MDEV-6376: InnoDB: Assertion failure in thread 139995225970432

in file buf0mtflu.cc line 570.

Analysis: Real timing bug, we should take the mutex before we
try to send those shutdown messages, that would make sure
that threads doing a unfinished flush (they have acquired
this mutex) have time to do their work before we add shutdown
messages to work queue. Currently, we just add those shutdown
messages to work queue and code assumes that at flush, there
is constant number of items to be processed and thus
leading to assertion.

2 files changed, 31 insertions, 0 deletions

diff --git a/storage/innobase/buf/buf0mtflu.cc b/storage/innobase/buf/buf0mtflu.cc
index 5a1769e3b70..ded24edc799 100644
--- a/storage/innobase/buf/buf0mtflu.cc
+++ b/storage/innobase/buf/buf0mtflu.cc
@@ -378,6 +378,20 @@ buf_mtflu_io_thread_exit(void)
 	fprintf(stderr, "InnoDB: [Note]: Signal mtflush_io_threads to exit [%lu]\n",
 		srv_mtflush_threads);
 
+	/* This lock is to safequard against timing bug: flush request take
+	this mutex before sending work items to be processed by flush
+	threads. Inside flush thread we assume that work queue contains only
+	a constant number of items. Thus, we may not install new work items
+	below before all previous ones are processed. This mutex is released
+	by flush request after all work items sent to flush threads have
+	been processed. Thus, we can get this mutex if and only if work
+	queue is empty. */
+
+	os_fast_mutex_lock(&mtflush_mtx);
+
+	/* Make sure the work queue is empty */
+	ut_a(ib_wqueue_is_empty(mtflush_io->wq));
+
 	/* Send one exit work item/thread */
 	for (i=0; i < srv_mtflush_threads; i++) {
 		work_item[i].tsk = MT_WRK_NONE;
@@ -399,6 +413,9 @@ buf_mtflu_io_thread_exit(void)
 
 	ut_a(ib_wqueue_is_empty(mtflush_io->wq));
 
+	/* Requests sent */
+	os_fast_mutex_unlock(&mtflush_mtx);
+
 	/* Collect all work done items */
 	for (i=0; i < srv_mtflush_threads;) {
 		wrk_t* work_item = NULL;
diff --git a/storage/xtradb/buf/buf0mtflu.cc b/storage/xtradb/buf/buf0mtflu.cc
index b14b83aa5d0..945bd93a1d3 100644
--- a/storage/xtradb/buf/buf0mtflu.cc
+++ b/storage/xtradb/buf/buf0mtflu.cc
@@ -385,6 +385,17 @@ buf_mtflu_io_thread_exit(void)
 	fprintf(stderr, "InnoDB: [Note]: Signal mtflush_io_threads to exit [%lu]\n",
 		srv_mtflush_threads);
 
+	/* This lock is to safequard against timing bug: flush request take
+	this mutex before sending work items to be processed by flush
+	threads. Inside flush thread we assume that work queue contains only
+	a constant number of items. Thus, we may not install new work items
+	below before all previous ones are processed. This mutex is released
+	by flush request after all work items sent to flush threads have
+	been processed. Thus, we can get this mutex if and only if work
+	queue is empty. */
+
+	os_fast_mutex_lock(&mtflush_mtx);
+
 	/* Send one exit work item/thread */
 	for (i=0; i < srv_mtflush_threads; i++) {
 		work_item[i].tsk = MT_WRK_NONE;
@@ -406,6 +417,9 @@ buf_mtflu_io_thread_exit(void)
 
 	ut_a(ib_wqueue_is_empty(mtflush_io->wq));
 
+	/* Requests sent */
+	os_fast_mutex_unlock(&mtflush_mtx);
+
 	/* Collect all work done items */
 	for (i=0; i < srv_mtflush_threads;) {
 		wrk_t* work_item = NULL;