Unfortunately, it looks like this fix causes a regression. Some programs rely on being able to seek() and write to a file on multiple threads. For example, py.test captures the standard streams by redirecting them to a tmpfile and then truncating+seeking to 0 after every test. This change broke that situation when multiple threads are logging. Anyway, there's no real fundamental reason to prevent concurrent access, since the underlying stdio implementation is threadsafe. I think we'll have to resurrect my PR.