SMOTE within a recipe versus SMOTE in trainControl

Question

I am trying to understand where exactly SMOTE-ing should occur when training a model with cross-validation. I understand that all pre-processing steps should occur for each fold