changeset 8831 gets sockets, cores, and siblings wrong

changeset 8831 gets sockets, cores, and siblings wrong

[Andrew Theurer]

FWIW, I noticed that changeset 8831 gets the sockets/cores/siblings 
incorrect on my system.  On 8830 I get 4 sockets per node, 2 cores per 
socket, and 2 threads per core.  On 8831 I get 1 socket per node, 2 
cores per socket, and 8 threads per core.  Xen boot messages still 
appear to show four distinct physical cpus (sockets).  Just wondering if 
anyone has noticed this behavior.

-Andrew

Re: changeset 8831 gets sockets, cores, and siblings wrong

[Keir Fraser]

On 15 Feb 2006, at 21:21, Andrew Theurer wrote:

> FWIW, I noticed that changeset 8831 gets the sockets/cores/siblings 
> incorrect on my system.  On 8830 I get 4 sockets per node, 2 cores per 
> socket, and 2 threads per core.  On 8831 I get 1 socket per node, 2 
> cores per socket, and 8 threads per core.  Xen boot messages still 
> appear to show four distinct physical cpus (sockets).  Just wondering 
> if anyone has noticed this behavior.

This all stems from an incorrect value for smp_num_siblings. This could 
come from one of two places:
  1. Do you get a message 'WARNING: 8 siblings found for CPUx, should be 
2' during boot? If so, smpboot.c is screwing up the smp_num_siblings 
calculation.
  2. Otherwise, detect_ht() in arch/x86/cpu/common.c must be the 
culprit, and you'll need to add some tracing to it to find out where 
the value '8' is coming from.

  -- Keir

Re: [PATCH] changeset 8831 gets sockets, cores, and siblings wrong

[Andrew Theurer]

[-- Attachment #1: Type: text/plain, Size: 1242 bytes --]

Keir Fraser wrote:
>
> On 15 Feb 2006, at 21:21, Andrew Theurer wrote:
>
>> FWIW, I noticed that changeset 8831 gets the sockets/cores/siblings 
>> incorrect on my system.  On 8830 I get 4 sockets per node, 2 cores 
>> per socket, and 2 threads per core.  On 8831 I get 1 socket per node, 
>> 2 cores per socket, and 8 threads per core.  Xen boot messages still 
>> appear to show four distinct physical cpus (sockets).  Just wondering 
>> if anyone has noticed this behavior.
>
> This all stems from an incorrect value for smp_num_siblings. This 
> could come from one of two places:
>  1. Do you get a message 'WARNING: 8 siblings found for CPUx, should 
> be 2' during boot? If so, smpboot.c is screwing up the 
> smp_num_siblings calculation.
>  2. Otherwise, detect_ht() in arch/x86/cpu/common.c must be the 
> culprit, and you'll need to add some tracing to it to find out where 
> the value '8' is coming from.
It looks like with changeset 8831 cpu_core_ids are no longer unique 
across phys cpus.  So, when we search for siblings to populate 
cpu_sibling_map, just checking for same core IDs are not good enough.  
Both core ID and phys cpu ID have to match.  Patch attached fixes this.

-Andrew

<signed-off-by: habanero@us.ibm.com>

[-- Attachment #2: fix_sibling_calc-8849.patch --]
[-- Type: text/x-patch, Size: 685 bytes --]

diff -Naurp xen-unstable.hg-8849/xen/arch/x86/smpboot.c xen-unstable.hg-8849-sibling_fix/xen/arch/x86/smpboot.c
--- xen-unstable.hg-8849/xen/arch/x86/smpboot.c	2006-02-16 12:15:53.000000000 -0600
+++ xen-unstable.hg-8849-sibling_fix/xen/arch/x86/smpboot.c	2006-02-16 12:11:33.000000000 -0600
@@ -1093,7 +1093,8 @@ static void __init smp_boot_cpus(unsigne
 			for (i = 0; i < NR_CPUS; i++) {
 				if (!cpu_isset(i, cpu_callout_map))
 					continue;
-				if (cpu_core_id[cpu] == cpu_core_id[i]) {
+				if (cpu_core_id[cpu] == cpu_core_id[i] && 
+                                    phys_proc_id[cpu] == phys_proc_id[i]) {
 					siblings++;
 					cpu_set(i, cpu_sibling_map[cpu]);
 				}

[-- Attachment #3: Type: text/plain, Size: 138 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Re: [PATCH] changeset 8831 gets sockets, cores, and siblings wrong

[Keir Fraser]

On 16 Feb 2006, at 18:18, Andrew Theurer wrote:

> It looks like with changeset 8831 cpu_core_ids are no longer unique 
> across phys cpus.  So, when we search for siblings to populate 
> cpu_sibling_map, just checking for same core IDs are not good enough.  
> Both core ID and phys cpu ID have to match.  Patch attached fixes 
> this.

Thanks. It turns out all that code in smpboot.c disappeared in the 
latest 2.6 codebase, so I upgraded to that and the problem has gone 
away. :-)

  -- Keir