Re: [PATCH -next] arm64: Export __cpu_logical_map

Re: [PATCH -next] arm64: Export __cpu_logical_map

[Sumit Gupta]

>>>>> ERROR: modpost: "__cpu_logical_map" [drivers/cpufreq/tegra194-cpufreq.ko] undefined!
>>>>>
>>>>> ARM64 tegra194-cpufreq driver use cpu_logical_map, export
>>>>> __cpu_logical_map to fix build issue.
>>>>>
>>>
>>> I wonder why like other instances in the drivers, the mpidr is not get
>>> directly from the cpu. The cpufreq_driver->init call happens when the cpu
>>> is being brought online and is executed on the required cpu IIUC.
>>>
>> Yes, this occurs during hotplug case.
>> But in the case of system boot, 'cpufreq_driver->init' is called later
>> during cpufreq platform driver's probe. The value of CPU in 'policy->cpu'
>> can be different from the current CPU. That's why read_cpuid_mpidr() can't
>> be used.
>>
> 
> Fair enough, why not do cross call like in set_target ? Since it is one-off
> in init, I don't see any issue when you are doing it runtime for set_target.
> 
>>> read_cpuid_mpidr() is inline and avoids having to export the logical_cpu_map.
>>> Though we may not add physical hotplug anytime soon, less dependency
>>> on this cpu_logical_map is better given that we can resolve this without
>>> the need to access the map.
>>>
> 
> To be honest, we have tried to remove all the dependency on cluster id
> in generic code as it is not well defined. This one is tegra specific
> driver so should be fine. But I am still bit nervous to export
> cpu_logical_map as we have no clue what that would mean for physical
> hotplug.
> 
As suggested, I have done below change to get the cluster number using 
read_cpuid_mpidr(). Please review and suggest if this looks ok?
I will send formal patch if the change is fine.

Thanks,
Sumit

----

diff --git a/drivers/cpufreq/tegra194-cpufreq.c 
b/drivers/cpufreq/tegra194-cpufreq.c
index bae527e..06f5ccf 100644
--- a/drivers/cpufreq/tegra194-cpufreq.c
+++ b/drivers/cpufreq/tegra194-cpufreq.c
@@ -56,9 +56,11 @@ struct read_counters_work {

  static struct workqueue_struct *read_counters_wq;

-static enum cluster get_cpu_cluster(u8 cpu)
+static void get_cpu_cluster(void *cluster)
  {
-       return MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu), 1);
+       u64 mpidr = read_cpuid_mpidr() & MPIDR_HWID_BITMASK;
+
+       *((uint32_t *) cluster) = MPIDR_AFFINITY_LEVEL(mpidr, 1);
  }

  /*
@@ -186,8 +188,10 @@ static unsigned int tegra194_get_speed(u32 cpu)
  static int tegra194_cpufreq_init(struct cpufreq_policy *policy)
  {
         struct tegra194_cpufreq_data *data = cpufreq_get_driver_data();
-       int cl = get_cpu_cluster(policy->cpu);
         u32 cpu;
+       u32 cl;
+
+       smp_call_function_single(policy->cpu, get_cpu_cluster, &cl, true);


> 
> --
> Regards,
> Sudeep
> 

Re: [PATCH -next] arm64: Export __cpu_logical_map

[Sudeep Holla]

On Sat, Aug 01, 2020 at 05:46:43PM +0530, Sumit Gupta wrote:
>
> > > > > > ERROR: modpost: "__cpu_logical_map" [drivers/cpufreq/tegra194-cpufreq.ko] undefined!
> > > > > >
> > > > > > ARM64 tegra194-cpufreq driver use cpu_logical_map, export
> > > > > > __cpu_logical_map to fix build issue.
> > > > > >
> > > >
> > > > I wonder why like other instances in the drivers, the mpidr is not get
> > > > directly from the cpu. The cpufreq_driver->init call happens when the cpu
> > > > is being brought online and is executed on the required cpu IIUC.
> > > >
> > > Yes, this occurs during hotplug case.
> > > But in the case of system boot, 'cpufreq_driver->init' is called later
> > > during cpufreq platform driver's probe. The value of CPU in 'policy->cpu'
> > > can be different from the current CPU. That's why read_cpuid_mpidr() can't
> > > be used.
> > >
> >
> > Fair enough, why not do cross call like in set_target ? Since it is one-off
> > in init, I don't see any issue when you are doing it runtime for set_target.
> >
> > > > read_cpuid_mpidr() is inline and avoids having to export the logical_cpu_map.
> > > > Though we may not add physical hotplug anytime soon, less dependency
> > > > on this cpu_logical_map is better given that we can resolve this without
> > > > the need to access the map.
> > > >
> >
> > To be honest, we have tried to remove all the dependency on cluster id
> > in generic code as it is not well defined. This one is tegra specific
> > driver so should be fine. But I am still bit nervous to export
> > cpu_logical_map as we have no clue what that would mean for physical
> > hotplug.
> >
> As suggested, I have done below change to get the cluster number using
> read_cpuid_mpidr(). Please review and suggest if this looks ok?
> I will send formal patch if the change is fine.
>
> Thanks,
> Sumit
>
> ----
>
> diff --git a/drivers/cpufreq/tegra194-cpufreq.c
> b/drivers/cpufreq/tegra194-cpufreq.c
> index bae527e..06f5ccf 100644
> --- a/drivers/cpufreq/tegra194-cpufreq.c
> +++ b/drivers/cpufreq/tegra194-cpufreq.c
> @@ -56,9 +56,11 @@ struct read_counters_work {
>
>  static struct workqueue_struct *read_counters_wq;
>
> -static enum cluster get_cpu_cluster(u8 cpu)
> +static void get_cpu_cluster(void *cluster)
>  {
> -       return MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu), 1);
> +       u64 mpidr = read_cpuid_mpidr() & MPIDR_HWID_BITMASK;
> +
> +       *((uint32_t *) cluster) = MPIDR_AFFINITY_LEVEL(mpidr, 1);
>  }
>
>  /*
> @@ -186,8 +188,10 @@ static unsigned int tegra194_get_speed(u32 cpu)
>  static int tegra194_cpufreq_init(struct cpufreq_policy *policy)
>  {
>         struct tegra194_cpufreq_data *data = cpufreq_get_driver_data();
> -       int cl = get_cpu_cluster(policy->cpu);
>         u32 cpu;
> +       u32 cl;
> +
> +       smp_call_function_single(policy->cpu, get_cpu_cluster, &cl, true);

Thanks for this, looks good to me. You can add:

Reviewed-by: Sudeep Holla <sudeep.holla@arm.com>

--
Regards,
Sudeep

Re: [PATCH -next] arm64: Export __cpu_logical_map

[Catalin Marinas]

On Mon, Aug 10, 2020 at 08:49:56AM +0100, Sudeep Holla wrote:
> On Sat, Aug 01, 2020 at 05:46:43PM +0530, Sumit Gupta wrote:
> > > > > > > ERROR: modpost: "__cpu_logical_map" [drivers/cpufreq/tegra194-cpufreq.ko] undefined!
> > > > > > >
> > > > > > > ARM64 tegra194-cpufreq driver use cpu_logical_map, export
> > > > > > > __cpu_logical_map to fix build issue.
> > > > > > >
> > > > >
> > > > > I wonder why like other instances in the drivers, the mpidr is not get
> > > > > directly from the cpu. The cpufreq_driver->init call happens when the cpu
> > > > > is being brought online and is executed on the required cpu IIUC.
> > > >
> > > > Yes, this occurs during hotplug case.
> > > > But in the case of system boot, 'cpufreq_driver->init' is called later
> > > > during cpufreq platform driver's probe. The value of CPU in 'policy->cpu'
> > > > can be different from the current CPU. That's why read_cpuid_mpidr() can't
> > > > be used.
> > >
> > > Fair enough, why not do cross call like in set_target ? Since it is one-off
> > > in init, I don't see any issue when you are doing it runtime for set_target.
> > >
> > > > > read_cpuid_mpidr() is inline and avoids having to export the logical_cpu_map.
> > > > > Though we may not add physical hotplug anytime soon, less dependency
> > > > > on this cpu_logical_map is better given that we can resolve this without
> > > > > the need to access the map.
> > >
> > > To be honest, we have tried to remove all the dependency on cluster id
> > > in generic code as it is not well defined. This one is tegra specific
> > > driver so should be fine. But I am still bit nervous to export
> > > cpu_logical_map as we have no clue what that would mean for physical
> > > hotplug.
> >
> > As suggested, I have done below change to get the cluster number using
> > read_cpuid_mpidr(). Please review and suggest if this looks ok?
> > I will send formal patch if the change is fine.
> >
> > diff --git a/drivers/cpufreq/tegra194-cpufreq.c
> > b/drivers/cpufreq/tegra194-cpufreq.c
> > index bae527e..06f5ccf 100644
> > --- a/drivers/cpufreq/tegra194-cpufreq.c
> > +++ b/drivers/cpufreq/tegra194-cpufreq.c
> > @@ -56,9 +56,11 @@ struct read_counters_work {
> >
> >  static struct workqueue_struct *read_counters_wq;
> >
> > -static enum cluster get_cpu_cluster(u8 cpu)
> > +static void get_cpu_cluster(void *cluster)
> >  {
> > -       return MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu), 1);
> > +       u64 mpidr = read_cpuid_mpidr() & MPIDR_HWID_BITMASK;
> > +
> > +       *((uint32_t *) cluster) = MPIDR_AFFINITY_LEVEL(mpidr, 1);
> >  }
> >
> >  /*
> > @@ -186,8 +188,10 @@ static unsigned int tegra194_get_speed(u32 cpu)
> >  static int tegra194_cpufreq_init(struct cpufreq_policy *policy)
> >  {
> >         struct tegra194_cpufreq_data *data = cpufreq_get_driver_data();
> > -       int cl = get_cpu_cluster(policy->cpu);
> >         u32 cpu;
> > +       u32 cl;
> > +
> > +       smp_call_function_single(policy->cpu, get_cpu_cluster, &cl, true);
> 
> Thanks for this, looks good to me. You can add:
> 
> Reviewed-by: Sudeep Holla <sudeep.holla@arm.com>

I already merged Kefeng's __cpu_logical_map fix (commit eaecca9e7710)
but if the above goes in, I can drop the EXPORT_SYMBOL part (and keep
the rest as it's a good refactoring).

-- 
Catalin

Re: [PATCH -next] arm64: Export __cpu_logical_map

[Sudeep Holla]

On Mon, Aug 10, 2020 at 11:19:55AM +0100, Catalin Marinas wrote:
> On Mon, Aug 10, 2020 at 08:49:56AM +0100, Sudeep Holla wrote:
> > On Sat, Aug 01, 2020 at 05:46:43PM +0530, Sumit Gupta wrote:
> > > > > > > > ERROR: modpost: "__cpu_logical_map" [drivers/cpufreq/tegra194-cpufreq.ko] undefined!
> > > > > > > >
> > > > > > > > ARM64 tegra194-cpufreq driver use cpu_logical_map, export
> > > > > > > > __cpu_logical_map to fix build issue.
> > > > > > > >
> > > > > >
> > > > > > I wonder why like other instances in the drivers, the mpidr is not get
> > > > > > directly from the cpu. The cpufreq_driver->init call happens when the cpu
> > > > > > is being brought online and is executed on the required cpu IIUC.
> > > > >
> > > > > Yes, this occurs during hotplug case.
> > > > > But in the case of system boot, 'cpufreq_driver->init' is called later
> > > > > during cpufreq platform driver's probe. The value of CPU in 'policy->cpu'
> > > > > can be different from the current CPU. That's why read_cpuid_mpidr() can't
> > > > > be used.
> > > >
> > > > Fair enough, why not do cross call like in set_target ? Since it is one-off
> > > > in init, I don't see any issue when you are doing it runtime for set_target.
> > > >
> > > > > > read_cpuid_mpidr() is inline and avoids having to export the logical_cpu_map.
> > > > > > Though we may not add physical hotplug anytime soon, less dependency
> > > > > > on this cpu_logical_map is better given that we can resolve this without
> > > > > > the need to access the map.
> > > >
> > > > To be honest, we have tried to remove all the dependency on cluster id
> > > > in generic code as it is not well defined. This one is tegra specific
> > > > driver so should be fine. But I am still bit nervous to export
> > > > cpu_logical_map as we have no clue what that would mean for physical
> > > > hotplug.
> > >
> > > As suggested, I have done below change to get the cluster number using
> > > read_cpuid_mpidr(). Please review and suggest if this looks ok?
> > > I will send formal patch if the change is fine.
> > >
> > > diff --git a/drivers/cpufreq/tegra194-cpufreq.c
> > > b/drivers/cpufreq/tegra194-cpufreq.c
> > > index bae527e..06f5ccf 100644
> > > --- a/drivers/cpufreq/tegra194-cpufreq.c
> > > +++ b/drivers/cpufreq/tegra194-cpufreq.c
> > > @@ -56,9 +56,11 @@ struct read_counters_work {
> > >
> > >  static struct workqueue_struct *read_counters_wq;
> > >
> > > -static enum cluster get_cpu_cluster(u8 cpu)
> > > +static void get_cpu_cluster(void *cluster)
> > >  {
> > > -       return MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu), 1);
> > > +       u64 mpidr = read_cpuid_mpidr() & MPIDR_HWID_BITMASK;
> > > +
> > > +       *((uint32_t *) cluster) = MPIDR_AFFINITY_LEVEL(mpidr, 1);
> > >  }
> > >
> > >  /*
> > > @@ -186,8 +188,10 @@ static unsigned int tegra194_get_speed(u32 cpu)
> > >  static int tegra194_cpufreq_init(struct cpufreq_policy *policy)
> > >  {
> > >         struct tegra194_cpufreq_data *data = cpufreq_get_driver_data();
> > > -       int cl = get_cpu_cluster(policy->cpu);
> > >         u32 cpu;
> > > +       u32 cl;
> > > +
> > > +       smp_call_function_single(policy->cpu, get_cpu_cluster, &cl, true);
> > 
> > Thanks for this, looks good to me. You can add:
> > 
> > Reviewed-by: Sudeep Holla <sudeep.holla@arm.com>
> 
> I already merged Kefeng's __cpu_logical_map fix (commit eaecca9e7710)
> but if the above goes in, I can drop the EXPORT_SYMBOL part (and keep
> the rest as it's a good refactoring).
> 

OK, I will keep an eye on this and we can drop export once this is merged.

-- 
Regards,
Sudeep

Re: [PATCH -next] arm64: Export __cpu_logical_map

[Sumit Gupta]

>>>>>>>> ERROR: modpost: "__cpu_logical_map" [drivers/cpufreq/tegra194-cpufreq.ko] undefined!
>>>>>>>>
>>>>>>>> ARM64 tegra194-cpufreq driver use cpu_logical_map, export
>>>>>>>> __cpu_logical_map to fix build issue.
>>>>>>>>
>>>>>>
>>>>>> I wonder why like other instances in the drivers, the mpidr is not get
>>>>>> directly from the cpu. The cpufreq_driver->init call happens when the cpu
>>>>>> is being brought online and is executed on the required cpu IIUC.
>>>>>
>>>>> Yes, this occurs during hotplug case.
>>>>> But in the case of system boot, 'cpufreq_driver->init' is called later
>>>>> during cpufreq platform driver's probe. The value of CPU in 'policy->cpu'
>>>>> can be different from the current CPU. That's why read_cpuid_mpidr() can't
>>>>> be used.
>>>>
>>>> Fair enough, why not do cross call like in set_target ? Since it is one-off
>>>> in init, I don't see any issue when you are doing it runtime for set_target.
>>>>
>>>>>> read_cpuid_mpidr() is inline and avoids having to export the logical_cpu_map.
>>>>>> Though we may not add physical hotplug anytime soon, less dependency
>>>>>> on this cpu_logical_map is better given that we can resolve this without
>>>>>> the need to access the map.
>>>>
>>>> To be honest, we have tried to remove all the dependency on cluster id
>>>> in generic code as it is not well defined. This one is tegra specific
>>>> driver so should be fine. But I am still bit nervous to export
>>>> cpu_logical_map as we have no clue what that would mean for physical
>>>> hotplug.
>>>
>>> As suggested, I have done below change to get the cluster number using
>>> read_cpuid_mpidr(). Please review and suggest if this looks ok?
>>> I will send formal patch if the change is fine.
>>>
>>> diff --git a/drivers/cpufreq/tegra194-cpufreq.c
>>> b/drivers/cpufreq/tegra194-cpufreq.c
>>> index bae527e..06f5ccf 100644
>>> --- a/drivers/cpufreq/tegra194-cpufreq.c
>>> +++ b/drivers/cpufreq/tegra194-cpufreq.c
>>> @@ -56,9 +56,11 @@ struct read_counters_work {
>>>
>>>   static struct workqueue_struct *read_counters_wq;
>>>
>>> -static enum cluster get_cpu_cluster(u8 cpu)
>>> +static void get_cpu_cluster(void *cluster)
>>>   {
>>> -       return MPIDR_AFFINITY_LEVEL(cpu_logical_map(cpu), 1);
>>> +       u64 mpidr = read_cpuid_mpidr() & MPIDR_HWID_BITMASK;
>>> +
>>> +       *((uint32_t *) cluster) = MPIDR_AFFINITY_LEVEL(mpidr, 1);
>>>   }
>>>
>>>   /*
>>> @@ -186,8 +188,10 @@ static unsigned int tegra194_get_speed(u32 cpu)
>>>   static int tegra194_cpufreq_init(struct cpufreq_policy *policy)
>>>   {
>>>          struct tegra194_cpufreq_data *data = cpufreq_get_driver_data();
>>> -       int cl = get_cpu_cluster(policy->cpu);
>>>          u32 cpu;
>>> +       u32 cl;
>>> +
>>> +       smp_call_function_single(policy->cpu, get_cpu_cluster, &cl, true);
>>
>> Thanks for this, looks good to me. You can add:
>>
>> Reviewed-by: Sudeep Holla <sudeep.holla@arm.com>
> 
> I already merged Kefeng's __cpu_logical_map fix (commit eaecca9e7710)
> but if the above goes in, I can drop the EXPORT_SYMBOL part (and keep
> the rest as it's a good refactoring).
> 
I have posted the formal patch.

Thanks,
Sumit

> --
> Catalin
>