Re: lguest rebroken in 2.6.22-rc3-mm1

[Rusty Russell <rusty@rustcorp.com.au>]

On Mon, 2007-06-04 at 20:12 +0200, Andi Kleen wrote:
> > >
> > > Looks like this one got lost in rc3-mm1.
> >
> > Andi said that he fixed the zero-divide by other means?
> 
> I determined it cannot happen in my source tree. When notsc
> is passed TSC CPUID is cleared and sched-clock works.
> 
> I suspect what happens is that lguest forgets to clear the TSC cpuid
> bit when it disables TSC. Then the TSC frequency doesn't get computed
> and sched-clock can divide by zero.That's purely a lguest bug that needs
> to be fixed in lguest with a 
> clear_bit(X86_FEATURE_TSC, &boot_cpu_data.x86_capability) 
> somewhere

It's not quite that simple; lguest's paravirt_ops->cpuid sets TSC off,
and indeed X86_FEATURE_TSC isn't set in boot_cpu_data.x86_capability.

But TSC is a "required feature", so "cpu_has_tsc" is always true.

How about this patch:
===
Don't try to disable the TSC: it's a required feature under modern
configurations, so just mark the sched clock unstable which has the
same effect.

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
---
 drivers/lguest/lguest.c |    4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

===================================================================
--- a/drivers/lguest/lguest.c
+++ b/drivers/lguest/lguest.c
@@ -37,6 +37,7 @@
 #include <asm/e820.h>
 #include <asm/mce.h>
 #include <asm/io.h>
+#include <asm/sched-clock.h>
 
 /* Declarations for definitions in lguest_guest.S */
 extern char lguest_noirq_start[], lguest_noirq_end[];
@@ -508,7 +509,8 @@ __init void lguest_init(void *boot)
 	/* Math is always hard! */
 	new_cpu_data.hard_math = 1;
 
-	tsc_disable = 1;
+	/* Sched clock is unusable: you'll just hurt yourself if you try. */
+	__get_cpu_var(sc_data).unstable++;
 
 #ifdef CONFIG_X86_MCE
 	mce_disabled = 1;