Right now we drop the write lock between finishing a batch and updating the oplog (and notifying upstream that we've applied ops).
This allows queued readers to run, and will delay our reporting of our progress upstream. This may in turn needlessly delay write concern getLastError calls on the primary.