Linux NFSv4 security issue: client presents wrong user's credentials to NFS server

Post by Andrew J. Romero
Hi
It is common in certain cases for
a Linux workstation administrator to
create a shared account with a
corresponding .k5login file.
Kerberos principal strings for users
are added to the .k5login and the corresponding
users can logon as the local shared account using
their Kerberos credentials.
Once logged on to the workstation, the user's
local identity (local to the workstation) is
the UID of the local shared account; however,
when accessing kerberized resources external
to the workstation (for example NFSv4 volumes),
the user's identity is expected to be the
individual user's kerberos identity.
(who are listed sharusr's .k5login) all logon to
workstation1 as sharusr and are all running sessions
simultaneously, then, the GSS Context / kerberos credentials
stored in and used by the NFS-client kernel code when
processing NFS requests on behalf of a user,
will be correct for 1 user and incorrect for N-1 users.
Each of the N-1 users will not be able to
access NFS server resources that only they have been
granted access to; but, they will be able
to access all NFS resources that the 1 user
has been granted access to. (even if their
kerberos identity has not been granted access).
Even after each of the N-1 users destroys their credentials
they still retain access to all NFS resources
that the 1 user has been granted access to.
Access will not be denied until the 1 user also destroys
his / her credentials.
It is my understanding that the RPC GSS kernel module
on an NFS client system, keeps a GSS security context
(copy of a user's credentials) that it uses when it
communicates with the NFS server on the user's behalf.
I believe that the kernel security context
relates back to the user mode process solely by UID number.

That is correct.

Post by Andrew J. Romero
Would adding session ID to the relation resolve the issue ?
Each Session for each user would then have a unique tuple
(UID,SessionID) that would map to the *correct* GSS Security
context in the kernel.

I think what could be done is to hang gss contexts off either session
or process keyrings - then each process group would be able to own an
identity independently of other processes running with the same UID.

To do this the gssd upcall could be moved to a request-key mechanism, as
there would need to be a way to pass along which krb5 credentials ought
to be used to establish the context. I suppose that if the process
already has krb5 credentials in a keyring ccache, those could be used -
but I don't think that we could count on always having a keyring ccache
available, and the fallback would be to try to guess which ccache to use
based on UID as things are today. There would additionally be the
situation where keyring ccaches cannot be used, and the main problem
becomes how to have a process tell gssd which credential cache to use.

Maybe instead of establishing contexts at file access time, they could
be established deliberately beforehand with a userspace command that
would specify which krb5 credential to use to create the context, and
then the context would hang off the process keyrings. The client could
then check to see if there's an existing keyring context, and if so use
that, and if not do the upcall.

Ben
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo-***@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html

Trond Myklebust

2014-10-07 13:36:12 UTC

That is correct.

The problem with the request key interface is that it is completely
broken when applied to containers, since it only runs in the global
init namespace context. Fixing that is a non-trivial exercise; you'd
have to not only carry a full namespace context with the RPC
credential, but also somehow apply it to the upcall thread.

The rpc.gssd model is much easier to manage, since the daemon itself
is started from the container's local init process.

Post by Benjamin Coddington
Maybe instead of establishing contexts at file access time, they could
be established deliberately beforehand with a userspace command that
would specify which krb5 credential to use to create the context, and
then the context would hang off the process keyrings. The client could
then check to see if there's an existing keyring context, and if so use
that, and if not do the upcall.

The server is free to drop any GSS session at any time, depending on
internal resource management requirements. A client must be able to
recover from that scenario or risk being unable to write back data.

Cheers
Trond
--
Trond Myklebust

Linux NFS client maintainer, PrimaryData

trond.myklebust-7I+n7zu2hftEKMMhf/***@public.gmane.org
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo-***@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html

J. Bruce Fields

2014-10-07 14:29:34 UTC

Post by Trond Myklebust
The problem with the request key interface is that it is completely
broken when applied to containers, since it only runs in the global
init namespace context. Fixing that is a non-trivial exercise; you'd
have to not only carry a full namespace context with the RPC
credential, but also somehow apply it to the upcall thread.

This one's really keeping me up at night. Mainly because of the server
reboot recovery stuff. The client v4 idmapping has this problem too,
doesn't it?

How are we going to decide if the user-helper containerization is doable
or if it's just a hopeless case?

Jeff seems optimistic about fixing it, and that'd be great, but if it's
not going to happen then we need to give up and use daemons for this
stuff.

--b.
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo-***@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html

Benjamin Coddington

2014-10-08 19:39:40 UTC

That is correct.

The problem with the request key interface is that it is completely
broken when applied to containers, since it only runs in the global
init namespace context. Fixing that is a non-trivial exercise; you'd
have to not only carry a full namespace context with the RPC
credential, but also somehow apply it to the upcall thread.
The rpc.gssd model is much easier to manage, since the daemon itself
is started from the container's local init process.

That seems like something that might be fixed in the keys stuff instead
of dragging the namespace along on the rpc cred, because it seems like a
more general problem that request-key upcalls can execute outside their
container.. but maybe that's exactly what request-key is supposed to
do. I'll ask David what he thinks about it.

Alternatively, the contexts could still hang off session or process
keyrings but still use the existing upcall...

How did the idmap upcall get moved over to the request-key interface
without taking the container problem into account?

Ah, right. Thank you for pointing that out.

Ben
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo-***@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html

Trond Myklebust

2014-10-08 19:47:51 UTC

That is correct.

The problem with the request key interface is that it is completely
broken when applied to containers, since it only runs in the global
init namespace context. Fixing that is a non-trivial exercise; you'd
have to not only carry a full namespace context with the RPC
credential, but also somehow apply it to the upcall thread.
The rpc.gssd model is much easier to manage, since the daemon itself
is started from the container's local init process.

That seems like something that might be fixed in the keys stuff instead
of dragging the namespace along on the rpc cred, because it seems like a
more general problem that request-key upcalls can execute outside their
container.. but maybe that's exactly what request-key is supposed to
do. I'll ask David what he thinks about it.
Alternatively, the contexts could still hang off session or process
keyrings but still use the existing upcall...
How did the idmap upcall get moved over to the request-key interface
without taking the container problem into account?

Timing is everything.

We converted idmapper in 2010 just before the Parallels people came
and wanted to do net namespaces in 2011. However you will note that we
still do support the rpc.idmapd daemon, and that the interfaces it
uses have been converted to support net namespace.