WL#8356 "Improve scalability by not acquiring unnecessary locks for internal temp tables".

Do not acquire LOCK_plugin when doing ha_lock_engine() for builtin
engines in production builds. This is possible thanks to the fact
that in such builds for builtin plugins we don't do reference counting.
To implement this optimization introduced new builtin_htons[] array
to be able easily identify builtin handlertons.

This patch solves the problem with LOCK_plugin mutex showing up as
a scalability bottleneck in workloads which often create internal
temporary tables, such as 1-table Sysbench SELECT_DISTINCT/InnoDB
test.

Author: Dmitry Lenev

sql/handler.cc

@@ -175,6 +175,12 @@ inline double log2(double x)
 */
 st_plugin_int *hton2plugin[MAX_HA];
 
+/**
+  Array allowing to check if handlerton is builtin without
+  acquiring LOCK_plugin.
+*/
+static bool builtin_htons[MAX_HA];
+
 const char *ha_resolve_storage_engine_name(const handlerton *db_type)
 {
   return db_type == NULL ? "UNKNOWN" : hton2plugin[db_type->slot]->name.str;
@@ -496,10 +502,29 @@ plugin_ref ha_lock_engine(THD *thd, const handlerton *hton)
   if (hton)
   {
     st_plugin_int **plugin= hton2plugin + hton->slot;
-    
+
 #ifdef DBUG_OFF
+    /*
+      Take a shortcut for builtin engines -- return pointer to plugin
+      without acquiring LOCK_plugin mutex. This is safe safe since such
+      plugins are not deleted until shutdown and we don't do reference
+      counting in non-debug builds for them.
+
+      Since we have reference to handlerton on our hands, this method
+      can't be called concurrently to non-builtin handlerton initialization/
+      deinitialization. So it is safe to access builtin_htons[] without
+      additional locking.
+     */
+    if (builtin_htons[hton->slot])
+      return *plugin;
+
     return my_plugin_lock(thd, plugin);
 #else
+    /*
+      We can't take shortcut in debug builds.
+      At least assert that builtin_htons[slot] is set correctly.
+    */
+    DBUG_ASSERT(builtin_htons[hton->slot] == (plugin[0]->plugin_dl == NULL));
     return my_plugin_lock(thd, &plugin);
 #endif
   }
@@ -720,6 +745,7 @@ int ha_finalize_handlerton(st_plugin_int *plugin)
     DBUG_ASSERT(hton2plugin[hton->slot] == plugin);
     DBUG_ASSERT(hton->slot < MAX_HA);
     hton2plugin[hton->slot]= NULL;
+    builtin_htons[hton->slot]= false; /* Extra correctness. */
   }
 
   my_free(hton);
@@ -817,6 +843,7 @@ int ha_initialize_handlerton(st_plugin_int *plugin)
       hton->savepoint_offset= savepoint_alloc_size;
       savepoint_alloc_size+= tmp;
       hton2plugin[hton->slot]=plugin;
+      builtin_htons[hton->slot]= (plugin->plugin_dl == NULL);
       if (hton->prepare)
         total_ha_2pc++;
       break;