[PATCH] cpu hotplug ppc64 bug

[PATCH] cpu hotplug ppc64 bug

[Joel Schopp]

[-- Attachment #1: Type: text/plain, Size: 532 bytes --]

On Power4 and earlier hardware there is no need to clear the CPPR (see 
RPAp 479 section 18.5.4.7.2 for what little info there is on the CPPR) 
when stopping a cpu. On hardware that uses Power5 an undocumented change 
has been made that requires the CPPR to be cleared if an isolate is to 
be done on the stopped cpu. So the following patch lets cpu hotplug work 
on the recent hardware.

I sent this patch to the ppc64-dev list back in mid April and Suse 
picked it up then for SLES9 so it has been well tested for several months.

[-- Attachment #2: jul26.patch --]
[-- Type: text/plain, Size: 1984 bytes --]

# This is a BitKeeper generated patch for the following project:
# Project Name: Linux kernel tree
# This patch format is intended for GNU patch command version 2.5 or higher.
# This patch includes the following deltas:
#	           ChangeSet	1.1883  -> 1.1884 
#	include/asm-ppc64/xics.h	1.5     -> 1.6    
#	arch/ppc64/kernel/xics.c	1.45    -> 1.46   
#	arch/ppc64/kernel/smp.c	1.72    -> 1.73   
#
# The following is the BitKeeper ChangeSet Log
# --------------------------------------------
# 04/07/26	jschopp@zippy.ltc.austin.ibm.com	1.1884
# Power5 has a requirement to clear the CPPR before doing a stop_self or isolates will fail later on.  Clearing the CPPR is OK on Power4 so just 
# added it to the path.
# --------------------------------------------
#
diff -Nru a/arch/ppc64/kernel/smp.c b/arch/ppc64/kernel/smp.c
--- a/arch/ppc64/kernel/smp.c	Mon Jul 26 18:26:45 2004
+++ b/arch/ppc64/kernel/smp.c	Mon Jul 26 18:26:45 2004
@@ -300,6 +300,10 @@
 void cpu_die(void)
 {
 	local_irq_disable();
+	/* Some hardware requires clearing the CPPR, while other hardware does not
+	 * it is safe either way
+	 */
+	pSeriesLP_cppr_info(0, 0);
 	rtas_stop_self();
 	/* Should never get here... */
 	BUG();
diff -Nru a/arch/ppc64/kernel/xics.c b/arch/ppc64/kernel/xics.c
--- a/arch/ppc64/kernel/xics.c	Mon Jul 26 18:26:45 2004
+++ b/arch/ppc64/kernel/xics.c	Mon Jul 26 18:26:45 2004
@@ -190,7 +190,7 @@
 		      val64); 
 }
 
-static void pSeriesLP_cppr_info(int n_cpu, u8 value)
+void pSeriesLP_cppr_info(int n_cpu, u8 value)
 {
 	unsigned long lpar_rc;
 
diff -Nru a/include/asm-ppc64/xics.h b/include/asm-ppc64/xics.h
--- a/include/asm-ppc64/xics.h	Mon Jul 26 18:26:45 2004
+++ b/include/asm-ppc64/xics.h	Mon Jul 26 18:26:45 2004
@@ -19,6 +19,9 @@
 void xics_setup_cpu(void);
 void xics_cause_IPI(int cpu);
 
+/* first argument is ignored for now*/
+void pSeriesLP_cppr_info(int n_cpu, u8 value);
+
 struct xics_ipi_struct {
 	volatile unsigned long value;
 } ____cacheline_aligned;

Re: [PATCH] cpu hotplug ppc64 bug

[Paul Mackerras]

Joel Schopp writes:

> On Power4 and earlier hardware there is no need to clear the CPPR (see 
> RPAp 479 section 18.5.4.7.2 for what little info there is on the CPPR) 
> when stopping a cpu. On hardware that uses Power5 an undocumented change 
> has been made that requires the CPPR to be cleared if an isolate is to 
> be done on the stopped cpu. So the following patch lets cpu hotplug work 
> on the recent hardware.

>  void cpu_die(void)
>  {
>  	local_irq_disable();
> +	/* Some hardware requires clearing the CPPR, while other hardware does not
> +	 * it is safe either way
> +	 */
> +	pSeriesLP_cppr_info(0, 0);
>  	rtas_stop_self();

I wanted to do this a bit differently - I was going to make cpu_die be
a platform-specific function called via a ppc_md function pointer,
rather than putting very pseries-specific stuff in smp.c, which is
used on all platforms.  But having been on vacation and then
travelling, I haven't got to it yet.

Paul.

Re: [PATCH] cpu hotplug ppc64 bug

[Joel Schopp]

I think that function pointers are a good idea too.  But I'm not sure 
why that should hold up this patch.  As it stands now all ppc64 
platforms that do hotplug happen to be pSeries.  I don't know of any 
other ppc64 platforms that are not pSeries who plan on doing this in the 
next few years.  The function is already pSeries specific anyway with 
the rtas_stop_self call.

Meanwhile if we do not put this patch in all Power5 machines will be 
unable to do cpu hotplug with a mainline kernel.



> I wanted to do this a bit differently - I was going to make cpu_die be
> a platform-specific function called via a ppc_md function pointer,
> rather than putting very pseries-specific stuff in smp.c, which is
> used on all platforms.  But having been on vacation and then
> travelling, I haven't got to it yet.
> 
> Paul.
>