[RFC] AUDIT: do not panic when printk loses messages

On the latest kernels if one was to load about 15 rules, set the failure
state to panic, and then run service auditd stop the kernel will panic.
This is because auditd stops, then the script deletes all of the rules.
These deletions are sent as audit messages out of the printk kernel
interface which is already known to be lossy.  These will overun the
default kernel rate limiting (10 really fast messages) and will call
audit_panic().  The same effect can happen if a slew of avc's come
through while auditd is stopped.

This can be fixed a number of ways but this patch fixes the problem by
just not panicing if auditd is not running.  We know printk is lossy and
if the user chooses to set the failure mode to panic and tries to use
printk we can't make any promises no matter how hard we try, so why try?
At least in this way we continue to get lost message accounting and will
eventually know that things went bad.

The other change is to add a new call to audit_log_lost() if auditd
disappears.  We already pulled the skb off the queue and couldn't send
it so that message is lost.  At least this way we will account for the
last message and panic if the machine is configured to panic.  This code
path should only be run if auditd dies for unforeseen reasons.  If
auditd closes correctly audit_pid will get set to 0 and we won't walk
this code path.
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

kernel/audit.c

@@ -170,6 +170,8 @@ void audit_panic(const char *message)
 			printk(KERN_ERR "audit: %s\n", message);
 		break;
 	case AUDIT_FAIL_PANIC:
+		/* test audit_pid since printk is always losey, why bother? */
+		if (audit_pid)
 			panic("audit: %s\n", message);
 		break;
 	}
@@ -352,6 +354,7 @@ static int kauditd_thread(void *dummy)
 				if (err < 0) {
 					BUG_ON(err != -ECONNREFUSED); /* Shoudn't happen */
 					printk(KERN_ERR "audit: *NO* daemon at audit_pid=%d\n", audit_pid);
+					audit_log_lost("auditd dissapeared\n");
 					audit_pid = 0;
 				}
 			} else {