audit: use proper refcount locking on audit_sock
authorRichard Guy Briggs <rgb@redhat.com>
Tue, 13 Dec 2016 15:03:01 +0000 (10:03 -0500)
committerPaul Moore <paul@paul-moore.com>
Wed, 14 Dec 2016 18:06:04 +0000 (13:06 -0500)
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@google.com> and Cong Wang
<xiyou.wangcong@gmail.com> on ideas how to fix it.

Signed-off-by: Richard Guy Briggs <rgb@redhat.com>
Reviewed-by: Cong Wang <xiyou.wangcong@gmail.com>
[PM: fixed the comment block text formatting for auditd_reset()]
Signed-off-by: Paul Moore <paul@paul-moore.com>
kernel/audit.c

index f20eee0db7e6ca820ad3bcac1fac2deeeb698a1d..41017685f9f2836754d2338dfd75882d4f9be49a 100644 (file)
@@ -445,15 +445,20 @@ static void kauditd_retry_skb(struct sk_buff *skb)
  *
  * Description:
  * Break the auditd/kauditd connection and move all the records in the retry
- * queue into the hold queue in case auditd reconnects.
+ * queue into the hold queue in case auditd reconnects.  The audit_cmd_mutex
+ * must be held when calling this function.
  */
 static void auditd_reset(void)
 {
        struct sk_buff *skb;
 
        /* break the connection */
+       if (audit_sock) {
+               sock_put(audit_sock);
+               audit_sock = NULL;
+       }
        audit_pid = 0;
-       audit_sock = NULL;
+       audit_nlk_portid = 0;
 
        /* flush all of the retry queue to the hold queue */
        while ((skb = skb_dequeue(&audit_retry_queue)))
@@ -579,7 +584,9 @@ static int kauditd_thread(void *dummy)
 
                                auditd = 0;
                                if (AUDITD_BAD(rc, reschedule)) {
+                                       mutex_lock(&audit_cmd_mutex);
                                        auditd_reset();
+                                       mutex_unlock(&audit_cmd_mutex);
                                        reschedule = 0;
                                }
                        } else
@@ -594,7 +601,9 @@ static int kauditd_thread(void *dummy)
                                auditd = 0;
                                if (AUDITD_BAD(rc, reschedule)) {
                                        kauditd_hold_skb(skb);
+                                       mutex_lock(&audit_cmd_mutex);
                                        auditd_reset();
+                                       mutex_unlock(&audit_cmd_mutex);
                                        reschedule = 0;
                                } else
                                        /* temporary problem (we hope), queue
@@ -623,7 +632,9 @@ quick_loop:
                                if (rc) {
                                        auditd = 0;
                                        if (AUDITD_BAD(rc, reschedule)) {
+                                               mutex_lock(&audit_cmd_mutex);
                                                auditd_reset();
+                                               mutex_unlock(&audit_cmd_mutex);
                                                reschedule = 0;
                                        }
 
@@ -1010,11 +1021,16 @@ static int audit_receive_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
                        }
                        if (audit_enabled != AUDIT_OFF)
                                audit_log_config_change("audit_pid", new_pid, audit_pid, 1);
-                       audit_pid = new_pid;
-                       audit_nlk_portid = NETLINK_CB(skb).portid;
-                       audit_sock = skb->sk;
-                       if (!new_pid)
+                       if (new_pid) {
+                               if (audit_sock)
+                                       sock_put(audit_sock);
+                               audit_pid = new_pid;
+                               audit_nlk_portid = NETLINK_CB(skb).portid;
+                               sock_hold(skb->sk);
+                               audit_sock = skb->sk;
+                       } else {
                                auditd_reset();
+                       }
                        wake_up_interruptible(&kauditd_wait);
                }
                if (s.mask & AUDIT_STATUS_RATE_LIMIT) {
@@ -1283,8 +1299,10 @@ static void __net_exit audit_net_exit(struct net *net)
 {
        struct audit_net *aunet = net_generic(net, audit_net_id);
        struct sock *sock = aunet->nlsk;
+       mutex_lock(&audit_cmd_mutex);
        if (sock == audit_sock)
                auditd_reset();
+       mutex_unlock(&audit_cmd_mutex);
 
        RCU_INIT_POINTER(aunet->nlsk, NULL);
        synchronize_net();