During ReplicationCoordinatorImpl startup, we initialize the LocalOplogInfo oplog collection by calling acquireOplogForLogging

https://github.com/mongodb/mongo/blame/b70be78ecfe773a1c9b5685b4af9cd25ea214eb5/src/mongo/db/repl/oplog.cpp#L2548

This acquires an IX lock on the oplog collection, and oplogInfo itself is not protected at all. Thus a simultaneous call which reads the oplogInfo pointer with only an IS global lock can cause a data race.

Possibly we should acquire global X here, as the catalog code does when it uses the similar establishOplogForLogging().