[Patch] remove unneeded test for cpu in cpupool

[Patch] remove unneeded test for cpu in cpupool

[Juergen Gross]

[-- Attachment #1: Type: text/plain, Size: 671 bytes --]

Hi,

attached patch removes an unneeded check in vcpu_migrate().

Andre, could you please give it a try on your 48 core machine with

xl cpupool-numa-split

I'd like to make sure the check is really not needed. On my machines it worked
okay, but your machine seems to trigger races more easily.


Juergen

-- 
Juergen Gross                 Principal Developer Operating Systems
TSP ES&S SWE OS6                       Telephone: +49 (0) 89 3222 2967
Fujitsu Technology Solutions              e-mail: juergen.gross@ts.fujitsu.com
Domagkstr. 28                           Internet: ts.fujitsu.com
D-80807 Muenchen                 Company details: ts.fujitsu.com/imprint.html

[-- Attachment #2: cputest.patch --]
[-- Type: text/x-patch, Size: 1240 bytes --]

This patch removes an unneeded check in vcpu_migrate() for the picked cpu to
be in the current cpupool. pick_cpu should only return cpus in the correct
cpupool.

Signed-off-by: juergen.gross@ts.fujitsu.com

diff -r 3caed2112c65 xen/common/schedule.c
--- a/xen/common/schedule.c	Tue Mar 15 10:14:27 2011 +0000
+++ b/xen/common/schedule.c	Wed Mar 16 07:26:38 2011 +0100
@@ -431,7 +431,7 @@ static void vcpu_migrate(struct vcpu *v)
         if ( old_lock == per_cpu(schedule_data, old_cpu).schedule_lock )
         {
             /*
-             * If we selected a CPU on the previosu iteration, check if it
+             * If we selected a CPU on the previous iteration, check if it
              * remains suitable for running this vCPU.
              */
             if ( pick_called &&
@@ -442,8 +442,7 @@ static void vcpu_migrate(struct vcpu *v)
 
             /* Select a new CPU. */
             new_cpu = SCHED_OP(VCPU2OP(v), pick_cpu, v);
-            if ( (new_lock == per_cpu(schedule_data, new_cpu).schedule_lock) &&
-                 cpu_isset(new_cpu, v->domain->cpupool->cpu_valid) )
+            if ( new_lock == per_cpu(schedule_data, new_cpu).schedule_lock )
                 break;
             pick_called = 1;
         }

[-- 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] remove unneeded test for cpu in cpupool

[Keir Fraser]

On 16/03/2011 06:34, "Juergen Gross" <juergen.gross@ts.fujitsu.com> wrote:

> Hi,
> 
> attached patch removes an unneeded check in vcpu_migrate().
> 
> Andre, could you please give it a try on your 48 core machine with
> 
> xl cpupool-numa-split
> 
> I'd like to make sure the check is really not needed. On my machines it worked
> okay, but your machine seems to trigger races more easily.

Could you BUG_ON failing this test when we exit the retry loop? It seems a
suitable loop post-condition.

 -- Keir

> 
> Juergen

Re: [Patch] remove unneeded test for cpu in cpupool

[Juergen Gross]

On 03/16/11 11:13, Keir Fraser wrote:
> On 16/03/2011 06:34, "Juergen Gross"<juergen.gross@ts.fujitsu.com>  wrote:
>
>> Hi,
>>
>> attached patch removes an unneeded check in vcpu_migrate().
>>
>> Andre, could you please give it a try on your 48 core machine with
>>
>> xl cpupool-numa-split
>>
>> I'd like to make sure the check is really not needed. On my machines it worked
>> okay, but your machine seems to trigger races more easily.
>
> Could you BUG_ON failing this test when we exit the retry loop? It seems a
> suitable loop post-condition.

Not easily. The cpu_valid mask of a cpupool might change between the call of
pick_cpu and the test. In the normal case this is okay, as removing a cpu
from a cpupool (which is the critical case) will call cpu_disable_scheduler()
after removing the cpu from the mask, which will migrate the vcpu away.

Using a saved copy of cpu_valid for the test isn't working either, if a cpu
is added to and removed from the cpupool in a very short time. I would have
to use a generation count for the cpu_valid mask and do the check only if
the count didn't change. I think this is little bit of overkill here, but if
you'd prefer it, I can do a patch.


Juergen

-- 
Juergen Gross                 Principal Developer Operating Systems
TSP ES&S SWE OS6                       Telephone: +49 (0) 89 3222 2967
Fujitsu Technology Solutions              e-mail: juergen.gross@ts.fujitsu.com
Domagkstr. 28                           Internet: ts.fujitsu.com
D-80807 Muenchen                 Company details: ts.fujitsu.com/imprint.html

Re: [Patch] remove unneeded test for cpu in cpupool

[Keir Fraser]

On 16/03/2011 10:36, "Juergen Gross" <juergen.gross@ts.fujitsu.com> wrote:

> On 03/16/11 11:13, Keir Fraser wrote:
>> On 16/03/2011 06:34, "Juergen Gross"<juergen.gross@ts.fujitsu.com>  wrote:
>> 
>>> Hi,
>>> 
>>> attached patch removes an unneeded check in vcpu_migrate().
>>> 
>>> Andre, could you please give it a try on your 48 core machine with
>>> 
>>> xl cpupool-numa-split
>>> 
>>> I'd like to make sure the check is really not needed. On my machines it
>>> worked
>>> okay, but your machine seems to trigger races more easily.
>> 
>> Could you BUG_ON failing this test when we exit the retry loop? It seems a
>> suitable loop post-condition.
> 
> Not easily. The cpu_valid mask of a cpupool might change between the call of
> pick_cpu and the test. In the normal case this is okay, as removing a cpu
> from a cpupool (which is the critical case) will call cpu_disable_scheduler()
> after removing the cpu from the mask, which will migrate the vcpu away.
> 
> Using a saved copy of cpu_valid for the test isn't working either, if a cpu
> is added to and removed from the cpupool in a very short time. I would have
> to use a generation count for the cpu_valid mask and do the check only if
> the count didn't change. I think this is little bit of overkill here, but if
> you'd prefer it, I can do a patch.

Ah no, if it's more subtle than it appears, and it's not a straighjtforward
post-condition, then I wouldn't bother.

 -- Keir

> 
> Juergen