MDEV-532: Async InnoDB commit checkpoint.

Make the commit checkpoint inside InnoDB be asynchroneous.
Implement a background thread in binlog to do the writing and flushing of
binlog checkpoint events to disk.

mysql-test/suite/binlog/r/binlog_checkpoint.result

@@ -70,8 +70,14 @@ show binlog events in 'master-bin.000003' from <binlog_start>;
 Log_name	Pos	Event_type	Server_id	End_log_pos	Info
 master-bin.000003	#	Format_desc	#	#	SERVER_VERSION, BINLOG_VERSION
 master-bin.000003	#	Binlog_checkpoint	#	#	master-bin.000001
+SET DEBUG_SYNC= "RESET";
+SET @old_dbug= @@global.DEBUG_DBUG;
+SET GLOBAL debug_dbug="+d,binlog_background_checkpoint_processed";
 SET DEBUG_SYNC= "now SIGNAL con2_continue";
 con1 is still pending, no new binlog checkpoint should have been logged.
+SET DEBUG_SYNC= "now WAIT_FOR binlog_background_checkpoint_processed";
+SET GLOBAL debug_dbug= @old_dbug;
+SET DEBUG_SYNC= "RESET";
 show binlog events in 'master-bin.000003' from <binlog_start>;
 Log_name	Pos	Event_type	Server_id	End_log_pos	Info
 master-bin.000003	#	Format_desc	#	#	SERVER_VERSION, BINLOG_VERSION

mysql-test/suite/binlog/r/binlog_xa_recover.result

@@ -118,7 +118,11 @@ master-bin.00000<binlog_start>	#	Table_map	#	#	table_id: # (test.t1)
 master-bin.00000<binlog_start>	#	Write_rows	#	#	table_id: # flags: STMT_END_F
 master-bin.00000<binlog_start>	#	Xid	#	#	COMMIT /* XID */
 SET DEBUG_SYNC= "now SIGNAL con10_cont";
+SET @old_dbug= @@global.DEBUG_DBUG;
+SET GLOBAL debug_dbug="+d,binlog_background_checkpoint_processed";
 SET DEBUG_SYNC= "now SIGNAL con12_cont";
+SET DEBUG_SYNC= "now WAIT_FOR binlog_background_checkpoint_processed";
+SET GLOBAL debug_dbug= @old_dbug;
 SET DEBUG_SYNC= "now SIGNAL con11_cont";
 Checking that master-bin.000004 is the last binlog checkpoint
 show binlog events in 'master-bin.00000<binlog_start>' from <binlog_start>;

mysql-test/suite/binlog/t/binlog_checkpoint.test

@@ -71,6 +71,12 @@ SET DEBUG_SYNC= "now WAIT_FOR con2_ready";
 --let $binlog_file= master-bin.000003
 --source include/show_binlog_events.inc
 
+# We need to sync the test case with the background processing of the
+# commit checkpoint, otherwise we get nondeterministic results.
+SET DEBUG_SYNC= "RESET";
+SET @old_dbug= @@global.DEBUG_DBUG;
+SET GLOBAL debug_dbug="+d,binlog_background_checkpoint_processed";
+
 SET DEBUG_SYNC= "now SIGNAL con2_continue";
 
 connection con2;
@@ -78,6 +84,12 @@ reap;
 
 connection default;
 --echo con1 is still pending, no new binlog checkpoint should have been logged.
+# Make sure commit checkpoint is processed before we check that no checkpoint
+# event has been binlogged.
+SET DEBUG_SYNC= "now WAIT_FOR binlog_background_checkpoint_processed";
+SET GLOBAL debug_dbug= @old_dbug;
+SET DEBUG_SYNC= "RESET";
+
 --let $binlog_file= master-bin.000003
 --source include/show_binlog_events.inc
 

mysql-test/suite/binlog/t/binlog_xa_recover.test

@@ -14,8 +14,24 @@ CREATE TABLE t1 (a INT PRIMARY KEY, b MEDIUMTEXT) ENGINE=Innodb;
 # Insert some data to force a couple binlog rotations (3), so we get some
 # normal binlog checkpoints before starting the test.
 INSERT INTO t1 VALUES (100, REPEAT("x", 4100));
+# Wait for the master-bin.000002 binlog checkpoint to appear.
+--let $wait_for_all= 0
+--let $show_statement= SHOW BINLOG EVENTS IN "master-bin.000002"
+--let $field= Info
+--let $condition= = "master-bin.000002"
+--source include/wait_show_condition.inc
 INSERT INTO t1 VALUES (101, REPEAT("x", 4100));
+--let $wait_for_all= 0
+--let $show_statement= SHOW BINLOG EVENTS IN "master-bin.000003"
+--let $field= Info
+--let $condition= = "master-bin.000003"
+--source include/wait_show_condition.inc
 INSERT INTO t1 VALUES (102, REPEAT("x", 4100));
+--let $wait_for_all= 0
+--let $show_statement= SHOW BINLOG EVENTS IN "master-bin.000004"
+--let $field= Info
+--let $condition= = "master-bin.000004"
+--source include/wait_show_condition.inc
 
 # Now start a bunch of transactions that span multiple binlog
 # files. Leave then in the state prepared-but-not-committed in the engine
@@ -153,10 +169,19 @@ SET DEBUG_SYNC= "now SIGNAL con10_cont";
 connection con10;
 reap;
 connection default;
+
+# We need to sync the test case with the background processing of the
+# commit checkpoint, otherwise we get nondeterministic results.
+SET @old_dbug= @@global.DEBUG_DBUG;
+SET GLOBAL debug_dbug="+d,binlog_background_checkpoint_processed";
+
 SET DEBUG_SYNC= "now SIGNAL con12_cont";
 connection con12;
 reap;
 connection default;
+SET DEBUG_SYNC= "now WAIT_FOR binlog_background_checkpoint_processed";
+SET GLOBAL debug_dbug= @old_dbug;
+
 SET DEBUG_SYNC= "now SIGNAL con11_cont";
 connection con11;
 reap;
@@ -210,7 +235,20 @@ RESET MASTER;
 # crash recovery fails due to the error insert used for previous test.
 INSERT INTO t1 VALUES (21, REPEAT("x", 4100));
 INSERT INTO t1 VALUES (22, REPEAT("x", 4100));
+# Wait for the master-bin.000003 binlog checkpoint to appear.
+--let $wait_for_all= 0
+--let $show_statement= SHOW BINLOG EVENTS IN "master-bin.000003"
+--let $field= Info
+--let $condition= = "master-bin.000003"
+--source include/wait_show_condition.inc
 INSERT INTO t1 VALUES (23, REPEAT("x", 4100));
+# Wait for the last (master-bin.000004) binlog checkpoint to appear.
+--let $wait_for_all= 0
+--let $show_statement= SHOW BINLOG EVENTS IN "master-bin.000004"
+--let $field= Info
+--let $condition= = "master-bin.000004"
+--source include/wait_show_condition.inc
+
 --write_file $MYSQLTEST_VARDIR/tmp/mysqld.1.expect
 wait-binlog_xa_recover.test
 EOF

mysql-test/suite/perfschema/r/all_instances.result

@@ -76,6 +76,7 @@ wait/synch/mutex/sql/Master_info::run_lock
 wait/synch/mutex/sql/Master_info::sleep_lock
 wait/synch/mutex/sql/MDL_map::mutex
 wait/synch/mutex/sql/MDL_wait::LOCK_wait_status
+wait/synch/mutex/sql/MYSQL_BIN_LOG::LOCK_binlog_background_thread
 wait/synch/mutex/sql/MYSQL_BIN_LOG::LOCK_index
 wait/synch/mutex/sql/MYSQL_BIN_LOG::LOCK_xid_list
 wait/synch/mutex/sql/MYSQL_RELAY_LOG::LOCK_index
@@ -129,6 +130,8 @@ wait/synch/cond/sql/Master_info::sleep_cond
 wait/synch/cond/sql/Master_info::start_cond
 wait/synch/cond/sql/Master_info::stop_cond
 wait/synch/cond/sql/MDL_context::COND_wait_status
+wait/synch/cond/sql/MYSQL_BIN_LOG::COND_binlog_background_thread
+wait/synch/cond/sql/MYSQL_BIN_LOG::COND_binlog_background_thread_end
 wait/synch/cond/sql/MYSQL_BIN_LOG::COND_queue_busy
 wait/synch/cond/sql/MYSQL_BIN_LOG::COND_xid_list
 wait/synch/cond/sql/MYSQL_BIN_LOG::update_cond

mysql-test/suite/perfschema/r/relaylog.result

@@ -56,8 +56,11 @@ where event_name like "%MYSQL_BIN_LOG%"
   and event_name not like "%MYSQL_BIN_LOG::update_cond"
   order by event_name;
 EVENT_NAME	COUNT_STAR
+wait/synch/cond/sql/MYSQL_BIN_LOG::COND_binlog_background_thread	NONE
+wait/synch/cond/sql/MYSQL_BIN_LOG::COND_binlog_background_thread_end	NONE
 wait/synch/cond/sql/MYSQL_BIN_LOG::COND_queue_busy	NONE
 wait/synch/cond/sql/MYSQL_BIN_LOG::COND_xid_list	NONE
+wait/synch/mutex/sql/MYSQL_BIN_LOG::LOCK_binlog_background_thread	MANY
 wait/synch/mutex/sql/MYSQL_BIN_LOG::LOCK_index	MANY
 wait/synch/mutex/sql/MYSQL_BIN_LOG::LOCK_xid_list	MANY
 "Expect no slave relay log"
@@ -131,8 +134,11 @@ where event_name like "%MYSQL_BIN_LOG%"
   and event_name not like "%MYSQL_BIN_LOG::update_cond"
   order by event_name;
 EVENT_NAME	COUNT_STAR
+wait/synch/cond/sql/MYSQL_BIN_LOG::COND_binlog_background_thread	MANY
+wait/synch/cond/sql/MYSQL_BIN_LOG::COND_binlog_background_thread_end	NONE
 wait/synch/cond/sql/MYSQL_BIN_LOG::COND_queue_busy	NONE
-wait/synch/cond/sql/MYSQL_BIN_LOG::COND_xid_list	NONE
+wait/synch/cond/sql/MYSQL_BIN_LOG::COND_xid_list	MANY
+wait/synch/mutex/sql/MYSQL_BIN_LOG::LOCK_binlog_background_thread	MANY
 wait/synch/mutex/sql/MYSQL_BIN_LOG::LOCK_index	MANY
 wait/synch/mutex/sql/MYSQL_BIN_LOG::LOCK_xid_list	MANY
 "Expect a slave relay log"

sql/debug_sync.cc

@@ -984,6 +984,7 @@ static bool debug_sync_eval_action(THD *thd, char *action_str)
   DBUG_ENTER("debug_sync_eval_action");
   DBUG_ASSERT(thd);
   DBUG_ASSERT(action_str);
+  DBUG_PRINT("debug_sync", ("action_str='%s'", action_str));
 
   /*
     Get debug sync point name. Or a special command.

sql/log.cc

@@ -54,6 +54,7 @@
 #include "rpl_handler.h"
 #include "debug_sync.h"
 #include "sql_show.h"
+#include "my_pthread.h"
 
 /* max size of the log message */
 #define MAX_LOG_BUFFER_SIZE 1024
@@ -107,6 +108,17 @@ static SHOW_VAR binlog_status_vars_detail[]=
   {NullS, NullS, SHOW_LONG}
 };
 
+/*
+  Variables for the binlog background thread.
+  Protected by the MYSQL_BIN_LOG::LOCK_binlog_background_thread mutex.
+ */
+static bool binlog_background_thread_started= false;
+static bool binlog_background_thread_stop= false;
+static MYSQL_BIN_LOG::xid_count_per_binlog *
+    binlog_background_thread_queue= NULL;
+
+static bool start_binlog_background_thread();
+
 
 /**
    purge logs, master and slave sides both, related error code
@@ -2958,12 +2970,28 @@ void MYSQL_BIN_LOG::cleanup()
       my_free(b);
     }
 
+    /* Wait for the binlog background thread to stop. */
+    if (!is_relay_log && binlog_background_thread_started)
+    {
+      mysql_mutex_lock(&LOCK_binlog_background_thread);
+      binlog_background_thread_stop= true;
+      mysql_cond_signal(&COND_binlog_background_thread);
+      while (binlog_background_thread_stop)
+        mysql_cond_wait(&COND_binlog_background_thread_end,
+                        &LOCK_binlog_background_thread);
+      mysql_mutex_unlock(&LOCK_binlog_background_thread);
+      binlog_background_thread_started= false;
+    }
+
     mysql_mutex_destroy(&LOCK_log);
     mysql_mutex_destroy(&LOCK_index);
     mysql_mutex_destroy(&LOCK_xid_list);
+    mysql_mutex_destroy(&LOCK_binlog_background_thread);
     mysql_cond_destroy(&update_cond);
     mysql_cond_destroy(&COND_queue_busy);
     mysql_cond_destroy(&COND_xid_list);
+    mysql_cond_destroy(&COND_binlog_background_thread);
+    mysql_cond_destroy(&COND_binlog_background_thread_end);
   }
   DBUG_VOID_RETURN;
 }
@@ -2989,6 +3017,13 @@ void MYSQL_BIN_LOG::init_pthread_objects()
   mysql_cond_init(m_key_update_cond, &update_cond, 0);
   mysql_cond_init(m_key_COND_queue_busy, &COND_queue_busy, 0);
   mysql_cond_init(key_BINLOG_COND_xid_list, &COND_xid_list, 0);
+
+  mysql_mutex_init(key_BINLOG_LOCK_binlog_background_thread,
+                   &LOCK_binlog_background_thread, MY_MUTEX_INIT_FAST);
+  mysql_cond_init(key_BINLOG_COND_binlog_background_thread,
+                  &COND_binlog_background_thread, 0);
+  mysql_cond_init(key_BINLOG_COND_binlog_background_thread_end,
+                  &COND_binlog_background_thread_end, 0);
 }
 
 
@@ -3086,6 +3121,10 @@ bool MYSQL_BIN_LOG::open(const char *log_name,
   DBUG_ENTER("MYSQL_BIN_LOG::open");
   DBUG_PRINT("enter",("log_type: %d",(int) log_type_arg));
 
+  if (!is_relay_log && !binlog_background_thread_started &&
+      start_binlog_background_thread())
+    DBUG_RETURN(1);
+
   if (init_and_set_log_file_name(log_name, new_name, log_type_arg,
                                  io_cache_type_arg))
   {
@@ -5541,11 +5580,7 @@ bool general_log_write(THD *thd, enum enum_server_command command,
 }
 
 
-/*
-  I would like to make this function static, but this causes compiler warnings
-  when it is declared as friend function in log.h.
-*/
-void
+static void
 binlog_checkpoint_callback(void *cookie)
 {
   MYSQL_BIN_LOG::xid_count_per_binlog *entry=
@@ -8135,9 +8170,129 @@ int TC_LOG_BINLOG::unlog(ulong cookie, my_xid xid)
 void
 TC_LOG_BINLOG::commit_checkpoint_notify(void *cookie)
 {
-  mark_xid_done(((xid_count_per_binlog *)cookie)->binlog_id, true);
+  xid_count_per_binlog *entry= static_cast<xid_count_per_binlog *>(cookie);
+  mysql_mutex_lock(&LOCK_binlog_background_thread);
+  entry->next_in_queue= binlog_background_thread_queue;
+  binlog_background_thread_queue= entry;
+  mysql_cond_signal(&COND_binlog_background_thread);
+  mysql_mutex_unlock(&LOCK_binlog_background_thread);
 }
 
+/*
+  Binlog background thread.
+
+  This thread is used to log binlog checkpoints in the background, rather than
+  in the context of random storage engine threads that happen to call
+  commit_checkpoint_notify_ha() and may not like the delays while syncing
+  binlog to disk or may not be setup with all my_thread_init() and other
+  necessary stuff.
+
+  In the future, this thread could also be used to do log rotation in the
+  background, which could elimiate all stalls around binlog rotations.
+*/
+pthread_handler_t
+binlog_background_thread(void *arg __attribute__((unused)))
+{
+  bool stop;
+  MYSQL_BIN_LOG::xid_count_per_binlog *queue, *next;
+  THD *thd;
+
+  my_thread_init();
+  thd= new THD;
+  thd->system_thread= SYSTEM_THREAD_BINLOG_BACKGROUND;
+  thd->thread_stack= (char*) &thd;           /* Set approximate stack start */
+  mysql_mutex_lock(&LOCK_thread_count);
+  thd->thread_id= thread_id++;
+  mysql_mutex_unlock(&LOCK_thread_count);
+  thd->store_globals();
+
+  for (;;)
+  {
+    /*
+      Wait until there is something in the queue to process, or we are asked
+      to shut down.
+    */
+    thd_proc_info(thd, "Waiting for background binlog tasks");
+    mysql_mutex_lock(&mysql_bin_log.LOCK_binlog_background_thread);
+    for (;;)
+    {
+      stop= binlog_background_thread_stop;
+      queue= binlog_background_thread_queue;
+      if (stop || queue)
+        break;
+      mysql_cond_wait(&mysql_bin_log.COND_binlog_background_thread,
+                      &mysql_bin_log.LOCK_binlog_background_thread);
+    }
+    /* Grab the queue, if any. */
+    binlog_background_thread_queue= NULL;
+    mysql_mutex_unlock(&mysql_bin_log.LOCK_binlog_background_thread);
+
+    /* Process any incoming commit_checkpoint_notify() calls. */
+    while (queue)
+    {
+      thd_proc_info(thd, "Processing binlog checkpoint notification");
+      /* Grab next pointer first, as mark_xid_done() may free the element. */
+      next= queue->next_in_queue;
+      mysql_bin_log.mark_xid_done(queue->binlog_id, true);
+      queue= next;
+
+      DBUG_EXECUTE_IF("binlog_background_checkpoint_processed",
+        DBUG_ASSERT(!debug_sync_set_action(
+          thd,
+          STRING_WITH_LEN("now SIGNAL binlog_background_checkpoint_processed")));
+        );
+    }
+
+    if (stop)
+      break;
+  }
+
+  thd_proc_info(thd, "Stopping binlog background thread");
+
+  mysql_mutex_lock(&LOCK_thread_count);
+  delete thd;
+  mysql_mutex_unlock(&LOCK_thread_count);
+
+  my_thread_end();
+
+  /* Signal that we are (almost) stopped. */
+  mysql_mutex_lock(&mysql_bin_log.LOCK_binlog_background_thread);
+  binlog_background_thread_stop= false;
+  mysql_cond_signal(&mysql_bin_log.COND_binlog_background_thread_end);
+  mysql_mutex_unlock(&mysql_bin_log.LOCK_binlog_background_thread);
+
+  return 0;
+}
+
+#ifdef HAVE_PSI_INTERFACE
+static PSI_thread_key key_thread_binlog;
+
+static PSI_thread_info all_binlog_threads[]=
+{
+  { &key_thread_binlog, "binlog_background", PSI_FLAG_GLOBAL},
+};
+#endif /* HAVE_PSI_INTERFACE */
+
+static bool
+start_binlog_background_thread()
+{
+  pthread_t th;
+
+#ifdef HAVE_PSI_INTERFACE
+  if (PSI_server)
+    PSI_server->register_thread("sql", all_binlog_threads,
+                                array_elements(all_binlog_threads));
+#endif
+
+  if (mysql_thread_create(key_thread_binlog, &th, NULL,
+                          binlog_background_thread, NULL))
+    return 1;
+
+  binlog_background_thread_started= true;
+  return 0;
+}
+
+
 int TC_LOG_BINLOG::recover(LOG_INFO *linfo, const char *last_log_name,
                            IO_CACHE *first_log,
                            Format_description_log_event *fdle)

sql/log.h

@@ -395,8 +395,6 @@ private:
 #define BINLOG_COOKIE_IS_DUMMY(c) \
   ( ((ulong)(c)>>1) == BINLOG_COOKIE_DUMMY_ID )
 
-void binlog_checkpoint_callback(void *cookie);
-
 class binlog_cache_mngr;
 class MYSQL_BIN_LOG: public TC_LOG, private MYSQL_LOG
 {
@@ -450,27 +448,6 @@ class MYSQL_BIN_LOG: public TC_LOG, private MYSQL_LOG
     ulong binlog_id;
   };
 
-  /*
-    A list of struct xid_count_per_binlog is used to keep track of how many
-    XIDs are in prepared, but not committed, state in each binlog. And how
-    many commit_checkpoint_request()'s are pending.
-
-    When count drops to zero in a binlog after rotation, it means that there
-    are no more XIDs in prepared state, so that binlog is no longer needed
-    for XA crash recovery, and we can log a new binlog checkpoint event.
-
-    The list is protected against simultaneous access from multiple
-    threads by LOCK_xid_list.
-  */
-  struct xid_count_per_binlog : public ilink {
-    char *binlog_name;
-    uint binlog_name_len;
-    ulong binlog_id;
-    /* Total prepared XIDs and pending checkpoint requests in this binlog. */
-    long xid_count;
-    xid_count_per_binlog();   /* Give link error if constructor used. */
-  };
-  I_List<xid_count_per_binlog> binlog_xid_count_list;
   /*
     When this is set, a RESET MASTER is in progress.
 
@@ -480,7 +457,6 @@ class MYSQL_BIN_LOG: public TC_LOG, private MYSQL_LOG
     checkpoint arrives - when all have arrived, RESET MASTER will complete.
   */
   bool reset_master_pending;
-  friend void binlog_checkpoint_callback(void *cookie);
 
   /* LOCK_log and LOCK_index are inited by init_pthread_objects() */
   mysql_mutex_t LOCK_index;
@@ -550,10 +526,35 @@ class MYSQL_BIN_LOG: public TC_LOG, private MYSQL_LOG
   int write_transaction_or_stmt(group_commit_entry *entry);
   bool write_transaction_to_binlog_events(group_commit_entry *entry);
   void trx_group_commit_leader(group_commit_entry *leader);
-  void mark_xid_done(ulong cookie, bool write_checkpoint);
-  void mark_xids_active(ulong cookie, uint xid_count);
 
 public:
+  /*
+    A list of struct xid_count_per_binlog is used to keep track of how many
+    XIDs are in prepared, but not committed, state in each binlog. And how
+    many commit_checkpoint_request()'s are pending.
+
+    When count drops to zero in a binlog after rotation, it means that there
+    are no more XIDs in prepared state, so that binlog is no longer needed
+    for XA crash recovery, and we can log a new binlog checkpoint event.
+
+    The list is protected against simultaneous access from multiple
+    threads by LOCK_xid_list.
+  */
+  struct xid_count_per_binlog : public ilink {
+    char *binlog_name;
+    uint binlog_name_len;
+    ulong binlog_id;
+    /* Total prepared XIDs and pending checkpoint requests in this binlog. */
+    long xid_count;
+    /* For linking in requests to the binlog background thread. */
+    xid_count_per_binlog *next_in_queue;
+    xid_count_per_binlog();   /* Give link error if constructor used. */
+  };
+  I_List<xid_count_per_binlog> binlog_xid_count_list;
+  mysql_mutex_t LOCK_binlog_background_thread;
+  mysql_cond_t COND_binlog_background_thread;
+  mysql_cond_t COND_binlog_background_thread_end;
+
   using MYSQL_LOG::generate_name;
   using MYSQL_LOG::is_open;
 
@@ -709,6 +710,8 @@ public:
   bool appendv(const char* buf,uint len,...);
   bool append(Log_event* ev);
 
+  void mark_xids_active(ulong cookie, uint xid_count);
+  void mark_xid_done(ulong cookie, bool write_checkpoint);
   void make_log_name(char* buf, const char* log_ident);
   bool is_active(const char* log_file_name);
   bool can_purge_log(const char *log_file_name);

sql/mysqld.cc

@@ -726,6 +726,7 @@ PSI_mutex_key key_LOCK_des_key_file;
 #endif /* HAVE_OPENSSL */
 
 PSI_mutex_key key_BINLOG_LOCK_index, key_BINLOG_LOCK_xid_list,
+  key_BINLOG_LOCK_binlog_background_thread,
   key_delayed_insert_mutex, key_hash_filo_lock, key_LOCK_active_mi,
   key_LOCK_connection_count, key_LOCK_crypt, key_LOCK_delayed_create,
   key_LOCK_delayed_insert, key_LOCK_delayed_status, key_LOCK_error_log,
@@ -768,6 +769,7 @@ static PSI_mutex_info all_server_mutexes[]=
 
   { &key_BINLOG_LOCK_index, "MYSQL_BIN_LOG::LOCK_index", 0},
   { &key_BINLOG_LOCK_xid_list, "MYSQL_BIN_LOG::LOCK_xid_list", 0},
+  { &key_BINLOG_LOCK_binlog_background_thread, "MYSQL_BIN_LOG::LOCK_binlog_background_thread", 0},
   { &key_RELAYLOG_LOCK_index, "MYSQL_RELAY_LOG::LOCK_index", 0},
   { &key_delayed_insert_mutex, "Delayed_insert::mutex", 0},
   { &key_hash_filo_lock, "hash_filo::lock", 0},
@@ -836,6 +838,8 @@ PSI_cond_key key_PAGE_cond, key_COND_active, key_COND_pool;
 #endif /* HAVE_MMAP */
 
 PSI_cond_key key_BINLOG_COND_xid_list, key_BINLOG_update_cond,
+  key_BINLOG_COND_binlog_background_thread,
+  key_BINLOG_COND_binlog_background_thread_end,
   key_COND_cache_status_changed, key_COND_manager,
   key_COND_rpl_status, key_COND_server_started,
   key_delayed_insert_cond, key_delayed_insert_cond_client,
@@ -865,6 +869,8 @@ static PSI_cond_info all_server_conds[]=
 #endif /* HAVE_MMAP */
   { &key_BINLOG_COND_xid_list, "MYSQL_BIN_LOG::COND_xid_list", 0},
   { &key_BINLOG_update_cond, "MYSQL_BIN_LOG::update_cond", 0},
+  { &key_BINLOG_COND_binlog_background_thread, "MYSQL_BIN_LOG::COND_binlog_background_thread", 0},
+  { &key_BINLOG_COND_binlog_background_thread_end, "MYSQL_BIN_LOG::COND_binlog_background_thread_end", 0},
   { &key_BINLOG_COND_queue_busy, "MYSQL_BIN_LOG::COND_queue_busy", 0},
   { &key_RELAYLOG_update_cond, "MYSQL_RELAY_LOG::update_cond", 0},
   { &key_RELAYLOG_COND_queue_busy, "MYSQL_RELAY_LOG::COND_queue_busy", 0},

sql/mysqld.h

@@ -228,6 +228,7 @@ extern PSI_mutex_key key_LOCK_des_key_file;
 #endif
 
 extern PSI_mutex_key key_BINLOG_LOCK_index, key_BINLOG_LOCK_xid_list,
+  key_BINLOG_LOCK_binlog_background_thread,
   key_delayed_insert_mutex, key_hash_filo_lock, key_LOCK_active_mi,
   key_LOCK_connection_count, key_LOCK_crypt, key_LOCK_delayed_create,
   key_LOCK_delayed_insert, key_LOCK_delayed_status, key_LOCK_error_log,
@@ -259,6 +260,8 @@ extern PSI_cond_key key_PAGE_cond, key_COND_active, key_COND_pool;
 #endif /* HAVE_MMAP */
 
 extern PSI_cond_key key_BINLOG_COND_xid_list, key_BINLOG_update_cond,
+  key_BINLOG_COND_binlog_background_thread,
+  key_BINLOG_COND_binlog_background_thread_end,
   key_COND_cache_status_changed, key_COND_manager,
   key_COND_rpl_status, key_COND_server_started,
   key_delayed_insert_cond, key_delayed_insert_cond_client,

sql/rpl_rli.cc

@@ -58,6 +58,7 @@ Relay_log_info::Relay_log_info(bool is_slave_recovery)
 {
   DBUG_ENTER("Relay_log_info::Relay_log_info");
 
+  relay_log.is_relay_log= TRUE;
 #ifdef HAVE_PSI_INTERFACE
   relay_log.set_psi_keys(key_RELAYLOG_LOCK_index,
                          key_RELAYLOG_update_cond,
@@ -216,8 +217,6 @@ a file name for --relay-log-index option", opt_relaylog_index_name);
                               &mi->connection_name);
     }
 
-    rli->relay_log.is_relay_log= TRUE;
-
     /*
       note, that if open() fails, we'll still have index file open
       but a destructor will take care of that

sql/sql_class.h

@@ -1255,7 +1255,8 @@ enum enum_thread_type
   SYSTEM_THREAD_SLAVE_SQL= 4,
   SYSTEM_THREAD_NDBCLUSTER_BINLOG= 8,
   SYSTEM_THREAD_EVENT_SCHEDULER= 16,
-  SYSTEM_THREAD_EVENT_WORKER= 32
+  SYSTEM_THREAD_EVENT_WORKER= 32,
+  SYSTEM_THREAD_BINLOG_BACKGROUND= 64
 };
 
 inline char const *

storage/innobase/handler/ha_innodb.cc

@@ -106,6 +106,7 @@ static ulong commit_threads = 0;
 static mysql_mutex_t commit_threads_m;
 static mysql_cond_t commit_cond;
 static mysql_mutex_t commit_cond_m;
+static mysql_mutex_t pending_checkpoint_mutex;
 static bool innodb_inited = 0;
 
 #define INSIDE_HA_INNOBASE_CC
@@ -222,11 +223,13 @@ static mysql_pfs_key_t	innobase_share_mutex_key;
 static mysql_pfs_key_t	commit_threads_m_key;
 static mysql_pfs_key_t	commit_cond_mutex_key;
 static mysql_pfs_key_t	commit_cond_key;
+static mysql_pfs_key_t	pending_checkpoint_mutex_key;
 
 static PSI_mutex_info	all_pthread_mutexes[] = {
         {&commit_threads_m_key, "commit_threads_m", 0},
         {&commit_cond_mutex_key, "commit_cond_mutex", 0},
-        {&innobase_share_mutex_key, "innobase_share_mutex", 0}
+        {&innobase_share_mutex_key, "innobase_share_mutex", 0},
+        {&pending_checkpoint_mutex_key, "pending_checkpoint_mutex", 0}
 };
 
 static PSI_cond_info	all_innodb_conds[] = {
@@ -2601,6 +2604,9 @@ innobase_change_buffering_inited_ok:
 	mysql_mutex_init(commit_cond_mutex_key,
 			 &commit_cond_m, MY_MUTEX_INIT_FAST);
 	mysql_cond_init(commit_cond_key, &commit_cond, NULL);
+	mysql_mutex_init(pending_checkpoint_mutex_key,
+			 &pending_checkpoint_mutex,
+			 MY_MUTEX_INIT_FAST);
 	innodb_inited= 1;
 #ifdef MYSQL_DYNAMIC_PLUGIN
 	if (innobase_hton != p) {
@@ -2648,6 +2654,7 @@ innobase_end(
 		mysql_mutex_destroy(&commit_threads_m);
 		mysql_mutex_destroy(&commit_cond_m);
 		mysql_cond_destroy(&commit_cond);
+		mysql_mutex_destroy(&pending_checkpoint_mutex);
 	}
 
 	DBUG_RETURN(err);
@@ -3017,17 +3024,145 @@ innobase_rollback_trx(
 	DBUG_RETURN(convert_error_code_to_mysql(error, 0, NULL));
 }
 
+
+struct pending_checkpoint {
+	struct pending_checkpoint *next;
+	handlerton *hton;
+	void *cookie;
+	ib_uint64_t lsn;
+};
+static struct pending_checkpoint *pending_checkpoint_list;
+static struct pending_checkpoint *pending_checkpoint_list_end;
+
 /*****************************************************************//**
 Handle a commit checkpoint request from server layer.
-We simply flush the redo log immediately and do the notify call.*/
+We put the request in a queue, so that we can notify upper layer about
+checkpoint complete when we have flushed the redo log.
+If we have already flushed all relevant redo log, we notify immediately.*/
 static
 void
 innobase_checkpoint_request(
 	handlerton *hton,
 	void *cookie)
 {
-	log_buffer_flush_to_disk();
-	commit_checkpoint_notify_ha(hton, cookie);
+	ib_uint64_t			lsn;
+	ib_uint64_t			flush_lsn;
+	struct pending_checkpoint *	entry;
+
+	/* Do the allocation outside of lock to reduce contention. The normal
+	case is that not everything is flushed, so we will need to enqueue. */
+	entry = static_cast<struct pending_checkpoint *>
+		(my_malloc(sizeof(*entry), MYF(MY_WME)));
+	if (!entry) {
+		sql_print_error("Failed to allocate %u bytes."
+				" Commit checkpoint will be skipped.",
+				static_cast<unsigned>(sizeof(*entry)));
+		return;
+	}
+
+	entry->next = NULL;
+	entry->hton = hton;
+	entry->cookie = cookie;
+
+	mysql_mutex_lock(&pending_checkpoint_mutex);
+	lsn = log_get_lsn();
+	flush_lsn = log_get_flush_lsn();
+	if (lsn > flush_lsn) {
+		/* Put the request in queue.
+		When the log gets flushed past the lsn, we will remove the
+		entry from the queue and notify the upper layer. */
+		entry->lsn = lsn;
+		if (pending_checkpoint_list_end) {
+			pending_checkpoint_list_end->next = entry;
+			/* There is no need to order the entries in the list
+			by lsn. The upper layer can accept notifications in
+			any order, and short delays in notifications do not
+			significantly impact performance. */
+		} else {
+			pending_checkpoint_list = entry;
+		}
+		pending_checkpoint_list_end = entry;
+		entry = NULL;
+	}
+	mysql_mutex_unlock(&pending_checkpoint_mutex);
+
+	if (entry) {
+		/* We are already flushed. Notify the checkpoint immediately. */
+		commit_checkpoint_notify_ha(entry->hton, entry->cookie);
+		my_free(entry);
+	}
+}
+
+/*****************************************************************//**
+Log code calls this whenever log has been written and/or flushed up
+to a new position. We use this to notify upper layer of a new commit
+checkpoint when necessary.*/
+extern "C" UNIV_INTERN
+void
+innobase_mysql_log_notify(
+/*===============*/
+	ib_uint64_t	write_lsn,	/*!< in: LSN written to log file */
+	ib_uint64_t	flush_lsn)	/*!< in: LSN flushed to disk */
+{
+	struct pending_checkpoint *	pending;
+	struct pending_checkpoint *	entry;
+	struct pending_checkpoint *	last_ready;
+
+	/* It is safe to do a quick check for NULL first without lock.
+	Even if we should race, we will at most skip one checkpoint and
+	take the next one, which is harmless. */
+	if (!pending_checkpoint_list)
+		return;
+
+	mysql_mutex_lock(&pending_checkpoint_mutex);
+	pending = pending_checkpoint_list;
+	if (!pending)
+	{
+		mysql_mutex_unlock(&pending_checkpoint_mutex);
+		return;
+	}
+
+	last_ready = NULL;
+	for (entry = pending; entry != NULL; entry = entry -> next)
+	{
+		/* Notify checkpoints up until the first entry that has not
+		been fully flushed to the redo log. Since we do not maintain
+		the list ordered, in principle there could be more entries
+		later than were also flushed. But there is no harm in
+		delaying notifications for those a bit. And in practise, the
+		list is unlikely to have more than one element anyway, as we
+		flush the redo log at least once every second. */
+		if (entry->lsn > flush_lsn)
+			break;
+		last_ready = entry;
+	}
+
+	if (last_ready)
+	{
+		/* We found some pending checkpoints that are now flushed to
+		disk. So remove them from the list. */
+		pending_checkpoint_list = entry;
+		if (!entry)
+			pending_checkpoint_list_end = NULL;
+	}
+
+	mysql_mutex_unlock(&pending_checkpoint_mutex);
+
+	if (!last_ready)
+		return;
+
+	/* Now that we have released the lock, notify upper layer about all
+	commit checkpoints that have now completed. */
+	for (;;) {
+		entry = pending;
+		pending = pending->next;
+
+		commit_checkpoint_notify_ha(entry->hton, entry->cookie);
+
+		my_free(entry);
+		if (entry == last_ready)
+			break;
+	}
 }
 
 /*****************************************************************//**

storage/innobase/include/ha_prototypes.h

@@ -136,6 +136,17 @@ innobase_mysql_print_thd(
 	uint	max_query_len);	/*!< in: max query length to print, or 0 to
 				   use the default max length */
 
+/*****************************************************************//**
+Log code calls this whenever log has been written and/or flushed up
+to a new position. We use this to notify upper layer of a new commit
+checkpoint when necessary.*/
+UNIV_INTERN
+void
+innobase_mysql_log_notify(
+/*===============*/
+	ib_uint64_t	write_lsn,	/*!< in: LSN written to log file */
+	ib_uint64_t	flush_lsn);	/*!< in: LSN flushed to disk */
+
 /**************************************************************//**
 Converts a MySQL type to an InnoDB type. Note that this function returns
 the 'mtype' of InnoDB. InnoDB differentiates between MySQL's old <= 4.1

storage/innobase/include/log0log.h

@@ -151,6 +151,13 @@ UNIV_INLINE
 ib_uint64_t
 log_get_lsn(void);
 /*=============*/
+/************************************************************//**
+Gets the last lsn that is fully flushed to disk.
+@return	last flushed lsn */
+UNIV_INLINE
+ib_uint64_t
+log_get_flush_lsn(void);
+/*=============*/
 /****************************************************************
 Gets the log group capacity. It is OK to read the value without
 holding log_sys->mutex because it is constant.

storage/innobase/include/log0log.ic

@@ -411,6 +411,25 @@ log_get_lsn(void)
 	return(lsn);
 }
 
+/************************************************************//**
+Gets the last lsn that is fully flushed to disk.
+@return	last flushed lsn */
+UNIV_INLINE
+ib_uint64_t
+log_get_flush_lsn(void)
+/*=============*/
+{
+	ib_uint64_t	lsn;
+
+	mutex_enter(&(log_sys->mutex));
+
+	lsn = log_sys->flushed_to_disk_lsn;
+
+	mutex_exit(&(log_sys->mutex));
+
+	return(lsn);
+}
+
 /****************************************************************
 Gets the log group capacity. It is OK to read the value without
 holding log_sys->mutex because it is constant.

storage/innobase/log/log0log.c

@@ -1353,6 +1353,8 @@ log_write_up_to(
 	ulint		loop_count	= 0;
 #endif /* UNIV_DEBUG */
 	ulint		unlock;
+	ib_uint64_t	write_lsn;
+	ib_uint64_t	flush_lsn;
 
 	if (recv_no_ibuf_operations) {
 		/* Recovery is running and no operations on the log files are
@@ -1530,8 +1532,13 @@ loop:
 
 	log_flush_do_unlocks(unlock);
 
+	write_lsn = log_sys->write_lsn;
+	flush_lsn = log_sys->flushed_to_disk_lsn;
+
 	mutex_exit(&(log_sys->mutex));
 
+	innobase_mysql_log_notify(write_lsn, flush_lsn);
+
 	return;
 
 do_waits:

storage/xtradb/handler/ha_innodb.cc

@@ -121,6 +121,7 @@ static ulong commit_threads = 0;
 static mysql_mutex_t commit_threads_m;
 static mysql_cond_t commit_cond;
 static mysql_mutex_t commit_cond_m;
+static mysql_mutex_t pending_checkpoint_mutex;
 static bool innodb_inited = 0;
 
 
@@ -254,11 +255,13 @@ static mysql_pfs_key_t	innobase_share_mutex_key;
 static mysql_pfs_key_t	commit_threads_m_key;
 static mysql_pfs_key_t	commit_cond_mutex_key;
 static mysql_pfs_key_t	commit_cond_key;
+static mysql_pfs_key_t	pending_checkpoint_mutex_key;
 
 static PSI_mutex_info	all_pthread_mutexes[] = {
         {&commit_threads_m_key, "commit_threads_m", 0},
         {&commit_cond_mutex_key, "commit_cond_mutex", 0},
-        {&innobase_share_mutex_key, "innobase_share_mutex", 0}
+        {&innobase_share_mutex_key, "innobase_share_mutex", 0},
+        {&pending_checkpoint_mutex_key, "pending_checkpoint_mutex", 0}
 };
 
 static PSI_cond_info	all_innodb_conds[] = {
@@ -3088,6 +3091,9 @@ skip_overwrite:
 	mysql_mutex_init(commit_cond_mutex_key,
 			 &commit_cond_m, MY_MUTEX_INIT_FAST);
 	mysql_cond_init(commit_cond_key, &commit_cond, NULL);
+	mysql_mutex_init(pending_checkpoint_mutex_key,
+			 &pending_checkpoint_mutex,
+			 MY_MUTEX_INIT_FAST);
 	innodb_inited= 1;
 #ifdef MYSQL_DYNAMIC_PLUGIN
 	if (innobase_hton != p) {
@@ -3135,6 +3141,7 @@ innobase_end(
 		mysql_mutex_destroy(&commit_threads_m);
 		mysql_mutex_destroy(&commit_cond_m);
 		mysql_cond_destroy(&commit_cond);
+		mysql_mutex_destroy(&pending_checkpoint_mutex);
 	}
 
 	DBUG_RETURN(err);
@@ -3530,17 +3537,145 @@ innobase_rollback_trx(
 	DBUG_RETURN(convert_error_code_to_mysql(error, 0, NULL));
 }
 
+
+struct pending_checkpoint {
+	struct pending_checkpoint *next;
+	handlerton *hton;
+	void *cookie;
+	ib_uint64_t lsn;
+};
+static struct pending_checkpoint *pending_checkpoint_list;
+static struct pending_checkpoint *pending_checkpoint_list_end;
+
 /*****************************************************************//**
 Handle a commit checkpoint request from server layer.
-We simply flush the redo log immediately and do the notify call.*/
+We put the request in a queue, so that we can notify upper layer about
+checkpoint complete when we have flushed the redo log.
+If we have already flushed all relevant redo log, we notify immediately.*/
 static
 void
 innobase_checkpoint_request(
 	handlerton *hton,
 	void *cookie)
 {
-	log_buffer_flush_to_disk();
-	commit_checkpoint_notify_ha(hton, cookie);
+	ib_uint64_t			lsn;
+	ib_uint64_t			flush_lsn;
+	struct pending_checkpoint *	entry;
+
+	/* Do the allocation outside of lock to reduce contention. The normal
+	case is that not everything is flushed, so we will need to enqueue. */
+	entry = static_cast<struct pending_checkpoint *>
+		(my_malloc(sizeof(*entry), MYF(MY_WME)));
+	if (!entry) {
+		sql_print_error("Failed to allocate %u bytes."
+				" Commit checkpoint will be skipped.",
+				static_cast<unsigned>(sizeof(*entry)));
+		return;
+	}
+
+	entry->next = NULL;
+	entry->hton = hton;
+	entry->cookie = cookie;
+
+	mysql_mutex_lock(&pending_checkpoint_mutex);
+	lsn = log_get_lsn();
+	flush_lsn = log_get_flush_lsn();
+	if (lsn > flush_lsn) {
+		/* Put the request in queue.
+		When the log gets flushed past the lsn, we will remove the
+		entry from the queue and notify the upper layer. */
+		entry->lsn = lsn;
+		if (pending_checkpoint_list_end) {
+			pending_checkpoint_list_end->next = entry;
+			/* There is no need to order the entries in the list
+			by lsn. The upper layer can accept notifications in
+			any order, and short delays in notifications do not
+			significantly impact performance. */
+		} else {
+			pending_checkpoint_list = entry;
+		}
+		pending_checkpoint_list_end = entry;
+		entry = NULL;
+	}
+	mysql_mutex_unlock(&pending_checkpoint_mutex);
+
+	if (entry) {
+		/* We are already flushed. Notify the checkpoint immediately. */
+		commit_checkpoint_notify_ha(entry->hton, entry->cookie);
+		my_free(entry);
+	}
+}
+
+/*****************************************************************//**
+Log code calls this whenever log has been written and/or flushed up
+to a new position. We use this to notify upper layer of a new commit
+checkpoint when necessary.*/
+extern "C" UNIV_INTERN
+void
+innobase_mysql_log_notify(
+/*===============*/
+	ib_uint64_t	write_lsn,	/*!< in: LSN written to log file */
+	ib_uint64_t	flush_lsn)	/*!< in: LSN flushed to disk */
+{
+	struct pending_checkpoint *	pending;
+	struct pending_checkpoint *	entry;
+	struct pending_checkpoint *	last_ready;
+
+	/* It is safe to do a quick check for NULL first without lock.
+	Even if we should race, we will at most skip one checkpoint and
+	take the next one, which is harmless. */
+	if (!pending_checkpoint_list)
+		return;
+
+	mysql_mutex_lock(&pending_checkpoint_mutex);
+	pending = pending_checkpoint_list;
+	if (!pending)
+	{
+		mysql_mutex_unlock(&pending_checkpoint_mutex);
+		return;
+	}
+
+	last_ready = NULL;
+	for (entry = pending; entry != NULL; entry = entry -> next)
+	{
+		/* Notify checkpoints up until the first entry that has not
+		been fully flushed to the redo log. Since we do not maintain
+		the list ordered, in principle there could be more entries
+		later than were also flushed. But there is no harm in
+		delaying notifications for those a bit. And in practise, the
+		list is unlikely to have more than one element anyway, as we
+		flush the redo log at least once every second. */
+		if (entry->lsn > flush_lsn)
+			break;
+		last_ready = entry;
+	}
+
+	if (last_ready)
+	{
+		/* We found some pending checkpoints that are now flushed to
+		disk. So remove them from the list. */
+		pending_checkpoint_list = entry;
+		if (!entry)
+			pending_checkpoint_list_end = NULL;
+	}
+
+	mysql_mutex_unlock(&pending_checkpoint_mutex);
+
+	if (!last_ready)
+		return;
+
+	/* Now that we have released the lock, notify upper layer about all
+	commit checkpoints that have now completed. */
+	for (;;) {
+		entry = pending;
+		pending = pending->next;
+
+		commit_checkpoint_notify_ha(entry->hton, entry->cookie);
+
+		my_free(entry);
+		if (entry == last_ready)
+			break;
+	}
 }
 
 /*****************************************************************//**

storage/xtradb/include/ha_prototypes.h

@@ -136,6 +136,17 @@ innobase_mysql_print_thd(
 	uint	max_query_len);	/*!< in: max query length to print, or 0 to
 				   use the default max length */
 
+/*****************************************************************//**
+Log code calls this whenever log has been written and/or flushed up
+to a new position. We use this to notify upper layer of a new commit
+checkpoint when necessary.*/
+UNIV_INTERN
+void
+innobase_mysql_log_notify(
+/*===============*/
+	ib_uint64_t	write_lsn,	/*!< in: LSN written to log file */
+	ib_uint64_t	flush_lsn);	/*!< in: LSN flushed to disk */
+
 /**************************************************************//**
 Converts a MySQL type to an InnoDB type. Note that this function returns
 the 'mtype' of InnoDB. InnoDB differentiates between MySQL's old <= 4.1

storage/xtradb/include/log0log.h

@@ -151,6 +151,13 @@ UNIV_INLINE
 ib_uint64_t
 log_get_lsn(void);
 /*=============*/
+/************************************************************//**
+Gets the last lsn that is fully flushed to disk.
+@return	last flushed lsn */
+UNIV_INLINE
+ib_uint64_t
+log_get_flush_lsn(void);
+/*=============*/
 /****************************************************************
 Gets the log group capacity. It is OK to read the value without
 holding log_sys->mutex because it is constant.

storage/xtradb/include/log0log.ic

@@ -411,6 +411,25 @@ log_get_lsn(void)
 	return(lsn);
 }
 
+/************************************************************//**
+Gets the last lsn that is fully flushed to disk.
+@return	last flushed lsn */
+UNIV_INLINE
+ib_uint64_t
+log_get_flush_lsn(void)
+/*=============*/
+{
+	ib_uint64_t	lsn;
+
+	mutex_enter(&(log_sys->mutex));
+
+	lsn = log_sys->flushed_to_disk_lsn;
+
+	mutex_exit(&(log_sys->mutex));
+
+	return(lsn);
+}
+
 /****************************************************************
 Gets the log group capacity. It is OK to read the value without
 holding log_sys->mutex because it is constant.

storage/xtradb/log/log0log.c

@@ -1390,6 +1390,8 @@ log_write_up_to(
 	ulint		loop_count	= 0;
 #endif /* UNIV_DEBUG */
 	ulint		unlock;
+	ib_uint64_t	write_lsn;
+	ib_uint64_t	flush_lsn;
 
 	if (recv_no_ibuf_operations) {
 		/* Recovery is running and no operations on the log files are
@@ -1568,8 +1570,13 @@ loop:
 
 	log_flush_do_unlocks(unlock);
 
+	write_lsn = log_sys->write_lsn;
+	flush_lsn = log_sys->flushed_to_disk_lsn;
+
 	mutex_exit(&(log_sys->mutex));
 
+	innobase_mysql_log_notify(write_lsn, flush_lsn);
+
 	return;
 
 do_waits: