Accessing legacy data as virtual RDF stores is a key issue in the building of the Web of Data. In recent years, the MongoDB database has become a popular actor in the NoSQL market, making it a significant potential contributor to the Web of Linked Data. Therefore, in this paper we address the question of how to access arbitrary MongoDB documents with SPARQL. We propose a two-step method to (i) translate a SPARQL query into a pivot abstract query under MongoDB-to-RDF mappings represented in the xR2RML language, then (ii) translate the pivot query into a concrete MongoDB query. We elaborate on the discrepancy between the expressiveness of SPARQL and the MongoDB query language, and we show that we can always come up with a rewriting that shall produce all correct answers.