I don't like this solution. What speaks against building up the page array after the first buffer lookup failed, but before linking the buffer into the rbtree? That's the the inode cache and all the VFS caches do.