Kernel RFC proposes native UUID and GUID user IDs

An RFC posted to the Linux kernel list proposes treating user and group IDs as full UUIDs or Microsoft GUIDs, not only classic 32-bit values, so machines joined to Active Directory, LDAP, and similar services no longer need local mapping tables that diverge, collide, or rewrite ownership on shared storage.

Jan Koester's second-round series argues that directory services already name people by 128-bit objectGUIDs and entryUUIDs. Today winbind, sssd, nfsidmap, and related helpers compress those into 32-bit ranges or hashes. The numbers differ across hosts unless every mapping is identical, collisions remain possible, and a backup or NFS export records whoever wrote the file on that machine rather than the identity the directory issued. Putting the 128-bit id in the kernel would make the directory's id the owner on disk and in every credential.

Internally, kernel uid and gid types become a tagged union holding a legacy number, a UUID, or a GUID. There is no reverse lookup table: the kernel compares, hashes, and stores the whole value. New system calls with an _r suffix accept a 20-byte structure that carries the type and payload. Existing 32-bit calls keep working and report the overflow id for any UUID or GUID, the same way older 16-bit calls once treated large uids. The rich ids exist only in the initial user namespace; child namespaces see the overflow id and gain no privilege over them, while idmapped mounts pass them through unchanged.

The series extends the surfaces that already carry ownership: socket peer credentials and SCM messages, siginfo and signalfd, System V IPC permission records, keyrings, procfs status, audit, POSIX ACL xattrs (a new version format), ext4 inodes via a read-only-compatible feature, overlayfs upper layers, devpts, ksmbd login and SIDs, and NFSv4 owner strings. Filesystems without the feature refuse such owners or mount read-only; classic AUTH_SYS and some fixed-size SMB POSIX SID buffers still cannot carry them and continue to show the overflow id.

Documentation and selftests accompany the RFC. The work is explicitly a request for comments, not a merged change, and would need broad review before any stable kernel could ship native 128-bit identities.