The LRU lock and mmap_sem

I'm surprised that range locking is seen as insufficient. I've seen severe performance issues where a page fault gets stuck behind IO triggered by munmap() or similar on an entirely different VMA.