[RFC,v2,11/35] arch_topology: Make register_cpu_capacity_sysctl() tolerant to late CPUs

Commit Message

James Morse Sept. 13, 2023, 4:37 p.m. UTC

register_cpu_capacity_sysctl() adds a property to sysfs that describes
the CPUs capacity. This is done from a subsys_initcall() that assumes
all possible CPUs are registered.

With CPU hotplug, possible CPUs aren't registered until they become
present, (or for arm64 enabled). This leads to messages during boot:
| register_cpu_capacity_sysctl: too early to get CPU1 device!
and once these CPUs are added to the system, the file is missing.

Move this to a cpuhp callback, so that the file is created once
CPUs are brought online. This covers CPUs that are added late by
mechanisms like hotplug.
One observable difference is the file is now missing for offline CPUs.

Signed-off-by: James Morse <[e-mail]>
---
If the offline CPUs thing is a problem for the tools that consume
this value, we'd need to move cpu_capacity to be part of cpu.c's
common_cpu_attr_groups.
---
 drivers/base/arch_topology.c | 38 ++++++++++++++++++++++++------------
 1 file changed, 26 insertions(+), 12 deletions(-)

Comments

Jonathan Cameron Sept. 14, 2023, 12:01 p.m. UTC | #1

On Wed, 13 Sep 2023 16:37:59 +0000
James Morse <[e-mail]> wrote:

> register_cpu_capacity_sysctl() adds a property to sysfs that describes
> the CPUs capacity. This is done from a subsys_initcall() that assumes
> all possible CPUs are registered.
> 
> With CPU hotplug, possible CPUs aren't registered until they become
> present, (or for arm64 enabled). This leads to messages during boot:
> | register_cpu_capacity_sysctl: too early to get CPU1 device!
> and once these CPUs are added to the system, the file is missing.
> 
> Move this to a cpuhp callback, so that the file is created once
> CPUs are brought online. This covers CPUs that are added late by
> mechanisms like hotplug.
> One observable difference is the file is now missing for offline CPUs.
> 
> Signed-off-by: James Morse <[e-mail]>
> ---
> If the offline CPUs thing is a problem for the tools that consume
> this value, we'd need to move cpu_capacity to be part of cpu.c's
> common_cpu_attr_groups.

I think we should do that anyway and then use an is_visible() if we want to
change whether it is visible in offline cpus.

Dynamic sysfs file creation is horrible - particularly when done
from an totally different file from where the rest of the attributes
are registered.  I'm curious what the history behind that is.

Whilst here, why is there a common_cpu_attr_groups which is
identical to the hotpluggable_cpu_attr_groups in base/cpu.c?


+CC GregKH
Given changes in drivers/base/

> ---
>  drivers/base/arch_topology.c | 38 ++++++++++++++++++++++++------------
>  1 file changed, 26 insertions(+), 12 deletions(-)
> 
> diff --git a/drivers/base/arch_topology.c b/drivers/base/arch_topology.c
> index b741b5ba82bd..9ccb7daee78e 100644
> --- a/drivers/base/arch_topology.c
> +++ b/drivers/base/arch_topology.c
> @@ -220,20 +220,34 @@ static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
>  
>  static DEVICE_ATTR_RO(cpu_capacity);
>  
> +static int cpu_capacity_sysctl_add(unsigned int cpu)
> +{
> +	struct device *cpu_dev = get_cpu_device(cpu);
> +
> +	if (!cpu_dev)
> +		return -ENOENT;
> +
> +	device_create_file(cpu_dev, &dev_attr_cpu_capacity);
> +
> +	return 0;
> +}
> +
> +static int cpu_capacity_sysctl_remove(unsigned int cpu)
> +{
> +	struct device *cpu_dev = get_cpu_device(cpu);
> +
> +	if (!cpu_dev)
> +		return -ENOENT;
> +
> +	device_remove_file(cpu_dev, &dev_attr_cpu_capacity);
> +
> +	return 0;
> +}
> +
>  static int register_cpu_capacity_sysctl(void)
>  {
> -	int i;
> -	struct device *cpu;
> -
> -	for_each_possible_cpu(i) {
> -		cpu = get_cpu_device(i);
> -		if (!cpu) {
> -			pr_err("%s: too early to get CPU%d device!\n",
> -			       __func__, i);
> -			continue;
> -		}
> -		device_create_file(cpu, &dev_attr_cpu_capacity);
> -	}
> +	cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "topology/cpu-capacity",
> +			  cpu_capacity_sysctl_add, cpu_capacity_sysctl_remove);
>  
>  	return 0;
>  }

Russell King (Oracle) Oct. 20, 2023, 11:53 a.m. UTC | #2

On Thu, Sep 14, 2023 at 01:01:26PM +0100, Jonathan Cameron wrote:
> On Wed, 13 Sep 2023 16:37:59 +0000
> James Morse <[e-mail]> wrote:
> 
> > register_cpu_capacity_sysctl() adds a property to sysfs that describes
> > the CPUs capacity. This is done from a subsys_initcall() that assumes
> > all possible CPUs are registered.
> > 
> > With CPU hotplug, possible CPUs aren't registered until they become
> > present, (or for arm64 enabled). This leads to messages during boot:
> > | register_cpu_capacity_sysctl: too early to get CPU1 device!
> > and once these CPUs are added to the system, the file is missing.
> > 
> > Move this to a cpuhp callback, so that the file is created once
> > CPUs are brought online. This covers CPUs that are added late by
> > mechanisms like hotplug.
> > One observable difference is the file is now missing for offline CPUs.
> > 
> > Signed-off-by: James Morse <[e-mail]>
> > ---
> > If the offline CPUs thing is a problem for the tools that consume
> > this value, we'd need to move cpu_capacity to be part of cpu.c's
> > common_cpu_attr_groups.
> 
> I think we should do that anyway and then use an is_visible() if we want to
> change whether it is visible in offline cpus.
> 
> Dynamic sysfs file creation is horrible - particularly when done
> from an totally different file from where the rest of the attributes
> are registered.  I'm curious what the history behind that is.
> 
> Whilst here, why is there a common_cpu_attr_groups which is
> identical to the hotpluggable_cpu_attr_groups in base/cpu.c?
> 
> 
> +CC GregKH
> Given changes in drivers/base/

It would be good to have a comment on this from Greg before I post
an updated series of James' patches with most of the comments
addressed, possibly later today.

Thanks.

Russell King (Oracle) Oct. 20, 2023, 1:44 p.m. UTC | #3

On Thu, Sep 14, 2023 at 01:01:26PM +0100, Jonathan Cameron wrote:
> On Wed, 13 Sep 2023 16:37:59 +0000
> James Morse <[e-mail]> wrote:
> 
> > register_cpu_capacity_sysctl() adds a property to sysfs that describes
> > the CPUs capacity. This is done from a subsys_initcall() that assumes
> > all possible CPUs are registered.
> > 
> > With CPU hotplug, possible CPUs aren't registered until they become
> > present, (or for arm64 enabled). This leads to messages during boot:
> > | register_cpu_capacity_sysctl: too early to get CPU1 device!
> > and once these CPUs are added to the system, the file is missing.
> > 
> > Move this to a cpuhp callback, so that the file is created once
> > CPUs are brought online. This covers CPUs that are added late by
> > mechanisms like hotplug.
> > One observable difference is the file is now missing for offline CPUs.
> > 
> > Signed-off-by: James Morse <[e-mail]>
> > ---
> > If the offline CPUs thing is a problem for the tools that consume
> > this value, we'd need to move cpu_capacity to be part of cpu.c's
> > common_cpu_attr_groups.
> 
> I think we should do that anyway and then use an is_visible() if we want to
> change whether it is visible in offline cpus.
> 
> Dynamic sysfs file creation is horrible - particularly when done
> from an totally different file from where the rest of the attributes
> are registered.  I'm curious what the history behind that is.
> 
> Whilst here, why is there a common_cpu_attr_groups which is
> identical to the hotpluggable_cpu_attr_groups in base/cpu.c?

Looking into doing this, the easy bit is adding the attribute group
with an appropriate .is_visible dependent on cpu_present(), but we
need to be able to call sysfs_update_groups() when the state of the
.is_visible() changes.

Given the comment in sysfs_update_groups() about "if an error occurs",
rather than making this part of common_cpu_attr_groups, would it be
better that it's part of its own set of groups, thus limiting the
damage from a possible error? I suspect, however, that any error at
that point means that the system is rather fatally wounded.

This is what I have so far to implement your idea, less the necessary
sysfs_update_groups() call when we need to change the visibility of
the attributes.

diff --git a/drivers/base/arch_topology.c b/drivers/base/arch_topology.c
index 9ccb7daee78e..06c9fc6620d2 100644
--- a/drivers/base/arch_topology.c
+++ b/drivers/base/arch_topology.c
@@ -215,43 +215,24 @@ static ssize_t cpu_capacity_show(struct device *dev,
 	return sysfs_emit(buf, "%lu\n", topology_get_cpu_scale(cpu->dev.id));
 }
 
-static void update_topology_flags_workfn(struct work_struct *work);
-static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
-
 static DEVICE_ATTR_RO(cpu_capacity);
 
-static int cpu_capacity_sysctl_add(unsigned int cpu)
-{
-	struct device *cpu_dev = get_cpu_device(cpu);
-
-	if (!cpu_dev)
-		return -ENOENT;
-
-	device_create_file(cpu_dev, &dev_attr_cpu_capacity);
-
-	return 0;
-}
-
-static int cpu_capacity_sysctl_remove(unsigned int cpu)
+static umode_t cpu_present_attrs_visible(struct kobject *kobi,
+					 struct attribute *attr, int index)
 {
-	struct device *cpu_dev = get_cpu_device(cpu);
-
-	if (!cpu_dev)
-		return -ENOENT;
-
-	device_remove_file(cpu_dev, &dev_attr_cpu_capacity);
+	struct device *dev = kobj_to_dev(kobj);
+	struct cpu *cpu = container_of(dev, struct cpu, dev);
 
-	return 0;
+	return cpu_present(cpu->dev.id) ? attr->mode : 0;
 }
 
-static int register_cpu_capacity_sysctl(void)
-{
-	cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "topology/cpu-capacity",
-			  cpu_capacity_sysctl_add, cpu_capacity_sysctl_remove);
+const struct attribute_group cpu_capacity_attr_group = {
+	.is_visible = cpu_present_attrs_visible,
+	.attrs = cpu_capacity_attrs
+};
 
-	return 0;
-}
-subsys_initcall(register_cpu_capacity_sysctl);
+static void update_topology_flags_workfn(struct work_struct *work);
+static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
 
 static int update_topology;
 
diff --git a/drivers/base/cpu.c b/drivers/base/cpu.c
index a19a8be93102..954b045705c2 100644
--- a/drivers/base/cpu.c
+++ b/drivers/base/cpu.c
@@ -192,6 +192,9 @@ static const struct attribute_group crash_note_cpu_attr_group = {
 static const struct attribute_group *common_cpu_attr_groups[] = {
 #ifdef CONFIG_KEXEC
 	&crash_note_cpu_attr_group,
+#endif
+#ifdef CONFIG_GENERIC_ARCH_TOPOLOGY
+	&cpu_capacity_attr_group,
 #endif
 	NULL
 };
diff --git a/include/linux/cpu.h b/include/linux/cpu.h
index e117c06e0c6b..745ad21e3dc8 100644
--- a/include/linux/cpu.h
+++ b/include/linux/cpu.h
@@ -30,6 +30,8 @@ struct cpu {
 	struct device dev;
 };
 
+extern const struct attribute_group cpu_capacity_attr_group;
+
 extern void boot_cpu_init(void);
 extern void boot_cpu_hotplug_init(void);
 extern void cpu_init(void);

Greg Kroah-Hartman Oct. 21, 2023, 10:56 a.m. UTC | #4

On Fri, Oct 20, 2023 at 12:53:29PM +0100, Russell King (Oracle) wrote:
> On Thu, Sep 14, 2023 at 01:01:26PM +0100, Jonathan Cameron wrote:
> > On Wed, 13 Sep 2023 16:37:59 +0000
> > James Morse <[e-mail]> wrote:
> > 
> > > register_cpu_capacity_sysctl() adds a property to sysfs that describes
> > > the CPUs capacity. This is done from a subsys_initcall() that assumes
> > > all possible CPUs are registered.
> > > 
> > > With CPU hotplug, possible CPUs aren't registered until they become
> > > present, (or for arm64 enabled). This leads to messages during boot:
> > > | register_cpu_capacity_sysctl: too early to get CPU1 device!
> > > and once these CPUs are added to the system, the file is missing.
> > > 
> > > Move this to a cpuhp callback, so that the file is created once
> > > CPUs are brought online. This covers CPUs that are added late by
> > > mechanisms like hotplug.
> > > One observable difference is the file is now missing for offline CPUs.
> > > 
> > > Signed-off-by: James Morse <[e-mail]>
> > > ---
> > > If the offline CPUs thing is a problem for the tools that consume
> > > this value, we'd need to move cpu_capacity to be part of cpu.c's
> > > common_cpu_attr_groups.
> > 
> > I think we should do that anyway and then use an is_visible() if we want to
> > change whether it is visible in offline cpus.
> > 
> > Dynamic sysfs file creation is horrible - particularly when done
> > from an totally different file from where the rest of the attributes
> > are registered.  I'm curious what the history behind that is.
> > 
> > Whilst here, why is there a common_cpu_attr_groups which is
> > identical to the hotpluggable_cpu_attr_groups in base/cpu.c?
> > 
> > 
> > +CC GregKH
> > Given changes in drivers/base/
> 
> It would be good to have a comment on this from Greg before I post
> an updated series of James' patches with most of the comments
> addressed, possibly later today.

Sorry, I don't see what I am supposed to comment on, so please just send
a new series and I'll look at that.

thanks,

greg k-h

Russell King (Oracle) Nov. 14, 2023, 10:04 a.m. UTC | #5

Okay, after 25 days, I'm now changing James' comment to:

    If the offline CPUs thing is a problem for the tools that consume
    this value, we'd need to move cpu_capacity to be part of cpu.c's
    common_cpu_attr_groups. However, attempts to discuss this just end
    up in a black hole, so this is a non-starter. Thus, if this needs
    to be done, it can be done as a separate patch.

and thus I'm going to consider this patch acceptable to everyone.

On Fri, Oct 20, 2023 at 02:44:35PM +0100, Russell King (Oracle) wrote:
> On Thu, Sep 14, 2023 at 01:01:26PM +0100, Jonathan Cameron wrote:
> > On Wed, 13 Sep 2023 16:37:59 +0000
> > James Morse <[e-mail]> wrote:
> > 
> > > register_cpu_capacity_sysctl() adds a property to sysfs that describes
> > > the CPUs capacity. This is done from a subsys_initcall() that assumes
> > > all possible CPUs are registered.
> > > 
> > > With CPU hotplug, possible CPUs aren't registered until they become
> > > present, (or for arm64 enabled). This leads to messages during boot:
> > > | register_cpu_capacity_sysctl: too early to get CPU1 device!
> > > and once these CPUs are added to the system, the file is missing.
> > > 
> > > Move this to a cpuhp callback, so that the file is created once
> > > CPUs are brought online. This covers CPUs that are added late by
> > > mechanisms like hotplug.
> > > One observable difference is the file is now missing for offline CPUs.
> > > 
> > > Signed-off-by: James Morse <[e-mail]>
> > > ---
> > > If the offline CPUs thing is a problem for the tools that consume
> > > this value, we'd need to move cpu_capacity to be part of cpu.c's
> > > common_cpu_attr_groups.
> > 
> > I think we should do that anyway and then use an is_visible() if we want to
> > change whether it is visible in offline cpus.
> > 
> > Dynamic sysfs file creation is horrible - particularly when done
> > from an totally different file from where the rest of the attributes
> > are registered.  I'm curious what the history behind that is.
> > 
> > Whilst here, why is there a common_cpu_attr_groups which is
> > identical to the hotpluggable_cpu_attr_groups in base/cpu.c?
> 
> Looking into doing this, the easy bit is adding the attribute group
> with an appropriate .is_visible dependent on cpu_present(), but we
> need to be able to call sysfs_update_groups() when the state of the
> .is_visible() changes.
> 
> Given the comment in sysfs_update_groups() about "if an error occurs",
> rather than making this part of common_cpu_attr_groups, would it be
> better that it's part of its own set of groups, thus limiting the
> damage from a possible error? I suspect, however, that any error at
> that point means that the system is rather fatally wounded.
> 
> This is what I have so far to implement your idea, less the necessary
> sysfs_update_groups() call when we need to change the visibility of
> the attributes.
> 
> diff --git a/drivers/base/arch_topology.c b/drivers/base/arch_topology.c
> index 9ccb7daee78e..06c9fc6620d2 100644
> --- a/drivers/base/arch_topology.c
> +++ b/drivers/base/arch_topology.c
> @@ -215,43 +215,24 @@ static ssize_t cpu_capacity_show(struct device *dev,
>  	return sysfs_emit(buf, "%lu\n", topology_get_cpu_scale(cpu->dev.id));
>  }
>  
> -static void update_topology_flags_workfn(struct work_struct *work);
> -static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
> -
>  static DEVICE_ATTR_RO(cpu_capacity);
>  
> -static int cpu_capacity_sysctl_add(unsigned int cpu)
> -{
> -	struct device *cpu_dev = get_cpu_device(cpu);
> -
> -	if (!cpu_dev)
> -		return -ENOENT;
> -
> -	device_create_file(cpu_dev, &dev_attr_cpu_capacity);
> -
> -	return 0;
> -}
> -
> -static int cpu_capacity_sysctl_remove(unsigned int cpu)
> +static umode_t cpu_present_attrs_visible(struct kobject *kobi,
> +					 struct attribute *attr, int index)
>  {
> -	struct device *cpu_dev = get_cpu_device(cpu);
> -
> -	if (!cpu_dev)
> -		return -ENOENT;
> -
> -	device_remove_file(cpu_dev, &dev_attr_cpu_capacity);
> +	struct device *dev = kobj_to_dev(kobj);
> +	struct cpu *cpu = container_of(dev, struct cpu, dev);
>  
> -	return 0;
> +	return cpu_present(cpu->dev.id) ? attr->mode : 0;
>  }
>  
> -static int register_cpu_capacity_sysctl(void)
> -{
> -	cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "topology/cpu-capacity",
> -			  cpu_capacity_sysctl_add, cpu_capacity_sysctl_remove);
> +const struct attribute_group cpu_capacity_attr_group = {
> +	.is_visible = cpu_present_attrs_visible,
> +	.attrs = cpu_capacity_attrs
> +};
>  
> -	return 0;
> -}
> -subsys_initcall(register_cpu_capacity_sysctl);
> +static void update_topology_flags_workfn(struct work_struct *work);
> +static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
>  
>  static int update_topology;
>  
> diff --git a/drivers/base/cpu.c b/drivers/base/cpu.c
> index a19a8be93102..954b045705c2 100644
> --- a/drivers/base/cpu.c
> +++ b/drivers/base/cpu.c
> @@ -192,6 +192,9 @@ static const struct attribute_group crash_note_cpu_attr_group = {
>  static const struct attribute_group *common_cpu_attr_groups[] = {
>  #ifdef CONFIG_KEXEC
>  	&crash_note_cpu_attr_group,
> +#endif
> +#ifdef CONFIG_GENERIC_ARCH_TOPOLOGY
> +	&cpu_capacity_attr_group,
>  #endif
>  	NULL
>  };
> diff --git a/include/linux/cpu.h b/include/linux/cpu.h
> index e117c06e0c6b..745ad21e3dc8 100644
> --- a/include/linux/cpu.h
> +++ b/include/linux/cpu.h
> @@ -30,6 +30,8 @@ struct cpu {
>  	struct device dev;
>  };
>  
> +extern const struct attribute_group cpu_capacity_attr_group;
> +
>  extern void boot_cpu_init(void);
>  extern void boot_cpu_hotplug_init(void);
>  extern void cpu_init(void);
> 
> -- 
> RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
> FTTP is here! 80Mbps down 10Mbps up. Decent connectivity at last!

Jonathan Cameron Nov. 30, 2023, 4:46 p.m. UTC | #6

On Fri, 20 Oct 2023 14:44:35 +0100
"Russell King (Oracle)" <[e-mail]> wrote:

> On Thu, Sep 14, 2023 at 01:01:26PM +0100, Jonathan Cameron wrote:
> > On Wed, 13 Sep 2023 16:37:59 +0000
> > James Morse <[e-mail]> wrote:
> >   
> > > register_cpu_capacity_sysctl() adds a property to sysfs that describes
> > > the CPUs capacity. This is done from a subsys_initcall() that assumes
> > > all possible CPUs are registered.
> > > 
> > > With CPU hotplug, possible CPUs aren't registered until they become
> > > present, (or for arm64 enabled). This leads to messages during boot:
> > > | register_cpu_capacity_sysctl: too early to get CPU1 device!
> > > and once these CPUs are added to the system, the file is missing.
> > > 
> > > Move this to a cpuhp callback, so that the file is created once
> > > CPUs are brought online. This covers CPUs that are added late by
> > > mechanisms like hotplug.
> > > One observable difference is the file is now missing for offline CPUs.
> > > 
> > > Signed-off-by: James Morse <[e-mail]>
> > > ---
> > > If the offline CPUs thing is a problem for the tools that consume
> > > this value, we'd need to move cpu_capacity to be part of cpu.c's
> > > common_cpu_attr_groups.  
> > 
> > I think we should do that anyway and then use an is_visible() if we want to
> > change whether it is visible in offline cpus.
> > 
> > Dynamic sysfs file creation is horrible - particularly when done
> > from an totally different file from where the rest of the attributes
> > are registered.  I'm curious what the history behind that is.
> > 
> > Whilst here, why is there a common_cpu_attr_groups which is
> > identical to the hotpluggable_cpu_attr_groups in base/cpu.c?  
> 
> Looking into doing this, the easy bit is adding the attribute group
> with an appropriate .is_visible dependent on cpu_present(), but we
> need to be able to call sysfs_update_groups() when the state of the
> .is_visible() changes.
Hi Russell,

Sorry, somehow I missed this completely until you referred back to it :(

This is pretty much what I was thinking so thanks for doing it.

> 
> Given the comment in sysfs_update_groups() about "if an error occurs",
> rather than making this part of common_cpu_attr_groups, would it be
> better that it's part of its own set of groups, thus limiting the
> damage from a possible error? I suspect, however, that any error at
> that point means that the system is rather fatally wounded.
> 
> This is what I have so far to implement your idea, less the necessary
> sysfs_update_groups() call when we need to change the visibility of
> the attributes.

Fwiw (and I think you shouldn't add this to the critical path for your
main series for obvious reasons), I think you are right that it makes
sense to do this in a separate group, but that if we were going to see
an error I'd 'hope' we shouldn't see anything that hasn't occurred
when groups were originally added. Maybe that's overly optimistic.

Sorry again for lack of reply before now and thanks for pointing this
out.  I'd love to see this posted after the ARM vCPU HP stuff is in.

Jonathan


> 
> diff --git a/drivers/base/arch_topology.c b/drivers/base/arch_topology.c
> index 9ccb7daee78e..06c9fc6620d2 100644
> --- a/drivers/base/arch_topology.c
> +++ b/drivers/base/arch_topology.c
> @@ -215,43 +215,24 @@ static ssize_t cpu_capacity_show(struct device *dev,
>  	return sysfs_emit(buf, "%lu\n", topology_get_cpu_scale(cpu->dev.id));
>  }
>  
> -static void update_topology_flags_workfn(struct work_struct *work);
> -static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
> -
>  static DEVICE_ATTR_RO(cpu_capacity);
>  
> -static int cpu_capacity_sysctl_add(unsigned int cpu)
> -{
> -	struct device *cpu_dev = get_cpu_device(cpu);
> -
> -	if (!cpu_dev)
> -		return -ENOENT;
> -
> -	device_create_file(cpu_dev, &dev_attr_cpu_capacity);
> -
> -	return 0;
> -}
> -
> -static int cpu_capacity_sysctl_remove(unsigned int cpu)
> +static umode_t cpu_present_attrs_visible(struct kobject *kobi,
> +					 struct attribute *attr, int index)
>  {
> -	struct device *cpu_dev = get_cpu_device(cpu);
> -
> -	if (!cpu_dev)
> -		return -ENOENT;
> -
> -	device_remove_file(cpu_dev, &dev_attr_cpu_capacity);
> +	struct device *dev = kobj_to_dev(kobj);
> +	struct cpu *cpu = container_of(dev, struct cpu, dev);
>  
> -	return 0;
> +	return cpu_present(cpu->dev.id) ? attr->mode : 0;
>  }
>  
> -static int register_cpu_capacity_sysctl(void)
> -{
> -	cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "topology/cpu-capacity",
> -			  cpu_capacity_sysctl_add, cpu_capacity_sysctl_remove);
> +const struct attribute_group cpu_capacity_attr_group = {
> +	.is_visible = cpu_present_attrs_visible,
> +	.attrs = cpu_capacity_attrs
> +};
>  
> -	return 0;
> -}
> -subsys_initcall(register_cpu_capacity_sysctl);
> +static void update_topology_flags_workfn(struct work_struct *work);
> +static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
>  
>  static int update_topology;
>  
> diff --git a/drivers/base/cpu.c b/drivers/base/cpu.c
> index a19a8be93102..954b045705c2 100644
> --- a/drivers/base/cpu.c
> +++ b/drivers/base/cpu.c
> @@ -192,6 +192,9 @@ static const struct attribute_group crash_note_cpu_attr_group = {
>  static const struct attribute_group *common_cpu_attr_groups[] = {
>  #ifdef CONFIG_KEXEC
>  	&crash_note_cpu_attr_group,
> +#endif
> +#ifdef CONFIG_GENERIC_ARCH_TOPOLOGY
> +	&cpu_capacity_attr_group,
>  #endif
>  	NULL
>  };
> diff --git a/include/linux/cpu.h b/include/linux/cpu.h
> index e117c06e0c6b..745ad21e3dc8 100644
> --- a/include/linux/cpu.h
> +++ b/include/linux/cpu.h
> @@ -30,6 +30,8 @@ struct cpu {
>  	struct device dev;
>  };
>  
> +extern const struct attribute_group cpu_capacity_attr_group;
> +
>  extern void boot_cpu_init(void);
>  extern void boot_cpu_hotplug_init(void);
>  extern void cpu_init(void);
>

Patch

diff --git a/drivers/base/arch_topology.c b/drivers/base/arch_topology.c
index b741b5ba82bd..9ccb7daee78e 100644
--- a/drivers/base/arch_topology.c
+++ b/drivers/base/arch_topology.c
@@ -220,20 +220,34 @@  static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
 
 static DEVICE_ATTR_RO(cpu_capacity);
 
+static int cpu_capacity_sysctl_add(unsigned int cpu)
+{
+	struct device *cpu_dev = get_cpu_device(cpu);
+
+	if (!cpu_dev)
+		return -ENOENT;
+
+	device_create_file(cpu_dev, &dev_attr_cpu_capacity);
+
+	return 0;
+}
+
+static int cpu_capacity_sysctl_remove(unsigned int cpu)
+{
+	struct device *cpu_dev = get_cpu_device(cpu);
+
+	if (!cpu_dev)
+		return -ENOENT;
+
+	device_remove_file(cpu_dev, &dev_attr_cpu_capacity);
+
+	return 0;
+}
+
 static int register_cpu_capacity_sysctl(void)
 {
-	int i;
-	struct device *cpu;
-
-	for_each_possible_cpu(i) {
-		cpu = get_cpu_device(i);
-		if (!cpu) {
-			pr_err("%s: too early to get CPU%d device!\n",
-			       __func__, i);
-			continue;
-		}
-		device_create_file(cpu, &dev_attr_cpu_capacity);
-	}
+	cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "topology/cpu-capacity",
+			  cpu_capacity_sysctl_add, cpu_capacity_sysctl_remove);
 
 	return 0;
 }