Xen devel - Jul 2013 - [PATCH] xen/arm: disable cpuidle when linux is running as dom0

When linux is running as dom0, Xen doesn''t show the physical cpu but a
virtual CPU.
On some ARM SOC (for instance the exynos 5250), linux registers callbacks
for cpuidle. When these callbacks are called, they will modify
directly the physical cpu not the virtual one. It can impact the whole board
instead of dom0.

Signed-off-by: Julien Grall <julien.grall@linaro.org>
---
 arch/arm/xen/enlighten.c |    7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/arch/arm/xen/enlighten.c b/arch/arm/xen/enlighten.c
index 49839d8..a98999f 100644
--- a/arch/arm/xen/enlighten.c
+++ b/arch/arm/xen/enlighten.c
@@ -24,6 +24,8 @@
 #include <linux/of.h>
 #include <linux/of_irq.h>
 #include <linux/of_address.h>
+#include <linux/cpuidle.h>
+#include <linux/cpufreq.h>
 
 #include <linux/mm.h>
 
@@ -292,6 +294,11 @@ static int __init xen_pm_init(void)
 {
 	pm_power_off = xen_power_off;
 	arm_pm_restart = xen_restart;
+	/*
+	 * Making sure board specific code will not set up ops for
+	 * cpu idle.
+	 */
+	disable_cpuidle();
 
 	return 0;
 }
-- 
1.7.10.4

On Mon, Jul 15, 2013 at 03:21:41PM +0100, Julien Grall
wrote:
> When linux is running as dom0, Xen doesn''t show the physical cpu
but a
> virtual CPU.
> On some ARM SOC (for instance the exynos 5250), linux registers callbacks
> for cpuidle. When these callbacks are called, they will modify
> directly the physical cpu not the virtual one. It can impact the whole
board
> instead of dom0.
Should you also call disable_cpufreq() ?
> 
> Signed-off-by: Julien Grall <julien.grall@linaro.org>
> ---
>  arch/arm/xen/enlighten.c |    7 +++++++
>  1 file changed, 7 insertions(+)
> 
> diff --git a/arch/arm/xen/enlighten.c b/arch/arm/xen/enlighten.c
> index 49839d8..a98999f 100644
> --- a/arch/arm/xen/enlighten.c
> +++ b/arch/arm/xen/enlighten.c
> @@ -24,6 +24,8 @@
>  #include <linux/of.h>
>  #include <linux/of_irq.h>
>  #include <linux/of_address.h>
> +#include <linux/cpuidle.h>
> +#include <linux/cpufreq.h>
>  
>  #include <linux/mm.h>
>  
> @@ -292,6 +294,11 @@ static int __init xen_pm_init(void)
>  {
>  	pm_power_off = xen_power_off;
>  	arm_pm_restart = xen_restart;
> +	/*
> +	 * Making sure board specific code will not set up ops for
> +	 * cpu idle.
> +	 */
> +	disable_cpuidle();
>  
>  	return 0;
>  }
> -- 
> 1.7.10.4
> 
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xen.org
> http://lists.xen.org/xen-devel

On 07/15/2013 04:25 PM, Konrad Rzeszutek Wilk wrote:
> On Mon, Jul 15, 2013 at 03:21:41PM +0100, Julien Grall wrote:
>> When linux is running as dom0, Xen doesn''t show the physical
cpu but a
>> virtual CPU.
>> On some ARM SOC (for instance the exynos 5250), linux registers
callbacks
>> for cpuidle. When these callbacks are called, they will modify
>> directly the physical cpu not the virtual one. It can impact the whole
board
>> instead of dom0.
> 
> Should you also call disable_cpufreq() ?
I had some issue on the versatile express when cpufreq was disabled.
I will give another try and see the exact error.

-- 
Julien

On Mon, 15 Jul 2013, Konrad Rzeszutek Wilk wrote:
> On Mon, Jul 15, 2013 at 03:21:41PM +0100, Julien Grall wrote:
> > When linux is running as dom0, Xen doesn''t show the physical
cpu but a
> > virtual CPU.
> > On some ARM SOC (for instance the exynos 5250), linux registers
callbacks
> > for cpuidle. When these callbacks are called, they will modify
> > directly the physical cpu not the virtual one.
> It can impact the whole board
> > instead of dom0.
Certainly this is something that should be fixed at the hypervisor level
too. However I agree that Linux should try to avoid doing that when
running on Xen.


> Should you also call disable_cpufreq() ?
Sounds like a good idea.
Julien, could you add that to this patch?

> > 
> > Signed-off-by: Julien Grall <julien.grall@linaro.org>
> > ---
> >  arch/arm/xen/enlighten.c |    7 +++++++
> >  1 file changed, 7 insertions(+)
> > 
> > diff --git a/arch/arm/xen/enlighten.c b/arch/arm/xen/enlighten.c
> > index 49839d8..a98999f 100644
> > --- a/arch/arm/xen/enlighten.c
> > +++ b/arch/arm/xen/enlighten.c
> > @@ -24,6 +24,8 @@
> >  #include <linux/of.h>
> >  #include <linux/of_irq.h>
> >  #include <linux/of_address.h>
> > +#include <linux/cpuidle.h>
> > +#include <linux/cpufreq.h>
> >  
> >  #include <linux/mm.h>
> >  
> > @@ -292,6 +294,11 @@ static int __init xen_pm_init(void)
> >  {
> >  	pm_power_off = xen_power_off;
> >  	arm_pm_restart = xen_restart;
> > +	/*
> > +	 * Making sure board specific code will not set up ops for
> > +	 * cpu idle.
> > +	 */
> > +	disable_cpuidle();
> >  
> >  	return 0;
> >  }
> > -- 
> > 1.7.10.4
> > 
> > 
> > _______________________________________________
> > Xen-devel mailing list
> > Xen-devel@lists.xen.org
> > http://lists.xen.org/xen-devel
>

On 17 July 2013 14:28, Stefano Stabellini
<stefano.stabellini@eu.citrix.com> wrote:
> On Mon, 15 Jul 2013, Konrad Rzeszutek Wilk wrote:
>> On Mon, Jul 15, 2013 at 03:21:41PM +0100, Julien Grall wrote:
>> > When linux is running as dom0, Xen doesn''t show the
physical cpu but a
>> > virtual CPU.
>> > On some ARM SOC (for instance the exynos 5250), linux registers
callbacks
>> > for cpuidle. When these callbacks are called, they will modify
>> > directly the physical cpu not the virtual one.
>> It can impact the whole board
>> > instead of dom0.
>
> Certainly this is something that should be fixed at the hypervisor level
> too. However I agree that Linux should try to avoid doing that when
> running on Xen.
>
>
>
>> Should you also call disable_cpufreq() ?
>
> Sounds like a good idea.
> Julien, could you add that to this patch?
I have checked with both the Arndale and Versatile Express without any issue.
I will resend this patch with disable_cpufreq().

-- 
Julien Grall