MDEV-11802 innodb.innodb_bug14676111 fails

The function trx_purge_stop() was calling os_event_reset(purge_sys->event)
before calling rw_lock_x_lock(&purge_sys->latch). The os_event_set()
call in srv_purge_coordinator_suspend() is protected by that X-latch.

It would seem a good idea to consistently protect both os_event_set()
and os_event_reset() calls with a common mutex or rw-lock in those
cases where os_event_set() and os_event_reset() are used
like condition variables, tied to changes of shared state.

For each os_event_t, we try to document the mutex or rw-lock that is
being used. For some events, frequent calls to os_event_set() seem to
try to avoid hangs. Some events are never waited for infinitely, only
timed waits, and os_event_set() is used for early termination of these
waits.

os_aio_simulated_put_read_threads_to_sleep(): Define as a null macro
on other systems than Windows. TODO: remove this altogether and disable
innodb_use_native_aio on Windows.

os_aio_segment_wait_events[]: Initialize only if innodb_use_native_aio=0.

Author: dr-m

storage/innobase/buf/buf0dump.cc

@@ -1,6 +1,7 @@
 /*****************************************************************************
 
 Copyright (c) 2011, 2016, Oracle and/or its affiliates. All Rights Reserved.
+Copyright (c) 2017, MariaDB Corporation. All Rights Reserved.
 
 This program is free software; you can redistribute it and/or modify it under
 the terms of the GNU General Public License as published by the Free Software
@@ -52,8 +53,8 @@ enum status_severity {
 
 /* Flags that tell the buffer pool dump/load thread which action should it
 take after being waked up. */
-static ibool	buf_dump_should_start = FALSE;
-static ibool	buf_load_should_start = FALSE;
+static volatile bool	buf_dump_should_start;
+static volatile bool	buf_load_should_start;
 
 static ibool	buf_load_abort_flag = FALSE;
 
@@ -78,7 +79,7 @@ void
 buf_dump_start()
 /*============*/
 {
-	buf_dump_should_start = TRUE;
+	buf_dump_should_start = true;
 	os_event_set(srv_buf_dump_event);
 }
 
@@ -92,7 +93,7 @@ void
 buf_load_start()
 /*============*/
 {
-	buf_load_should_start = TRUE;
+	buf_load_should_start = true;
 	os_event_set(srv_buf_dump_event);
 }
 
@@ -695,15 +696,18 @@ DECLARE_THREAD(buf_dump_thread)(
 		os_event_wait(srv_buf_dump_event);
 
 		if (buf_dump_should_start) {
-			buf_dump_should_start = FALSE;
+			buf_dump_should_start = false;
 			buf_dump(TRUE /* quit on shutdown */);
 		}
 
 		if (buf_load_should_start) {
-			buf_load_should_start = FALSE;
+			buf_load_should_start = false;
 			buf_load();
 		}
 
+		if (buf_dump_should_start || buf_load_should_start) {
+			continue;
+		}
 		os_event_reset(srv_buf_dump_event);
 	}
 

storage/innobase/buf/buf0flu.cc

@@ -1,6 +1,7 @@
 /*****************************************************************************
 
 Copyright (c) 1995, 2016, Oracle and/or its affiliates. All Rights Reserved.
+Copyright (c) 2017, MariaDB Corporation. All Rights Reserved.
 
 This program is free software; you can redistribute it and/or modify it under
 the terms of the GNU General Public License as published by the Free Software
@@ -732,8 +733,7 @@ buf_flush_write_complete(
 	flush_type = buf_page_get_flush_type(bpage);
 	buf_pool->n_flush[flush_type]--;
 
-	/* fprintf(stderr, "n pending flush %lu\n",
-	buf_pool->n_flush[flush_type]); */
+	ut_ad(buf_pool_mutex_own(buf_pool));
 
 	if (buf_pool->n_flush[flush_type] == 0
 	    && buf_pool->init_flush[flush_type] == FALSE) {

storage/innobase/dict/dict0stats_bg.cc

@@ -1,6 +1,7 @@
 /*****************************************************************************
 
 Copyright (c) 2012, 2016, Oracle and/or its affiliates. All Rights Reserved.
+Copyright (c) 2017, MariaDB Corporation. All Rights Reserved.
 
 This program is free software; you can redistribute it and/or modify it under
 the terms of the GNU General Public License as published by the Free Software
@@ -39,8 +40,9 @@ Created Apr 25, 2012 Vasil Dimov
 
 #define SHUTTING_DOWN()		(srv_shutdown_state != SRV_SHUTDOWN_NONE)
 
-/** Event to wake up the stats thread */
-UNIV_INTERN os_event_t		dict_stats_event = NULL;
+/** Event to wake up dict_stats_thread on dict_stats_recalc_pool_add()
+or shutdown. Not protected by any mutex. */
+UNIV_INTERN os_event_t		dict_stats_event;
 
 /** This mutex protects the "recalc_pool" variable. */
 static ib_mutex_t		recalc_pool_mutex;

storage/innobase/include/buf0buf.h

@@ -1,6 +1,7 @@
 /*****************************************************************************
 
 Copyright (c) 1995, 2016, Oracle and/or its affiliates. All Rights Reserved.
+Copyright (c) 2017, MariaDB Corporation. All Rights Reserved.
 
 This program is free software; you can redistribute it and/or modify it under
 the terms of the GNU General Public License as published by the Free Software
@@ -1905,7 +1906,9 @@ struct buf_pool_t{
 	os_event_t	no_flush[BUF_FLUSH_N_TYPES];
 					/*!< this is in the set state
 					when there is no flush batch
-					of the given type running */
+					of the given type running;
+					os_event_set() and os_event_reset()
+					are protected by buf_pool_t::mutex */
 	ib_rbt_t*	flush_rbt;	/*!< a red-black tree is used
 					exclusively during recovery to
 					speed up insertions in the

storage/innobase/include/buf0dblwr.h

@@ -1,6 +1,7 @@
 /*****************************************************************************
 
 Copyright (c) 1995, 2014, Oracle and/or its affiliates. All Rights Reserved.
+Copyright (c) 2017, MariaDB Corporation. All Rights Reserved.
 
 This program is free software; you can redistribute it and/or modify it under
 the terms of the GNU General Public License as published by the Free Software
@@ -134,11 +135,13 @@ struct buf_dblwr_t{
 	ulint		b_reserved;/*!< number of slots currently reserved
 				for batch flush. */
 	os_event_t	b_event;/*!< event where threads wait for a
-				batch flush to end. */
+				batch flush to end;
+				os_event_set() and os_event_reset()
+				are protected by buf_dblwr_t::mutex */
 	ulint		s_reserved;/*!< number of slots currently
 				reserved for single page flushes. */
 	os_event_t	s_event;/*!< event where threads wait for a
-				single page flush slot. */
+				single page flush slot. Protected by mutex. */
 	bool*		in_use;	/*!< flag used to indicate if a slot is
 				in use. Only used for single page
 				flushes. */

storage/innobase/include/dict0stats_bg.h

@@ -1,6 +1,7 @@
 /*****************************************************************************
 
 Copyright (c) 2012, 2016, Oracle and/or its affiliates. All Rights Reserved.
+Copyright (c) 2017, MariaDB Corporation. All Rights Reserved.
 
 This program is free software; you can redistribute it and/or modify it under
 the terms of the GNU General Public License as published by the Free Software
@@ -32,7 +33,8 @@ Created Apr 26, 2012 Vasil Dimov
 #include "os0sync.h" /* os_event_t */
 #include "os0thread.h" /* DECLARE_THREAD */
 
-/** Event to wake up the stats thread */
+/** Event to wake up dict_stats_thread on dict_stats_recalc_pool_add()
+or shutdown. Not protected by any mutex. */
 extern os_event_t	dict_stats_event;
 
 /*****************************************************************//**

storage/innobase/include/fil0fil.h

@@ -1,6 +1,7 @@
 /*****************************************************************************
 
 Copyright (c) 1995, 2016, Oracle and/or its affiliates. All Rights Reserved.
+Copyright (c) 2017, MariaDB Corporation. All Rights Reserved.
 
 This program is free software; you can redistribute it and/or modify it under
 the terms of the GNU General Public License as published by the Free Software
@@ -206,7 +207,9 @@ struct fil_node_t {
 	ibool		open;	/*!< TRUE if file open */
 	os_file_t	handle;	/*!< OS handle to the file, if file open */
 	os_event_t	sync_event;/*!< Condition event to group and
-				serialize calls to fsync */
+				serialize calls to fsync;
+				os_event_set() and os_event_reset()
+				are protected by fil_system_t::mutex */
 	ibool		is_raw_disk;/*!< TRUE if the 'file' is actually a raw
 				device or a raw disk partition */
 	ulint		size;	/*!< size of the file in database pages, 0 if

storage/innobase/include/fsp0types.h

@@ -1,6 +1,7 @@
 /*****************************************************************************
 
 Copyright (c) 1995, 2009, Oracle and/or its affiliates. All Rights Reserved.
+Copyright (c) 2017, MariaDB Corporation. All Rights Reserved.
 
 This program is free software; you can redistribute it and/or modify it under
 the terms of the GNU General Public License as published by the Free Software

storage/innobase/include/fts0types.h

@@ -126,7 +126,9 @@ struct fts_sync_t {
 	bool		in_progress;	/*!< flag whether sync is in progress.*/
 	bool		unlock_cache;	/*!< flag whether unlock cache when
 					write fts node */
-	os_event_t	event;		/*!< sync finish event */
+	os_event_t	event;		/*!< sync finish event;
+					only os_event_set() and os_event_wait()
+					are used */
 };
 
 /** The cache for the FTS system. It is a memory-based inverted index

storage/innobase/include/lock0lock.h

@@ -918,7 +918,12 @@ struct lock_sys_t{
 	srv_slot_t*	waiting_threads;	/*!< Array  of user threads
 						suspended while waiting for
 						locks within InnoDB, protected
-						by the lock_sys->wait_mutex */
+						by the lock_sys->wait_mutex;
+						os_event_set() and
+						os_event_reset() on
+						waiting_threads[]->event
+						are protected by
+						trx_t::mutex */
 	srv_slot_t*	last_slot;		/*!< highest slot ever used
 						in the waiting_threads array,
 						protected by
@@ -931,10 +936,11 @@ struct lock_sys_t{
 
 	ulint		n_lock_max_wait_time;	/*!< Max wait time */
 
-	os_event_t	timeout_event;		/*!< Set to the event that is
-						created in the lock wait monitor
-						thread. A value of 0 means the
-						thread is not active */
+	os_event_t	timeout_event;		/*!< An event waited for by
+						lock_wait_timeout_thread.
+						Not protected by a mutex,
+						but the waits are timed.
+						Signaled on shutdown only. */
 
 	bool		timeout_thread_active;	/*!< True if the timeout thread
 						is running */