I moved this a little too late - we need to run populate_rootfs() before
running initcalls because some driver initcalls need to open files for
firmware.

The populate_rootfs() call is still coming after init_idle(), so it won't
knock the scheduler over.