[PATCH][2.5.71] fix x86-64 nmi.c oops on Simics

[PATCH][2.5.71] fix x86-64 nmi.c oops on Simics

[mikpe]

Andi,

In 2.5, x86-64's nmi.c will semi-enable the local APIC NMI watchdog
on Simics, which leads to a kernel oops if disable_lapic_nmi_watchdog()
is called. The reason is:

1. setup_apic_nmi_watchdog() calls setup_k7_watchdog() for incorrect
   CPU models. Simics is model 8 while we should only allow models 6
   and 15 (well, 6 can't occur).
2. setup_k7_watchdog() initialises nmi_watchdog_msr [which tells others
   that we're enabled], and then tries to wrmsr() the MSRs. This fails
   in Simics with a #GP, you catch the failure and bail out.
3. on return to setup_apic_nmi_watchdog(), the failure goes unnoticed
   and nmi_active is set to 1.
4. Later someone calls disable_lapic_nmi_watchdog(), which, since
   nmi_active == 1, goes ahead and wrmsr()s EVNTSEL0. Bang!

The patch below forces setup_apic_nmi_watchdog() to bail out on unknown
models, so we never get into this half-initialised situation on Simics.
This also merges the code with what's in i386's nmi.c.

With this in place you could remove the error-checking wrmsr()s, but
I didn't bother with that at this time.

/Mikael

diff -ruN linux-2.5.71-andi/arch/x86_64/kernel/nmi.c linux-2.5.71-mikpe/arch/x86_64/kernel/nmi.c
--- linux-2.5.71-andi/arch/x86_64/kernel/nmi.c	2003-06-16 16:08:35.000000000 +0200
+++ linux-2.5.71-mikpe/arch/x86_64/kernel/nmi.c	2003-06-16 16:20:29.000000000 +0200
@@ -233,7 +236,7 @@
 {
 	switch (boot_cpu_data.x86_vendor) {
 	case X86_VENDOR_AMD:
-		if (boot_cpu_data.x86 < 6)
+		if (boot_cpu_data.x86 != 6 && boot_cpu_data.x86 != 15)
 			return;
 		setup_k7_watchdog();
 		break;

Re: [PATCH][2.5.71] fix x86-64 nmi.c oops on Simics

[Andi Kleen]

On Mon, Jun 16, 2003 at 05:12:02PM +0200, mikpe@csd.uu.se wrote:
> Andi,
> 
> In 2.5, x86-64's nmi.c will semi-enable the local APIC NMI watchdog
> on Simics, which leads to a kernel oops if disable_lapic_nmi_watchdog()
> is called. The reason is:

Thanks. I used to test for Simics explicitely by checking the cpuid
name for "ScrewDriver" in other places. I think I will do this here too.

-Andi