audit: try harder to send to auditd upon netlink failure

There are several reports of the kernel losing contact with auditd when
it is, in fact, still running.  When this happens, kernel syslogs show:
	"audit: *NO* daemon at audit_pid=<pid>"
although auditd is still running, and is apparently happy, listening on
the netlink socket. The pid in the "*NO* daemon" message matches the pid
of the running auditd process.  Restarting auditd solves this.

The problem appears to happen randomly, and doesn't seem to be strongly
correlated to the rate of audit events being logged.  The problem
happens fairly regularly (every few days), but not yet reproduced to
order.

On production kernels, BUG_ON() is a no-op, so any error will trigger
this.

Commit 34eab0a7cd ("audit: prevent an older auditd shutdown from
orphaning a newer auditd startup") eliminates one possible cause.  This
isn't the case here, since the PID in the error message and the PID of
the running auditd match.

The primary expected cause of error here is -ECONNREFUSED when the audit
daemon goes away, when netlink_getsockbyportid() can't find the auditd
portid entry in the netlink audit table (or there is no receive
function).  If -EPERM is returned, that situation isn't likely to be
resolved in a timely fashion without administrator intervention.  In
both cases, reset the audit_pid.  This does not rule out a race
condition.  SELinux is expected to return zero since this isn't an INET
or INET6 socket.  Other LSMs may have other return codes.  Log the error
code for better diagnosis in the future.

In the case of -ENOMEM, the situation could be temporary, based on local
or general availability of buffers.  -EAGAIN should never happen since
the netlink audit (kernel) socket is set to MAX_SCHEDULE_TIMEOUT.
-ERESTARTSYS and -EINTR are not expected since this kernel thread is not
expected to receive signals.  In these cases (or any other unexpected
ones for now), report the error and re-schedule the thread, retrying up
to 5 times.

v2:
	Removed BUG_ON().
	Moved comma in pr_*() statements.
	Removed audit_strerror() text.

Reported-by: Vipin Rathor <v.rathor@gmail.com>
Reported-by: <ctcard@hotmail.com>
Signed-off-by: Richard Guy Briggs <rgb@redhat.com>
[PM: applied rgb's fixup patch to correct audit_log_lost() format issues]
Signed-off-by: Paul Moore <pmoore@redhat.com>

kernel/audit.c

@@ -407,16 +407,33 @@ static void audit_printk_skb(struct sk_buff *skb)
 static void kauditd_send_skb(struct sk_buff *skb)
 {
 	int err;
+	int attempts = 0;
+#define AUDITD_RETRIES 5
+
+restart:
 	/* take a reference in case we can't send it and we want to hold it */
 	skb_get(skb);
 	err = netlink_unicast(audit_sock, skb, audit_nlk_portid, 0);
 	if (err < 0) {
-		BUG_ON(err != -ECONNREFUSED); /* Shouldn't happen */
+		pr_err("netlink_unicast sending to audit_pid=%d returned error: %d\n",
+		       audit_pid, err);
 		if (audit_pid) {
-			pr_err("*NO* daemon at audit_pid=%d\n", audit_pid);
-			audit_log_lost("auditd disappeared");
-			audit_pid = 0;
-			audit_sock = NULL;
+			if (err == -ECONNREFUSED || err == -EPERM
+			    || ++attempts >= AUDITD_RETRIES) {
+				char s[32];
+
+				snprintf(s, sizeof(s), "audit_pid=%d reset", audit_pid);
+				audit_log_lost(s);
+				audit_pid = 0;
+				audit_sock = NULL;
+			} else {
+				pr_warn("re-scheduling(#%d) write to audit_pid=%d\n",
+					attempts, audit_pid);
+				set_current_state(TASK_INTERRUPTIBLE);
+				schedule();
+				__set_current_state(TASK_RUNNING);
+				goto restart;
+			}
 		}
 		/* we might get lucky and get this in the next auditd */
 		audit_hold_skb(skb);