diff mbox series

[v1,1/4] ACPI: processor: Reorder acpi_processor_driver_init()

Message ID 2885079.e9J7NaK4W3@kreacher
State Superseded
Headers show
Series thermal: core/ACPI: Fix processor cooling device regression | expand

Commit Message

Rafael J. Wysocki March 3, 2023, 7:19 p.m. UTC
From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

The cpufreq policy notifier in the ACPI processor driver may as
well be registered before the driver itself, which causes
acpi_processor_cpufreq_init to be true (unless the notifier
registration fails, which is unlikely at that point) when the
ACPI CPU thermal cooling devices are registered, so the
processor_get_max_state() result does not change while
acpi_processor_driver_init() is running.

Change the ordering in acpi_processor_driver_init() accordingly
to prevent the max_state value from remaining 0 permanently for all
ACPI CPU cooling devices.

Fixes: a365105c685c("thermal: sysfs: Reuse cdev->max_state")
Reported-by: Wang, Quanxian <quanxian.wang@intel.com>
Link: https://lore.kernel.org/linux-pm/53ec1f06f61c984100868926f282647e57ecfb2d.camel@intel.com/
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
---
 drivers/acpi/processor_driver.c |   12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

Comments

Zhang, Rui March 7, 2023, 4:51 p.m. UTC | #1
On Fri, 2023-03-03 at 20:19 +0100, Rafael J. Wysocki wrote:
> From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> 
> The cpufreq policy notifier in the ACPI processor driver may as
> well be registered before the driver itself, which causes
> acpi_processor_cpufreq_init to be true (unless the notifier
> registration fails, which is unlikely at that point) when the
> ACPI CPU thermal cooling devices are registered, so the
> processor_get_max_state() result does not change while
> acpi_processor_driver_init() is running.
> 
> Change the ordering in acpi_processor_driver_init() accordingly
> to prevent the max_state value from remaining 0 permanently for all
> ACPI CPU cooling devices.
> 
> Fixes: a365105c685c("thermal: sysfs: Reuse cdev->max_state")
> Reported-by: Wang, Quanxian <quanxian.wang@intel.com>
> Link: 
> https://lore.kernel.org/linux-pm/53ec1f06f61c984100868926f282647e57ecfb2d.camel@intel.com/
> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> ---
>  drivers/acpi/processor_driver.c |   12 ++++++------
>  1 file changed, 6 insertions(+), 6 deletions(-)
> 
> Index: linux-pm/drivers/acpi/processor_driver.c
> ===================================================================
> --- linux-pm.orig/drivers/acpi/processor_driver.c
> +++ linux-pm/drivers/acpi/processor_driver.c
> @@ -263,6 +263,12 @@ static int __init acpi_processor_driver_
>  	if (acpi_disabled)
>  		return 0;
>  
> +	if (!cpufreq_register_notifier(&acpi_processor_notifier_block,
> +				       CPUFREQ_POLICY_NOTIFIER)) {
> +		acpi_processor_cpufreq_init = true;
> +		acpi_processor_ignore_ppc_init();
> +	}
> +
>  	result = driver_register(&acpi_processor_driver);
>  	if (result < 0)
>  		return result;
> @@ -276,12 +282,6 @@ static int __init acpi_processor_driver_
>  	cpuhp_setup_state_nocalls(CPUHP_ACPI_CPUDRV_DEAD, "acpi/cpu-
> drv:dead",
>  				  NULL, acpi_soft_cpu_dead);
>  
> -	if (!cpufreq_register_notifier(&acpi_processor_notifier_block,
> -				       CPUFREQ_POLICY_NOTIFIER)) {
> -		acpi_processor_cpufreq_init = true;
> -		acpi_processor_ignore_ppc_init();
> -	}
> -
>  	acpi_processor_throttling_init();
>  	return 0;
>  err:
> 
Just FYI.
I need some time to ramp up on the ordering here to double confirm this
does not break any dependency, too many things are involved here :p.

I will test the whole patch series later this week.

thanks,
rui
Zhang, Rui March 12, 2023, 4:08 p.m. UTC | #2
On Fri, 2023-03-03 at 20:19 +0100, Rafael J. Wysocki wrote:
> From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> 
> The cpufreq policy notifier in the ACPI processor driver may as
> well be registered before the driver itself, which causes
> acpi_processor_cpufreq_init to be true (unless the notifier
> registration fails, which is unlikely at that point) when the
> ACPI CPU thermal cooling devices are registered, so the
> processor_get_max_state() result does not change while
> acpi_processor_driver_init() is running.
> 
> Change the ordering in acpi_processor_driver_init() accordingly
> to prevent the max_state value from remaining 0 permanently for all
> ACPI CPU cooling devices.
> 
> Fixes: a365105c685c("thermal: sysfs: Reuse cdev->max_state")
> Reported-by: Wang, Quanxian <quanxian.wang@intel.com>
> Link: 
> https://lore.kernel.org/linux-pm/53ec1f06f61c984100868926f282647e57ecfb2d.camel@intel.com/
> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

The full patch series fixes the problem but this one does not.

This is because,

static int cpu_has_cpufreq(unsigned int cpu)
{
        struct
cpufreq_policy *policy;

        if (!acpi_processor_cpufreq_init)
                return 0;

        policy = cpufreq_cpu_get(cpu);
        if (policy) {
                cpufreq_cpu_put(policy);
                return 1;
        }
        return 0;
}

Although acpi_processor_cpufreq_init is set to true with patch 1/4, but
we don't have cpufreq driver registered, thus cpufreq_cpu_get() return
NULL.

so acpi_processor_cpufreq_init is not the only dependency here. :(

thanks,
rui

> ---
>  drivers/acpi/processor_driver.c |   12 ++++++------
>  1 file changed, 6 insertions(+), 6 deletions(-)
> 
> Index: linux-pm/drivers/acpi/processor_driver.c
> ===================================================================
> --- linux-pm.orig/drivers/acpi/processor_driver.c
> +++ linux-pm/drivers/acpi/processor_driver.c
> @@ -263,6 +263,12 @@ static int __init acpi_processor_driver_
>  	if (acpi_disabled)
>  		return 0;
>  
> +	if (!cpufreq_register_notifier(&acpi_processor_notifier_block,
> +				       CPUFREQ_POLICY_NOTIFIER)) {
> +		acpi_processor_cpufreq_init = true;
> +		acpi_processor_ignore_ppc_init();
> +	}
> +
>  	result = driver_register(&acpi_processor_driver);
>  	if (result < 0)
>  		return result;
> @@ -276,12 +282,6 @@ static int __init acpi_processor_driver_
>  	cpuhp_setup_state_nocalls(CPUHP_ACPI_CPUDRV_DEAD, "acpi/cpu-
> drv:dead",
>  				  NULL, acpi_soft_cpu_dead);
>  
> -	if (!cpufreq_register_notifier(&acpi_processor_notifier_block,
> -				       CPUFREQ_POLICY_NOTIFIER)) {
> -		acpi_processor_cpufreq_init = true;
> -		acpi_processor_ignore_ppc_init();
> -	}
> -
>  	acpi_processor_throttling_init();
>  	return 0;
>  err:
> 
> 
>
Rafael J. Wysocki March 13, 2023, 1:48 p.m. UTC | #3
On Sun, Mar 12, 2023 at 5:09 PM Zhang, Rui <rui.zhang@intel.com> wrote:
>
> On Fri, 2023-03-03 at 20:19 +0100, Rafael J. Wysocki wrote:
> > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> >
> > The cpufreq policy notifier in the ACPI processor driver may as
> > well be registered before the driver itself, which causes
> > acpi_processor_cpufreq_init to be true (unless the notifier
> > registration fails, which is unlikely at that point) when the
> > ACPI CPU thermal cooling devices are registered, so the
> > processor_get_max_state() result does not change while
> > acpi_processor_driver_init() is running.
> >
> > Change the ordering in acpi_processor_driver_init() accordingly
> > to prevent the max_state value from remaining 0 permanently for all
> > ACPI CPU cooling devices.
> >
> > Fixes: a365105c685c("thermal: sysfs: Reuse cdev->max_state")
> > Reported-by: Wang, Quanxian <quanxian.wang@intel.com>
> > Link:
> > https://lore.kernel.org/linux-pm/53ec1f06f61c984100868926f282647e57ecfb2d.camel@intel.com/
> > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
>
> The full patch series fixes the problem but this one does not.

That is a correct observation, but the $subject patch fixes part of
the problem (which is not addressed by the rest of the series AFAICS)
and so it deserves a Fixes tag of its own IMO.

I guess I should clarify that in the changelog.

> This is because,
>
> static int cpu_has_cpufreq(unsigned int cpu)
> {
>         struct
> cpufreq_policy *policy;
>
>         if (!acpi_processor_cpufreq_init)
>                 return 0;
>
>         policy = cpufreq_cpu_get(cpu);
>         if (policy) {
>                 cpufreq_cpu_put(policy);
>                 return 1;
>         }
>         return 0;
> }
>
> Although acpi_processor_cpufreq_init is set to true with patch 1/4, but
> we don't have cpufreq driver registered, thus cpufreq_cpu_get() return
> NULL.
> so acpi_processor_cpufreq_init is not the only dependency here. :(

Right.  That's why the other patches in the series are needed too.

> > ---
> >  drivers/acpi/processor_driver.c |   12 ++++++------
> >  1 file changed, 6 insertions(+), 6 deletions(-)
> >
> > Index: linux-pm/drivers/acpi/processor_driver.c
> > ===================================================================
> > --- linux-pm.orig/drivers/acpi/processor_driver.c
> > +++ linux-pm/drivers/acpi/processor_driver.c
> > @@ -263,6 +263,12 @@ static int __init acpi_processor_driver_
> >       if (acpi_disabled)
> >               return 0;
> >
> > +     if (!cpufreq_register_notifier(&acpi_processor_notifier_block,
> > +                                    CPUFREQ_POLICY_NOTIFIER)) {
> > +             acpi_processor_cpufreq_init = true;
> > +             acpi_processor_ignore_ppc_init();
> > +     }
> > +
> >       result = driver_register(&acpi_processor_driver);
> >       if (result < 0)
> >               return result;
> > @@ -276,12 +282,6 @@ static int __init acpi_processor_driver_
> >       cpuhp_setup_state_nocalls(CPUHP_ACPI_CPUDRV_DEAD, "acpi/cpu-
> > drv:dead",
> >                                 NULL, acpi_soft_cpu_dead);
> >
> > -     if (!cpufreq_register_notifier(&acpi_processor_notifier_block,
> > -                                    CPUFREQ_POLICY_NOTIFIER)) {
> > -             acpi_processor_cpufreq_init = true;
> > -             acpi_processor_ignore_ppc_init();
> > -     }
> > -
> >       acpi_processor_throttling_init();
> >       return 0;
> >  err:
> >
> >
> >
Zhang, Rui March 13, 2023, 2:54 p.m. UTC | #4
On Mon, 2023-03-13 at 14:48 +0100, Rafael J. Wysocki wrote:
> On Sun, Mar 12, 2023 at 5:09 PM Zhang, Rui <rui.zhang@intel.com>
> wrote:
> > On Fri, 2023-03-03 at 20:19 +0100, Rafael J. Wysocki wrote:
> > > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > > 
> > > The cpufreq policy notifier in the ACPI processor driver may as
> > > well be registered before the driver itself, which causes
> > > acpi_processor_cpufreq_init to be true (unless the notifier
> > > registration fails, which is unlikely at that point) when the
> > > ACPI CPU thermal cooling devices are registered, so the
> > > processor_get_max_state() result does not change while
> > > acpi_processor_driver_init() is running.
> > > 
> > > Change the ordering in acpi_processor_driver_init() accordingly
> > > to prevent the max_state value from remaining 0 permanently for
> > > all
> > > ACPI CPU cooling devices.
> > > 
> > > Fixes: a365105c685c("thermal: sysfs: Reuse cdev->max_state")
> > > Reported-by: Wang, Quanxian <quanxian.wang@intel.com>
> > > Link:
> > > https://lore.kernel.org/linux-pm/53ec1f06f61c984100868926f282647e57ecfb2d.camel@intel.com/
> > > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > 
> > The full patch series fixes the problem but this one does not.
> 
> That is a correct observation, but the $subject patch fixes part of
> the problem (which is not addressed by the rest of the series AFAICS)
> and so it deserves a Fixes tag of its own IMO.

Am I understanding this correctly that this patch helps in below case?

cpufreq driver like intel_pstate is registered before we register the
notifier callback in processor_driver. In this case, we are not able to
catch the CPUFREQ_CREATE_POLICY notification and cpufreq should be
counted as part of cooling states when registering the ACPI CPU cooling
device. (acpi_processor_cpufreq_init must be set at this time)

Or else, in normal case, the ACPI CPU cdev->max_state always return 0
(when t-state not available) until we receive the CPUFREQ_CREATE_POLICY
notification and call thermal_cooling_device_update(), both with and
without this patch.

Patch 2,3,4 works on my test platform, without applying patch 1/4.

thanks,
rui

> 
> I guess I should clarify that in the changelog.
> 
> > This is because,
> > 
> > static int cpu_has_cpufreq(unsigned int cpu)
> > {
> >         struct
> > cpufreq_policy *policy;
> > 
> >         if (!acpi_processor_cpufreq_init)
> >                 return 0;
> > 
> >         policy = cpufreq_cpu_get(cpu);
> >         if (policy) {
> >                 cpufreq_cpu_put(policy);
> >                 return 1;
> >         }
> >         return 0;
> > }
> > 
> > Although acpi_processor_cpufreq_init is set to true with patch 1/4,
> > but
> > we don't have cpufreq driver registered, thus cpufreq_cpu_get()
> > return
> > NULL.
> > so acpi_processor_cpufreq_init is not the only dependency here. :(
> 
> Right.  That's why the other patches in the series are needed too.
> 
> > > ---
> > >  drivers/acpi/processor_driver.c |   12 ++++++------
> > >  1 file changed, 6 insertions(+), 6 deletions(-)
> > > 
> > > Index: linux-pm/drivers/acpi/processor_driver.c
> > > =================================================================
> > > ==
> > > --- linux-pm.orig/drivers/acpi/processor_driver.c
> > > +++ linux-pm/drivers/acpi/processor_driver.c
> > > @@ -263,6 +263,12 @@ static int __init acpi_processor_driver_
> > >       if (acpi_disabled)
> > >               return 0;
> > > 
> > > +     if
> > > (!cpufreq_register_notifier(&acpi_processor_notifier_block,
> > > +                                    CPUFREQ_POLICY_NOTIFIER)) {
> > > +             acpi_processor_cpufreq_init = true;
> > > +             acpi_processor_ignore_ppc_init();
> > > +     }
> > > +
> > >       result = driver_register(&acpi_processor_driver);
> > >       if (result < 0)
> > >               return result;
> > > @@ -276,12 +282,6 @@ static int __init acpi_processor_driver_
> > >       cpuhp_setup_state_nocalls(CPUHP_ACPI_CPUDRV_DEAD,
> > > "acpi/cpu-
> > > drv:dead",
> > >                                 NULL, acpi_soft_cpu_dead);
> > > 
> > > -     if
> > > (!cpufreq_register_notifier(&acpi_processor_notifier_block,
> > > -                                    CPUFREQ_POLICY_NOTIFIER)) {
> > > -             acpi_processor_cpufreq_init = true;
> > > -             acpi_processor_ignore_ppc_init();
> > > -     }
> > > -
> > >       acpi_processor_throttling_init();
> > >       return 0;
> > >  err:
> > > 
> > > 
> > >
Rafael J. Wysocki March 13, 2023, 2:58 p.m. UTC | #5
On Mon, Mar 13, 2023 at 3:54 PM Zhang, Rui <rui.zhang@intel.com> wrote:
>
> On Mon, 2023-03-13 at 14:48 +0100, Rafael J. Wysocki wrote:
> > On Sun, Mar 12, 2023 at 5:09 PM Zhang, Rui <rui.zhang@intel.com>
> > wrote:
> > > On Fri, 2023-03-03 at 20:19 +0100, Rafael J. Wysocki wrote:
> > > > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > > >
> > > > The cpufreq policy notifier in the ACPI processor driver may as
> > > > well be registered before the driver itself, which causes
> > > > acpi_processor_cpufreq_init to be true (unless the notifier
> > > > registration fails, which is unlikely at that point) when the
> > > > ACPI CPU thermal cooling devices are registered, so the
> > > > processor_get_max_state() result does not change while
> > > > acpi_processor_driver_init() is running.
> > > >
> > > > Change the ordering in acpi_processor_driver_init() accordingly
> > > > to prevent the max_state value from remaining 0 permanently for
> > > > all
> > > > ACPI CPU cooling devices.
> > > >
> > > > Fixes: a365105c685c("thermal: sysfs: Reuse cdev->max_state")
> > > > Reported-by: Wang, Quanxian <quanxian.wang@intel.com>
> > > > Link:
> > > > https://lore.kernel.org/linux-pm/53ec1f06f61c984100868926f282647e57ecfb2d.camel@intel.com/
> > > > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > >
> > > The full patch series fixes the problem but this one does not.
> >
> > That is a correct observation, but the $subject patch fixes part of
> > the problem (which is not addressed by the rest of the series AFAICS)
> > and so it deserves a Fixes tag of its own IMO.
>
> Am I understanding this correctly that this patch helps in below case?
>
> cpufreq driver like intel_pstate is registered before we register the
> notifier callback in processor_driver. In this case, we are not able to
> catch the CPUFREQ_CREATE_POLICY notification and cpufreq should be
> counted as part of cooling states when registering the ACPI CPU cooling
> device. (acpi_processor_cpufreq_init must be set at this time)

Yes.

> Or else, in normal case, the ACPI CPU cdev->max_state always return 0
> (when t-state not available) until we receive the CPUFREQ_CREATE_POLICY
> notification and call thermal_cooling_device_update(), both with and
> without this patch.
>
> Patch 2,3,4 works on my test platform, without applying patch 1/4.

OK

> > I guess I should clarify that in the changelog.
> >
> > > This is because,
> > >
> > > static int cpu_has_cpufreq(unsigned int cpu)
> > > {
> > >         struct
> > > cpufreq_policy *policy;
> > >
> > >         if (!acpi_processor_cpufreq_init)
> > >                 return 0;
> > >
> > >         policy = cpufreq_cpu_get(cpu);
> > >         if (policy) {
> > >                 cpufreq_cpu_put(policy);
> > >                 return 1;
> > >         }
> > >         return 0;
> > > }
> > >
> > > Although acpi_processor_cpufreq_init is set to true with patch 1/4,
> > > but
> > > we don't have cpufreq driver registered, thus cpufreq_cpu_get()
> > > return
> > > NULL.
> > > so acpi_processor_cpufreq_init is not the only dependency here. :(
> >
> > Right.  That's why the other patches in the series are needed too.
> >
> > > > ---
> > > >  drivers/acpi/processor_driver.c |   12 ++++++------
> > > >  1 file changed, 6 insertions(+), 6 deletions(-)
> > > >
> > > > Index: linux-pm/drivers/acpi/processor_driver.c
> > > > =================================================================
> > > > ==
> > > > --- linux-pm.orig/drivers/acpi/processor_driver.c
> > > > +++ linux-pm/drivers/acpi/processor_driver.c
> > > > @@ -263,6 +263,12 @@ static int __init acpi_processor_driver_
> > > >       if (acpi_disabled)
> > > >               return 0;
> > > >
> > > > +     if
> > > > (!cpufreq_register_notifier(&acpi_processor_notifier_block,
> > > > +                                    CPUFREQ_POLICY_NOTIFIER)) {
> > > > +             acpi_processor_cpufreq_init = true;
> > > > +             acpi_processor_ignore_ppc_init();
> > > > +     }
> > > > +
> > > >       result = driver_register(&acpi_processor_driver);
> > > >       if (result < 0)
> > > >               return result;
> > > > @@ -276,12 +282,6 @@ static int __init acpi_processor_driver_
> > > >       cpuhp_setup_state_nocalls(CPUHP_ACPI_CPUDRV_DEAD,
> > > > "acpi/cpu-
> > > > drv:dead",
> > > >                                 NULL, acpi_soft_cpu_dead);
> > > >
> > > > -     if
> > > > (!cpufreq_register_notifier(&acpi_processor_notifier_block,
> > > > -                                    CPUFREQ_POLICY_NOTIFIER)) {
> > > > -             acpi_processor_cpufreq_init = true;
> > > > -             acpi_processor_ignore_ppc_init();
> > > > -     }
> > > > -
> > > >       acpi_processor_throttling_init();
> > > >       return 0;
> > > >  err:
> > > >
> > > >
> > > >
diff mbox series

Patch

Index: linux-pm/drivers/acpi/processor_driver.c
===================================================================
--- linux-pm.orig/drivers/acpi/processor_driver.c
+++ linux-pm/drivers/acpi/processor_driver.c
@@ -263,6 +263,12 @@  static int __init acpi_processor_driver_
 	if (acpi_disabled)
 		return 0;
 
+	if (!cpufreq_register_notifier(&acpi_processor_notifier_block,
+				       CPUFREQ_POLICY_NOTIFIER)) {
+		acpi_processor_cpufreq_init = true;
+		acpi_processor_ignore_ppc_init();
+	}
+
 	result = driver_register(&acpi_processor_driver);
 	if (result < 0)
 		return result;
@@ -276,12 +282,6 @@  static int __init acpi_processor_driver_
 	cpuhp_setup_state_nocalls(CPUHP_ACPI_CPUDRV_DEAD, "acpi/cpu-drv:dead",
 				  NULL, acpi_soft_cpu_dead);
 
-	if (!cpufreq_register_notifier(&acpi_processor_notifier_block,
-				       CPUFREQ_POLICY_NOTIFIER)) {
-		acpi_processor_cpufreq_init = true;
-		acpi_processor_ignore_ppc_init();
-	}
-
 	acpi_processor_throttling_init();
 	return 0;
 err: