From: Rafael J. Wysocki <[email protected]>
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 <[email protected]>
Link: https://lore.kernel.org/linux-pm/[email protected]/
Signed-off-by: Rafael J. Wysocki <[email protected]>
---
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:
On Fri, 2023-03-03 at 20:19 +0100, Rafael J. Wysocki wrote:
> From: Rafael J. Wysocki <[email protected]>
>
> 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 <[email protected]>
> Link:
> https://lore.kernel.org/linux-pm/[email protected]/
> Signed-off-by: Rafael J. Wysocki <[email protected]>
> ---
> 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
On Tue, Mar 7, 2023 at 5:55 PM Zhang, Rui <[email protected]> wrote:
>
> On Fri, 2023-03-03 at 20:19 +0100, Rafael J. Wysocki wrote:
> > From: Rafael J. Wysocki <[email protected]>
> >
> > 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 <[email protected]>
> > Link:
> > https://lore.kernel.org/linux-pm/[email protected]/
> > Signed-off-by: Rafael J. Wysocki <[email protected]>
> > ---
> > 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.
Unless I've overlooked something tricky, it should be fine, but of
course verifying this independently won't hurt.
> I will test the whole patch series later this week.
Thank you!
On Fri, 2023-03-03 at 20:19 +0100, Rafael J. Wysocki wrote:
> From: Rafael J. Wysocki <[email protected]>
>
> 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 <[email protected]>
> Link:
> https://lore.kernel.org/linux-pm/[email protected]/
> Signed-off-by: Rafael J. Wysocki <[email protected]>
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:
>
>
>
On Sun, Mar 12, 2023 at 5:09 PM Zhang, Rui <[email protected]> wrote:
>
> On Fri, 2023-03-03 at 20:19 +0100, Rafael J. Wysocki wrote:
> > From: Rafael J. Wysocki <[email protected]>
> >
> > 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 <[email protected]>
> > Link:
> > https://lore.kernel.org/linux-pm/[email protected]/
> > Signed-off-by: Rafael J. Wysocki <[email protected]>
>
> 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:
> >
> >
> >
On Mon, 2023-03-13 at 14:48 +0100, Rafael J. Wysocki wrote:
> On Sun, Mar 12, 2023 at 5:09 PM Zhang, Rui <[email protected]>
> wrote:
> > On Fri, 2023-03-03 at 20:19 +0100, Rafael J. Wysocki wrote:
> > > From: Rafael J. Wysocki <[email protected]>
> > >
> > > 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 <[email protected]>
> > > Link:
> > > https://lore.kernel.org/linux-pm/[email protected]/
> > > Signed-off-by: Rafael J. Wysocki <[email protected]>
> >
> > 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:
> > >
> > >
> > >
On Mon, Mar 13, 2023 at 3:54 PM Zhang, Rui <[email protected]> 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 <[email protected]>
> > wrote:
> > > On Fri, 2023-03-03 at 20:19 +0100, Rafael J. Wysocki wrote:
> > > > From: Rafael J. Wysocki <[email protected]>
> > > >
> > > > 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 <[email protected]>
> > > > Link:
> > > > https://lore.kernel.org/linux-pm/[email protected]/
> > > > Signed-off-by: Rafael J. Wysocki <[email protected]>
> > >
> > > 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:
> > > >
> > > >
> > > >