wlad posted some flame graphs that show that ha_innobase::info_low() is spending time on waiting for a shared latch. It seems to me that the following is the culprit:

			space->s_lock();

			stats.delete_length = 1024

				* fsp_get_available_space_in_free_extents(

					*space);

			space->s_unlock();

This is being executed on every call for HA_STATUS_VARIABLE, even though apparently only callers with HA_STATUS_VARIABLE_EXTRA are interested in the result. This would mainly be SHOW TABLES and get_schema_tables_record().

In the flame graph, the unnecessary latch acquisition occurs in mysql_update():

  /* Update the table->file->stats.records number */

  table->file->info(HA_STATUS_VARIABLE | HA_STATUS_NO_LOCK);

  set_statistics_for_table(thd, table);