Following a request by David Chow on linux fsdevel, this patch causes
NFS read and write requests to take the inode from page->mapping->host
rather than relying on file->f_dentry->d_inode. Apparently this will
simplify some work he is doing on another filesystem.

In any case, it cleans up the current mix of sometimes doing one
thing, sometimes the other (historical cruft), and puts NFS client
behaviour on par with what is done in other filesystems...