Zachary Amsden wrote:

Do you agree it is better to be safe than sorry in this case?  The
kind of bugs introduced by getting this wrong are really hard to find,
and I would rather err on the side of an extra increment and decrement
of preempt_count that causing a regression.

I think this patch is the direction we should go.  I this this would
work equally well for the other pv implementations; it would probably go
into the common lazy mode logic when we get around to doing it.

    J

diff -r b3fcc228c531 arch/i386/xen/enlighten.c

--- a/arch/i386/xen/enlighten.c	Mon Aug 20 14:20:15 2007 -0700
+++ b/arch/i386/xen/enlighten.c	Mon Aug 27 13:40:24 2007 -0700

@@ -250,6 +250,9 @@ static void xen_halt(void)
 
 static void xen_set_lazy_mode(enum paravirt_lazy_mode mode)
 {
+	if (preemptible() && mode == PARAVIRT_LAZY_FLUSH)
+		return;		/* nothing to flush with preempt on */
+
 	BUG_ON(preemptible());
 
 	switch (mode) {