Idiocy in audit_serial()

[David Woodhouse]

On Fri, 2005-07-15 at 09:03 +0100, David Woodhouse wrote:
> Someone reported livelock on a 16-way itanic box, in audit_serial().
> Looking at it, that doesn't really surprise me much. An audit.78
> kernel is building with this...

There's still too much contention on it, even if we eliminate the
livelock. There's no real reason why we should be generating a serial
number even when the context isn't auditable, is there? I'm building
audit.79 with this patch. It _might_ complete before I leave for the
airport...

--- linux-2.6.9/kernel/audit.c~	2005-07-15 08:45:38.000000000 +0100
+++ linux-2.6.9/kernel/audit.c	2005-07-16 10:33:21.000000000 +0100
@@ -693,7 +693,9 @@ unsigned int audit_serial(void)
 	unsigned int ret;
 
 	spin_lock_irqsave(&serial_lock, flags);
-	ret = serial++;
+	do {
+		ret = ++serial;
+	} while (unlikely(!ret));
 	spin_unlock_irqrestore(&serial_lock, flags);
 
 	return ret;
--- linux-2.6.9/kernel/auditsc.c~	2005-07-13 22:38:50.000000000 +0100
+++ linux-2.6.9/kernel/auditsc.c	2005-07-16 10:34:06.000000000 +0100
@@ -1030,7 +1030,7 @@ void audit_syscall_entry(struct task_str
 	if (likely(state == AUDIT_DISABLED))
 		return;
 
-	context->serial     = audit_serial();
+	context->serial     = 0;
 	context->ctime      = CURRENT_TIME;
 	context->in_syscall = 1;
 	context->auditable  = !!(state == AUDIT_RECORD_CONTEXT);
@@ -1183,6 +1183,8 @@ void audit_inode(const char *name, const
 void auditsc_get_stamp(struct audit_context *ctx,
 		       struct timespec *t, unsigned int *serial)
 {
+	if (!ctx->serial)
+		ctx->serial = audit_serial();
 	t->tv_sec  = ctx->ctime.tv_sec;
 	t->tv_nsec = ctx->ctime.tv_nsec;
 	*serial    = ctx->serial;



-- 
dwmw2