LDAPI autobind authentication for services#

Overview#

This design proposed to use LDAPI autobind for some internal FreeIPA services, so they can be started and used before KRB5 KDC service is operational. The proposal makes use of 389 DS’ existing LDAPI autobind implementation and new LDAPI autobind DN rewriter feature, which will be available in early 2021.

LDAPI autobind is a form of authentication that uses the effective UID and GID of a process as credentials. It uses SO_PEERCRED feature of unix socket connection and SASL EXTERNAL mechanism. Autobind enables secure and fast authentication for local services without using KRB5 KDC.

Technical background#

Benefits#

LDAPI autobind does not depend on Kerberos/GSSAPI. This fact permits services to authenticate and operate before KRB5 KDC is up and running. For example DNS services can be started before KDC, so a time sync daemon can synchronize clocks before KDC starts.

It’s faster than GSSAPI. On a test system ldapwhoami command with -Y EXTERNAL takes around 12ms real time. On the same system -Y GSSAPI over LDAPI with a cached service ticket takes between 20ms to 500ms real time.

Drawbacks#

On Fedora and RHEL-based systems, FreeIPA’s keytabs are not only protected by DAC permissions (discrete access control, also known as Unix file permissions) but also by SELinux’s mandatory access control. For example named.keytab is only readable by user named or group members of group named and processes that are allowed to open files with SELinux type etc_t. SO_PEERCRED-based authentication no longer verifies the SELinux context of a process. It is possible to retrieve the security context of a peer process with SO_PEERSEC, but 389-DS does not implement the feature.

The risk is negligible for BIND named. Most process are allowed to read files with etc_t any way.

Containers#

SO_PEERCRED based authentication works across containers if and only if the 389-DS server process and LDAP client process can share a mount point and a common user namespace. The mount point is required to share the Unix socket file of the 389-DS process. The user namespace must be shared, because the Kernel translates UID and GID numbers between namespaces. Disjunct namespaces result in UID/GID -1.

Candidates#

  • BIND named service, so named and chronyd can be started before KDC.

  • krb5kdc and kadmin service (after dropping root privileges)

  • ipa-custodia (after dropping root privileges)

Implementation#

The implementation makes use of the new DN rewriter feature for autobind. The feature enables FreeIPA to store the UID/GID mapping in the local, non-replicated cn=config backend and then map succesful binds to a DN in the replicated domain database. The approach has two benefits:

  1. FreeIPA does not depend on reserved and uniform UID/GID allocation across all servers. Each server can map its local UID/GID assignment to a principal.

  2. FreeIPA can map UID/GID combination to a host-specific service principal. For example server srv1 can map UID 25 / GID 25 to DNS/srv1.ipa.example@IPA.EXAMPLE. srv2 can map the same UID and GID to its principal DNS/srv2.ipa.example@IPA.EXAMPLE.

global settings#

LDAPI mappings, search base for UID/GID, and mapping base for nsslapd-authenticateAsDN must be configured in cn=config.

dn: cn=config
nsslapd-ldapimaptoentries: on
nsslapd-ldapientrysearchbase: cn=auto_bind,cn=config
nsslapd-ldapidnmappingbase: cn=auto_bind,cn=config

service-specific settings#

LDAPI DN rewriter feature comes with a new object class for mapping UID/GID to another DN. For example a mapping for BIND named would look like this:

dn: cn=named,cn=auto_bind,cn=config
objectClass: top
objectClass: nsLDAPIFixedAuthMap
cn: named
uidNumber: $NAMED_UID
gidNumber: $NAMED_GID
nsslapd-authenticateAsDN: krbprincipalname=DNS/$FQDN@$REALM,cn=services,cn=accounts,$SUFFIX

NOTE 389-DS does not enforce referential integrity in cn=config. nsslapd-authenticateAsDN can reference DNs that do not exist yet. A missing target results into authenticate error.

reload LDAPI mappings#

389-DS has an internal cache for LDAPI mappings. The cache must be refreshed after mapping are added, removed, or changed. A refresh can be accomplished by either restarting the server process or running the reload ldapi mappings task. The implementation detail will be handled automatically by new API and ipa-ldap-updater framework.

backup / restore#

It is possible that UID or GID can change, when a server is reinstalled and restored from a backup. ipa-restore will refresh all mappings on restore.

References#