On Mon, Dec 12, 2016 at 5:03 AM, Richard Guy Briggs <rgb(a)redhat.com> wrote:
 Resetting audit_sock appears to be racy.
 audit_sock was being copied and dereferenced without using a refcount on
 the source sock.
 Bump the refcount on the underlying sock when we store a refrence in
 audit_sock and release it when we reset audit_sock.  audit_sock
 modification needs the audit_cmd_mutex.
 See: 
https://lkml.org/lkml/2016/11/26/232
 Thanks to Eric Dumazet <edumazet(a)google.com> and Cong Wang
 <xiyou.wangcong(a)gmail.com> on ideas how to fix it.
 Signed-off-by: Richard Guy Briggs <rgb(a)redhat.com>
 ---
 There has been a lot of change in the audit code that is about to go
 upstream to address audit queue issues.  This patch is based on the
 source tree: 
git://git.infradead.org/users/pcmoore/audit#next
 ---
  kernel/audit.c |   34 ++++++++++++++++++++++++++++------
  1 files changed, 28 insertions(+), 6 deletions(-) 
My previous question about testing still stands, but I took a closer
look and have some additional comments, see below ...
 diff --git a/kernel/audit.c b/kernel/audit.c
 index f20eee0..439f7f3 100644
 --- a/kernel/audit.c
 +++ b/kernel/audit.c
 @@ -452,7 +452,9 @@ static void auditd_reset(void)
         struct sk_buff *skb;
         /* break the connection */
 +       sock_put(audit_sock);
         audit_pid = 0;
 +       audit_nlk_portid = 0;
         audit_sock = NULL;
         /* flush all of the retry queue to the hold queue */
 @@ -478,6 +480,12 @@ static int kauditd_send_unicast_skb(struct sk_buff *skb)
         if (rc >= 0) {
                 consume_skb(skb);
                 rc = 0;
 +       } else {
 +               if (rc & (-ENOMEM|-EPERM|-ECONNREFUSED)) { 
I dislike the way you wrote this because instead of simply looking at
this to see if it correct I need to sort out all the bits and find out
if there are other error codes that could run afoul of this check ...
make it simple, e.g. (rc == -ENOMEM || rc == -EPERM || ...).
Actually, since EPERM is 1, -EPERM (-1 in two's compliment is
0xffffffff) is going to cause this to be true for pretty much any
value of rc, yes?
 +                       mutex_lock(&audit_cmd_mutex);
 +                       auditd_reset();
 +                       mutex_unlock(&audit_cmd_mutex);
 +               } 
The code in audit#next handles netlink_unicast() errors in
kauditd_thread() and you are adding error handling code here in
kauditd_send_unicast_skb() ... that's messy.  I don't care too much
where the auditd_reset() call is made, but let's only do it in one
function; FWIW, I originally put the error handling code in
kauditd_thread() because there was other error handling code that
needed to done in that scope so it resulted in cleaner code.
Related, I see you are now considering ENOMEM to be a fatal condition,
that differs from the AUDITD_BAD macro in kauditd_thread(); this
difference needs to be reconciled.
Finally, you should update the comment header block for auditd_reset()
that it needs to be called with the audit_cmd_mutex held.
 @@ -1004,17 +1018,22 @@ static int audit_receive_msg(struct sk_buff
*skb, struct nlmsghdr *nlh)
                                 return -EACCES;
                         }
                         if (audit_pid && new_pid &&
 -                           audit_replace(requesting_pid) != -ECONNREFUSED) {
 +                           (audit_replace(requesting_pid) &
(-ECONNREFUSED|-EPERM|-ENOMEM))) { 
Do we simply want to treat any error here as fatal, and not just
ECONN/EPERM/ENOMEM?  If not, let's come up with a single macro to
handle the fatal netlink_unicast() return codes so we have some chance
to keep things consistent in the future.
-- 
paul moore
www.paul-moore.com