialize the events. The timeout value is arbitrary.
 	We probably need to experiment with it a little. */
 	memset(events, 0, sizeof(*events) * seg_size);
@@ -4419,76 +4419,72 @@ retry:
 
 	ret = io_getevents(io_ctx, 1, seg_size, events, &timeout);
 
-	/* This error handling is for any error in collecting the
-	IO requests. The errors, if any, for any particular IO
-	request are simply passed on to the calling routine. */
-
-	/* Not enough resources! Try again. */
-	if (ret == -EAGAIN) {
-		goto retry;
-	}
-
-	/* Interrupted! I have tested the behaviour in case of an
-	interrupt. If we have some completed IOs available then
-	the return code will be the number of IOs. We get EINTR only
-	if there are no completed IOs and we have been interrupted. */
-	if (ret == -EINTR) {
-		goto retry;
-	}
-
-	/* No pending request! Go back and check again. */
-	if (ret == 0) {
-		goto retry;
-	}
+	if (ret > 0) {
+		for (i = 0; i < ret; i++) {
+			os_aio_slot_t*	slot;
+			struct iocb*	control;
 
-	/* All other errors! should cause a trap for now. */
-	if (UNIV_UNLIKELY(ret < 0)) {
-		ut_print_timestamp(stderr);
-		fprintf(stderr,
-			"  InnoDB: unexpected ret_code[%d] from"
-			" io_getevents()!\n", ret);
-		ut_error;
-	}
+			control = (struct iocb *)events[i].obj;
+			ut_a(control != NULL);
 
-	ut_a(ret > 0);
+			slot = (os_aio_slot_t *) control->data;
 
-	for (i = 0; i < ret; i++) {
-		os_aio_slot_t*	slot;
-		struct iocb*	control;
+			/* Some sanity checks. */
+			ut_a(slot != NULL);
+			ut_a(slot->reserved);
 
-		control = (struct iocb *)events[i].obj;
-		ut_a(control != NULL);
+#if defined(UNIV_AIO_DEBUG)
+			fprintf(stderr,
+				"io_getevents[%c]: slot[%p] ctx[%p]"
+				" seg[%lu]\n",
+				(slot->type == OS_FILE_WRITE) ? 'w' : 'r',
+				slot, io_ctx, segment);
+#endif
 
-		slot = (os_aio_slot_t *) control->data;
+			/* We are not scribbling previous segment. */
+			ut_a(slot->pos >= start_pos);
 
-		/* Some sanity checks. */
-		ut_a(slot != NULL);
-		ut_a(slot->reserved);
+			/* We have not overstepped to next segment. */
+			ut_a(slot->pos < end_pos);
 
-#if defined(UNIV_AIO_DEBUG)
-		fprintf(stderr,
-			"io_getevents[%c]: slot[%p] ctx[%p]"
-			" seg[%lu]\n",
-			(slot->type == OS_FILE_WRITE) ? 'w' : 'r',
-			slot, io_ctx, segment);
-#endif
+			/* Mark this request as completed. The error handling
+			will be done in the calling function. */
+			os_mutex_enter(array->mutex);
+			slot->n_bytes = events[i].res;
+			slot->ret = events[i].res2;
+			slot->io_already_done = TRUE;
+			os_mutex_exit(array->mutex);
+		}
+		return;
+	}
 
-		/* We are not scribbling previous segment. */
-		ut_a(slot->pos >= start_pos);
+	if (UNIV_UNLIKELY(srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS)) {
+		return;
+	}
 
-		/* We have not overstepped to next segment. */
-		ut_a(slot->pos < end_pos);
+	/* This error handling is for any error in collecting the
+	IO requests. The errors, if any, for any particular IO
+	request are simply passed on to the calling routine. */
 
-		/* Mark this request as completed. The error handling
-		will be done in the calling function. */
-		os_mutex_enter(array->mutex);
-		slot->n_bytes = events[i].res;
-		slot->ret = events[i].res2;
-		slot->io_already_done = TRUE;
-		os_mutex_exit(array->mutex);
+	switch (ret) {
+	case -EAGAIN:
+		/* Not enough resources! Try again. */
+	case -EINTR:
+		/* Interrupted! I have tested the behaviour in case of an
+		interrupt. If we have some completed IOs available then
+		the return code will be the number of IOs. We get EINTR only
+		if there are no completed IOs and we have been interrupted. */
+	case 0:
+		/* No pending request! Go back and check again. */
+		goto retry;
 	}
 
-	return;
+	/* All other errors should cause a trap for now. */
+	ut_print_timestamp(stderr);
+	fprintf(stderr,
+		"  InnoDB: unexpected ret_code[%d] from io_getevents()!\n",
+		ret);
+	ut_error;
 }
 
 /**********************************************************************//**
@@ -4532,20 +4528,35 @@ os_aio_linux_handle(
 
 	/* Loop until we have found a completed request. */
 	for (;;) {
+		ibool	any_reserved = FALSE;
 		os_mutex_enter(array->mutex);
 		for (i = 0; i < n; ++i) {
 			slot = os_aio_array_get_nth_slot(
-					array, i + segment * n);
-			if (slot->reserved && slot->io_already_done) {
+				array, i + segment * n);
+			if (!slot->reserved) {
+				continue;
+			} else if (slot->io_already_done) {
 				/* Something for us to work on. */
 				goto found;
+			} else {
+				any_reserved = TRUE;
 			}
 		}
 
 		os_mutex_exit(array->mutex);
 
-		/* We don't have any completed request.
-		Wait for some request. Note that we return
+		/* There is no completed request.
+		If there is no pending request at all,
+		and the system is being shut down, exit. */
+		if (UNIV_UNLIKELY
+		    (!any_reserved
+		     && srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS)) {
+			*message1 = NULL;
+			*message2 = NULL;
+			return(TRUE);
+		}
+
+		/* Wait for some request. Note that we return
 		from wait iff we have found a request. */
 
 		srv_set_io_thread_op_info(global_seg,
@@ -4641,6 +4652,7 @@ os_aio_simulated_handle(
 	byte*		combined_buf;
 	byte*		combined_buf2;
 	ibool		ret;
+	ibool		any_reserved;
 	ulint		n;
 	ulint		i;
 
@@ -4671,18 +4683,21 @@ restart:
 		goto recommended_sleep;
 	}
 
-	os_mutex_enter(array->mutex);
-
 	srv_set_io_thread_op_info(global_segment,
 				  "looking for i/o requests (b)");
 
 	/* Check if there is a slot for which the i/o has already been
 	done */
+	any_reserved = FALSE;
+
+	os_mutex_enter(array->mutex);
 
 	for (i = 0; i < n; i++) {
 		slot = os_aio_array_get_nth_slot(array, i + segment * n);
 
-		if (slot->reserved && slot->io_already_done) {
+		if (!slot->reserved) {
+			continue;
+		} else if (slot->io_already_done) {
 
 			if (os_aio_print_debug) {
 				fprintf(stderr,
@@ -4694,9 +4709,23 @@ restart:
 			ret = TRUE;
 
 			goto slot_io_done;
+		} else {
+			any_reserved = TRUE;
 		}
 	}
 
+	/* There is no completed request.
+	If there is no pending request at all,
+	and the system is being shut down, exit. */
+	if (UNIV_UNLIKELY
+	    (!any_reserved
+	     && srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS)) {
+		os_mutex_exit(array->mutex);
+		*message1 = NULL;
+		*message2 = NULL;
+		return(TRUE);
+	}
+
 	n_consecutive = 0;
 
 	/* If there are at least 2 seconds old requests, then pick the oldest

storage/innobase/os/os0sync.c

@@ -558,10 +558,7 @@ 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).
+Waits for an event object until it is 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.
@@ -586,8 +583,6 @@ os_event_wait_low(
 					returned by previous call of
 					os_event_reset(). */
 {
-	ib_int64_t	old_signal_count;
-
 #ifdef __WIN__
 	if(!srv_use_native_conditions) {
 		DWORD	err;
@@ -600,43 +595,25 @@ os_event_wait_low(
 		err = WaitForSingleObject(event->handle, INFINITE);
 
 		ut_a(err == WAIT_OBJECT_0);
-
-		if (srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS) {
-			os_thread_exit(NULL);
-		}
 		return;
 	}
 #endif
 
-	os_fast_mutex_lock(&(event->os_mutex));
+	os_fast_mutex_lock(&event->os_mutex);
 
-	if (reset_sig_count) {
-		old_signal_count = reset_sig_count;
-	} else {
-		old_signal_count = event->signal_count;
+	if (!reset_sig_count) {
+		reset_sig_count = event->signal_count;
 	}
 
-	for (;;) {
-		if (event->is_set == TRUE
-		    || event->signal_count != old_signal_count) {
-
-			os_fast_mutex_unlock(&(event->os_mutex));
-
-			if (srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS) {
-
-				os_thread_exit(NULL);
-			}
-			/* Ok, we may return */
-
-			return;
-		}
-
+	while (!event->is_set && event->signal_count == reset_sig_count) {
 		os_cond_wait(&(event->cond_var), &(event->os_mutex));
 
 		/* Solaris manual said that spurious wakeups may occur: we
 		have to check if the event really has been signaled after
 		we came here to wait */
 	}
+
+	os_fast_mutex_unlock(&event->os_mutex);
 }
 
 /**********************************************************//**
@@ -657,7 +634,6 @@ os_event_wait_time_low(
 
 {
 	ibool		timed_out = FALSE;
-	ib_int64_t	old_signal_count;
 
 #ifdef __WIN__
 	DWORD		time_in_ms;
@@ -727,15 +703,12 @@ os_event_wait_time_low(
 
 	os_fast_mutex_lock(&event->os_mutex);
 
-	if (reset_sig_count) {
-		old_signal_count = reset_sig_count;
-	} else {
-		old_signal_count = event->signal_count;
+	if (!reset_sig_count) {
+		reset_sig_count = event->signal_count;
 	}
 
 	do {
-		if (event->is_set == TRUE
-		    || event->signal_count != old_signal_count) {
+		if (event->is_set || event->signal_count != reset_sig_count) {
 
 			break;
 		}
@@ -753,11 +726,6 @@ os_event_wait_time_low(
 
 	os_fast_mutex_unlock(&event->os_mutex);
 
-	if (srv_shutdown_state == SRV_SHUTDOWN_EXIT_THREADS) {
-
-		os_thread_exit(NULL);
-	}
-
 	return(timed_out ? OS_SYNC_TIME_EXCEEDED : 0);
 }
 

storage/innobase/srv/srv0srv.c

@@ -3082,11 +3082,7 @@ suspend_thread:
 	os_event_wait(slot->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

storage/innobase/srv/srv0start.c

@@ -2122,17 +2122,9 @@ innobase_shutdown_for_mysql(void)
 
 	srv_shutdown_state = SRV_SHUTDOWN_EXIT_THREADS;
 
-	/* In a 'very fast' shutdown, we do not need to wait for these threads
-	to die; all which counts is that we flushed the log; a 'very fast'
-	shutdown is essentially a crash. */
-
-	if (srv_fast_shutdown == 2) {
-		return(DB_SUCCESS);
-	}
-
 	/* 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(). */
+	to the signaled state. Then the threads will exit themselves after
+	os_event_wait(). */
 
 	for (i = 0; i < 1000; i++) {
 		/* NOTE: IF YOU CREATE THREADS IN INNODB, YOU MUST EXIT THEM