[patch 2.6.17] Standardize i386/x86_64 handling of NMI_VECTOR

[patch 2.6.17] Standardize i386/x86_64 handling of NMI_VECTOR

[Keith Owens]

x86_64 and i386 behave inconsistently when sending an IPI on vector 2
(NMI_VECTOR).  Make both behave the same, so IPI 2 is sent as NMI.

The crash code was abusing send_IPI_allbutself() by passing a code
instead of a vector, it only worked because crash knew about the
internal code of send_IPI_allbutself().  Change crash to use NMI_VECTOR
instead, and remove the comment about how crash was abusing the function.

Signed-off-by: Keith Owens <kaos@sgi.com>

---
 arch/i386/kernel/crash.c   |    7 +------
 arch/i386/kernel/smp.c     |   12 +++++++++++-
 arch/x86_64/kernel/crash.c |    2 +-
 include/asm-i386/hw_irq.h  |    2 ++
 4 files changed, 15 insertions(+), 8 deletions(-)

Index: linux/arch/i386/kernel/crash.c
===================================================================
--- linux.orig/arch/i386/kernel/crash.c
+++ linux/arch/i386/kernel/crash.c
@@ -119,14 +119,9 @@ static int crash_nmi_callback(struct pt_
 	return 1;
 }
 
-/*
- * By using the NMI code instead of a vector we just sneak thru the
- * word generator coming out with just what we want.  AND it does
- * not matter if clustered_apic_mode is set or not.
- */
 static void smp_send_nmi_allbutself(void)
 {
-	send_IPI_allbutself(APIC_DM_NMI);
+	send_IPI_allbutself(NMI_VECTOR);
 }
 
 static void nmi_shootdown_cpus(void)
Index: linux/arch/i386/kernel/smp.c
===================================================================
--- linux.orig/arch/i386/kernel/smp.c
+++ linux/arch/i386/kernel/smp.c
@@ -114,7 +114,17 @@ DEFINE_PER_CPU(struct tlb_state, cpu_tlb
 
 static inline int __prepare_ICR (unsigned int shortcut, int vector)
 {
-	return APIC_DM_FIXED | shortcut | vector | APIC_DEST_LOGICAL;
+	unsigned int icr = shortcut | APIC_DEST_LOGICAL;
+
+	switch (vector) {
+	default:
+		icr |= APIC_DM_FIXED | vector;
+		break;
+	case NMI_VECTOR:
+		icr |= APIC_DM_NMI;
+		break;
+	}
+	return icr;
 }
 
 static inline int __prepare_ICR2 (unsigned int mask)
Index: linux/arch/x86_64/kernel/crash.c
===================================================================
--- linux.orig/arch/x86_64/kernel/crash.c
+++ linux/arch/x86_64/kernel/crash.c
@@ -118,7 +118,7 @@ static int crash_nmi_callback(struct pt_
 
 static void smp_send_nmi_allbutself(void)
 {
-	send_IPI_allbutself(APIC_DM_NMI);
+	send_IPI_allbutself(NMI_VECTOR);
 }
 
 /*
Index: linux/include/asm-i386/hw_irq.h
===================================================================
--- linux.orig/include/asm-i386/hw_irq.h
+++ linux/include/asm-i386/hw_irq.h
@@ -20,6 +20,8 @@
 
 struct hw_interrupt_type;
 
+#define NMI_VECTOR		0x02
+
 /*
  * Various low-level irq details needed by irq.c, process.c,
  * time.c, io_apic.c and smp.c

Re: [patch 2.6.17] Standardize i386/x86_64 handling of NMI_VECTOR

[Andi Kleen]

On Thursday 22 June 2006 11:01, Keith Owens wrote:
> x86_64 and i386 behave inconsistently when sending an IPI on vector 2
> (NMI_VECTOR).  Make both behave the same, so IPI 2 is sent as NMI.
> 
> The crash code was abusing send_IPI_allbutself() by passing a code
> instead of a vector, it only worked because crash knew about the
> internal code of send_IPI_allbutself().  Change crash to use NMI_VECTOR
> instead, and remove the comment about how crash was abusing the function.

Does that fix anything? 

-Andi

Re: [patch 2.6.17] Standardize i386/x86_64 handling of NMI_VECTOR

[Keith Owens]

Andi Kleen (on Thu, 22 Jun 2006 12:48:21 +0200) wrote:
>On Thursday 22 June 2006 11:01, Keith Owens wrote:
>> x86_64 and i386 behave inconsistently when sending an IPI on vector 2
>> (NMI_VECTOR).  Make both behave the same, so IPI 2 is sent as NMI.
>> 
>> The crash code was abusing send_IPI_allbutself() by passing a code
>> instead of a vector, it only worked because crash knew about the
>> internal code of send_IPI_allbutself().  Change crash to use NMI_VECTOR
>> instead, and remove the comment about how crash was abusing the function.
>
>Does that fix anything? 

This patch is a pre-requisite for fixing the problem where sending an
IPI as NMI would reboot some Dell Xeon systems.  I cannot fix that
problem while crash continus to abuse send_IPI_allbutself().

It also removes the inconsistency between i386 and x86_64 for
NMI_VECTOR.  That will simplify all the RAS code that needs to bring
all the cpus to a clean stop, even when one or more cpus are spinning
disabled.

Re: [patch 2.6.17] Standardize i386/x86_64 handling of NMI_VECTOR

[Andi Kleen]

On Thursday 22 June 2006 13:46, Keith Owens wrote:
> Andi Kleen (on Thu, 22 Jun 2006 12:48:21 +0200) wrote:
> >On Thursday 22 June 2006 11:01, Keith Owens wrote:
> >> x86_64 and i386 behave inconsistently when sending an IPI on vector 2
> >> (NMI_VECTOR).  Make both behave the same, so IPI 2 is sent as NMI.
> >> 
> >> The crash code was abusing send_IPI_allbutself() by passing a code
> >> instead of a vector, it only worked because crash knew about the
> >> internal code of send_IPI_allbutself().  Change crash to use NMI_VECTOR
> >> instead, and remove the comment about how crash was abusing the function.
> >
> >Does that fix anything? 
> 
> This patch is a pre-requisite for fixing the problem where sending an
> IPI as NMI would reboot some Dell Xeon systems.  I cannot fix that
> problem while crash continus to abuse send_IPI_allbutself().

I merged it with the updated description. Thanks.

-Andi

Re: [patch 2.6.17] Standardize i386/x86_64 handling of NMI_VECTOR

[Keith Owens]

Andi Kleen (on Thu, 22 Jun 2006 13:59:46 +0200) wrote:
>On Thursday 22 June 2006 13:46, Keith Owens wrote:
>> Andi Kleen (on Thu, 22 Jun 2006 12:48:21 +0200) wrote:
>> >On Thursday 22 June 2006 11:01, Keith Owens wrote:
>> >> x86_64 and i386 behave inconsistently when sending an IPI on vector 2
>> >> (NMI_VECTOR).  Make both behave the same, so IPI 2 is sent as NMI.
>> >> 
>> >> The crash code was abusing send_IPI_allbutself() by passing a code
>> >> instead of a vector, it only worked because crash knew about the
>> >> internal code of send_IPI_allbutself().  Change crash to use NMI_VECTOR
>> >> instead, and remove the comment about how crash was abusing the function.
>> >
>> >Does that fix anything? 
>> 
>> This patch is a pre-requisite for fixing the problem where sending an
>> IPI as NMI would reboot some Dell Xeon systems.  I cannot fix that
>> problem while crash continus to abuse send_IPI_allbutself().
>
>I merged it with the updated description. Thanks.

Merged how much and into which tree?  Most of the change is i386, with
a one line change to x86_64.  I have to ensure that the i386 bits get
into the main tree.