Added a coarse-grain lock to prevent concurrent mmap()/munmap()/mprotect().
We can implement something more fine-grained if this becomes a performance
bottleneck, but I doubt it ever would.