The LFS spec recommends this, too. They suggest an ino64_t in struct stat
and struct dirent.
4 milliards of inodes sounds like a lot. However, this style of argument
collapsed in the past many times. If we change the interface, we should
get it right once and for good (this also involves dev_t).
With the decoupling of types in glibc, the kernel interface does not
necessarily reflect in the application interface.
Regards,
Martin