keys: fix race with concurrent install_user_keyrings()
authorDavid Howells <dhowells@redhat.com>
Tue, 12 Mar 2013 05:44:31 +0000 (16:44 +1100)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Thu, 14 Mar 2013 18:26:36 +0000 (11:26 -0700)
commit71d039e231396724ffedc4cfe8a3476f810f4ff2
treef895dd2a02ed50f20857e5a5a1b9f241b695bd72
parent53be86cb15c4a7b1a0f5bbfa0c566c168891ae52
keys: fix race with concurrent install_user_keyrings()

commit 0da9dfdd2cd9889201bc6f6f43580c99165cd087 upstream.

This fixes CVE-2013-1792.

There is a race in install_user_keyrings() that can cause a NULL pointer
dereference when called concurrently for the same user if the uid and
uid-session keyrings are not yet created.  It might be possible for an
unprivileged user to trigger this by calling keyctl() from userspace in
parallel immediately after logging in.

Assume that we have two threads both executing lookup_user_key(), both
looking for KEY_SPEC_USER_SESSION_KEYRING.

THREAD A THREAD B
=============================== ===============================
==>call install_user_keyrings();
if (!cred->user->session_keyring)
==>call install_user_keyrings()
...
user->uid_keyring = uid_keyring;
if (user->uid_keyring)
return 0;
<==
key = cred->user->session_keyring [== NULL]
user->session_keyring = session_keyring;
atomic_inc(&key->usage); [oops]

At the point thread A dereferences cred->user->session_keyring, thread B
hasn't updated user->session_keyring yet, but thread A assumes it is
populated because install_user_keyrings() returned ok.

The race window is really small but can be exploited if, for example,
thread B is interrupted or preempted after initializing uid_keyring, but
before doing setting session_keyring.

This couldn't be reproduced on a stock kernel.  However, after placing
systemtap probe on 'user->session_keyring = session_keyring;' that
introduced some delay, the kernel could be crashed reliably.

Fix this by checking both pointers before deciding whether to return.
Alternatively, the test could be done away with entirely as it is checked
inside the mutex - but since the mutex is global, that may not be the best
way.

Signed-off-by: David Howells <dhowells@redhat.com>
Reported-by: Mateusz Guzik <mguzik@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: James Morris <james.l.morris@oracle.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
security/keys/process_keys.c