Add new scheduling algorithm for reducing tail latencies

storage/innobase/handler/ha_innodb.cc

@@ -282,6 +282,22 @@ static TYPELIB innodb_checksum_algorithm_typelib = {
 	NULL
 };
 
+/** Possible values of the parameter innodb_lock_schedule_algorithm */
+static const char* innodb_lock_schedule_algorithm_names[] = {
+	"fcfs",
+	"vats",
+	NullS
+};
+
+/** Used to define an enumerate type of the system variable
+innodb_lock_schedule_algorithm. */
+static TYPELIB innodb_lock_schedule_algorithm_typelib = {
+	array_elements(innodb_lock_schedule_algorithm_names) - 1,
+	"innodb_lock_schedule_algorithm_typelib",
+	innodb_lock_schedule_algorithm_names,
+	NULL
+};
+
 /** Possible values for system variable "innodb_default_row_format". */
 static const char* innodb_default_row_format_names[] = {
 	"redundant",
@@ -18492,6 +18508,24 @@ static MYSQL_SYSVAR_ENUM(checksum_algorithm, srv_checksum_algorithm,
   NULL, NULL, SRV_CHECKSUM_ALGORITHM_CRC32,
   &innodb_checksum_algorithm_typelib);
 
+static MYSQL_SYSVAR_ENUM(lock_schedule_algorithm, innodb_lock_schedule_algorithm,
+  PLUGIN_VAR_RQCMDARG,
+  "The algorithm Innodb uses for deciding which locks to grant next when"
+  " a lock is released. Possible values are"
+  " FCFS"
+    " grant the locks in First-Come-First-Served order;"
+  " VATS"
+    " use the Variance-Aware-Transaction-Scheduling algorithm, which"
+    " chooses between First-Come-First-Served and Eldest-Transaction-First.",
+  NULL, NULL, INNODB_LOCK_SCHEDULE_ALGORITHM_FCFS,
+  &innodb_lock_schedule_algorithm_typelib);
+
+
+static MYSQL_SYSVAR_DOUBLE(vats_wait_lock_pct, innodb_vats_wait_lock_pct,
+  PLUGIN_VAR_RQCMDARG,
+  "Percentage of wait locks to trigger Eldest-Transaction-First.",
+  NULL, NULL, 0.0, 0, 99.999, 0);
+
 static MYSQL_SYSVAR_BOOL(log_checksums, innodb_log_checksums,
   PLUGIN_VAR_RQCMDARG,
   "Whether to compute and require checksums for InnoDB redo log blocks",
@@ -19457,6 +19491,8 @@ static struct st_mysql_sys_var* innobase_system_variables[]= {
   MYSQL_SYSVAR(ft_sort_pll_degree),
   MYSQL_SYSVAR(large_prefix),
   MYSQL_SYSVAR(force_load_corrupted),
+  MYSQL_SYSVAR(lock_schedule_algorithm),
+  MYSQL_SYSVAR(vats_wait_lock_pct),
   MYSQL_SYSVAR(locks_unsafe_for_binlog),
   MYSQL_SYSVAR(lock_wait_timeout),
   MYSQL_SYSVAR(page_size),

storage/innobase/include/lock0lock.h

@@ -43,6 +43,16 @@ Created 5/7/1996 Heikki Tuuri
 // Forward declaration
 class ReadView;
 
+/** Alternatives for innodb_lock_schedule_algorithm, which can be changed by
+ setting innodb_lock_schedule_algorithm */
+enum innodb_lock_schedule_algorithm_t {
+    INNODB_LOCK_SCHEDULE_ALGORITHM_FCFS,		/*!< First Come First Served */
+    INNODB_LOCK_SCHEDULE_ALGORITHM_VATS         /*!< Variance-Aware-Transaction-Scheduling */
+};
+
+extern ulong innodb_lock_schedule_algorithm;
+extern double innodb_vats_wait_lock_pct;
+
 /*********************************************************************//**
 Gets the size of a lock struct.
 @return size in bytes */

storage/innobase/include/lock0priv.ic

@@ -207,6 +207,30 @@ lock_rec_get_next_const(
 	return(lock_rec_get_next(heap_no, (lock_t*) lock));
 }
 
+/*********************************************************************//**
+Gets the first explicit lock request on a record.
+@return	first lock, NULL if none exists */
+UNIV_INLINE
+lock_t*
+lock_rec_get_first(
+/*===============*/
+	hash_table_t*		hash,	/*!< in: hash chain the lock on */
+	ulint           space_id,	/*!< in: space id of the record */
+	ulint           page_no,    /*!< in: page number of the record */
+	ulint			heap_no)    /*!< in: heap number of the record */
+{
+	ut_ad(lock_mutex_own());
+
+	for (lock_t* lock = lock_rec_get_first_on_page_addr(hash, space_id, page_no); lock;
+	     lock = lock_rec_get_next_on_page(lock)) {
+		if (lock_rec_get_nth_bit(lock, heap_no)) {
+			return(lock);
+		}
+	}
+
+	return(NULL);
+}
+
 /*********************************************************************//**
 Gets the first explicit lock request on a record.
 @return	first lock, NULL if none exists */