Check that only the OLD data are deleted, not only the ones that are not the latest valid CP (as we can yet have saved data for future CP).

mpi/src/mpi_failure_tolerance/starpu_mpi_checkpoint_package.c

@@ -50,9 +50,9 @@ int checkpoint_package_data_del(int cp_id, int cp_inst, int rank)
 	while (checkpoint_data != _starpu_mpi_checkpoint_data_list_end(checkpoint_data_list))
 	{
 		next_checkpoint_data = _starpu_mpi_checkpoint_data_list_next(checkpoint_data);
-		if (!(checkpoint_data->cp_id==cp_id && checkpoint_data->cp_inst==cp_inst)
-//		if ((checkpoint_data->cp_id==cp_id && checkpoint_data->cp_inst==cp_inst)
-			&& checkpoint_data->rank==rank)
+		// I delete all the old data (i.e. the cp inst is strictly lower than the one of the just validated CP) only for
+		// the rank that initiated the CP
+		if (checkpoint_data->cp_inst<cp_inst && checkpoint_data->rank==rank)
 		{
 			if (checkpoint_data->type==STARPU_R)
 			{
@@ -64,6 +64,7 @@ int checkpoint_package_data_del(int cp_id, int cp_inst, int rank)
 				free(checkpoint_data->ptr);
 			}
 			_starpu_mpi_checkpoint_data_list_erase(checkpoint_data_list, checkpoint_data);
+			free(checkpoint_data);
 			done++;
 		}
 		checkpoint_data = next_checkpoint_data;