audit: drop audit_cmd_lock in AUDIT_USER family of cases
authorRichard Guy Briggs <rgb@redhat.com>
Mon, 2 Dec 2013 16:33:01 +0000 (11:33 -0500)
committerEric Paris <eparis@redhat.com>
Tue, 14 Jan 2014 03:32:11 +0000 (22:32 -0500)
We do not need to hold the audit_cmd_mutex for this family of cases.  The
possible exception to this is the call to audit_filter_user(), so drop the lock
immediately after.  To help in fixing the race we are trying to avoid, make
sure that nothing called by audit_filter_user() calls audit_log_start().  In
particular, watch out for *_audit_rule_match().

This fix will take care of systemd and anything USING audit.  It still means
that we could race with something configuring audit and auditd shutting down.

Signed-off-by: Richard Guy Briggs <rgb@redhat.com>
Reported-by: toshi.okajima@jp.fujitsu.com
Tested-by: toshi.okajima@jp.fujitsu.com
Signed-off-by: Eric Paris <eparis@redhat.com>
kernel/audit.c

index ce8514f..c05b602 100644 (file)
@@ -876,6 +876,7 @@ static int audit_receive_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
                                if (err)
                                        break;
                        }
+                       mutex_unlock(&audit_cmd_mutex);
                        audit_log_common_recv_msg(&ab, msg_type);
                        if (msg_type != AUDIT_USER_TTY)
                                audit_log_format(ab, " msg='%.*s'",
@@ -893,6 +894,7 @@ static int audit_receive_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
                        }
                        audit_set_portid(ab, NETLINK_CB(skb).portid);
                        audit_log_end(ab);
+                       mutex_lock(&audit_cmd_mutex);
                }
                break;
        case AUDIT_ADD_RULE: