[RFC] cpufreq: dt-platdev: Fix module autoloading

Commit Message

Javier Martinez Canillas Nov. 19, 2024, 11:18 a.m. UTC

This driver can be built as a module since commit 3b062a086984 ("cpufreq:
dt-platdev: Support building as module"), but unfortunately this caused
a regression because the cputfreq-dt-platdev.ko module does not autoload.

Usually, this is solved by just using the MODULE_DEVICE_TABLE() macro to
export all the device IDs as module aliases. But this driver is special
due how matches with devices and decides what platform supports.

There are two of_device_id lists, an allow list that are for CPU devices
that always match and a deny list that's for devices that must not match.

The driver registers a cpufreq-dt platform device for all the CPU device
nodes that either are in the allow list or contain an operating-points-v2
property and are not in the deny list.

For the former just add a MODULE_DEVICE_TABLE(), and for the latter add a
module alias. That way the driver would always be autoloaded when needed.

Reported-by: Radu Rendec <[e-mail]>
Fixes: 3b062a086984 ("cpufreq: dt-platdev: Support building as module")
Signed-off-by: Javier Martinez Canillas <[e-mail]>
---
Posting as an RFC because I don't have a platform that uses this driver
but I'll let Radu test since he reported by issue.

 drivers/cpufreq/cpufreq-dt-platdev.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)

Comments

Viresh Kumar Nov. 21, 2024, 7:11 a.m. UTC | #1

+Rob/Arnd,

On 19-11-24, 12:18, Javier Martinez Canillas wrote:
> This driver can be built as a module since commit 3b062a086984 ("cpufreq:
> dt-platdev: Support building as module"), but unfortunately this caused
> a regression because the cputfreq-dt-platdev.ko module does not autoload.
> 
> Usually, this is solved by just using the MODULE_DEVICE_TABLE() macro to
> export all the device IDs as module aliases. But this driver is special
> due how matches with devices and decides what platform supports.
> 
> There are two of_device_id lists, an allow list that are for CPU devices
> that always match and a deny list that's for devices that must not match.
> 
> The driver registers a cpufreq-dt platform device for all the CPU device
> nodes that either are in the allow list or contain an operating-points-v2
> property and are not in the deny list.
> 
> For the former just add a MODULE_DEVICE_TABLE(), and for the latter add a
> module alias. That way the driver would always be autoloaded when needed.
> 
> Reported-by: Radu Rendec <[e-mail]>
> Fixes: 3b062a086984 ("cpufreq: dt-platdev: Support building as module")
> Signed-off-by: Javier Martinez Canillas <[e-mail]>
> ---
> Posting as an RFC because I don't have a platform that uses this driver
> but I'll let Radu test since he reported by issue.
> 
>  drivers/cpufreq/cpufreq-dt-platdev.c | 13 +++++++++++++
>  1 file changed, 13 insertions(+)
> 
> diff --git a/drivers/cpufreq/cpufreq-dt-platdev.c b/drivers/cpufreq/cpufreq-dt-platdev.c
> index 2a3e8bd317c9..7ae7c897c249 100644
> --- a/drivers/cpufreq/cpufreq-dt-platdev.c
> +++ b/drivers/cpufreq/cpufreq-dt-platdev.c
> @@ -97,6 +97,7 @@ static const struct of_device_id allowlist[] __initconst = {
>  
>  	{ }
>  };
> +MODULE_DEVICE_TABLE(of, allowlist);

This is fine obviously.

>  /*
>   * Machines for which the cpufreq device is *not* created, mostly used for
> @@ -236,4 +237,16 @@ static int __init cpufreq_dt_platdev_init(void)
>  }
>  core_initcall(cpufreq_dt_platdev_init);
>  MODULE_DESCRIPTION("Generic DT based cpufreq platdev driver");
> +/*
> + * The module alias is needed because the driver automatically registers a
> + * platform device for any CPU device node that has an operating-points-v2
> + * property and is not in the block list.
> + *
> + * For this reason the MODULE_DEVICE_TABLE() macro can only export aliases
> + * of the devices in the allow list, which means that the driver will not
> + * autoload for devices whose cpufreq-dt will be registered automatically.
> + *
> + * Adding an "of:N*T*Coperating-points-v2" alias is a workaround for this.
> + */
> +MODULE_ALIAS("of:N*T*Coperating-points-v2");

How does this work? This will autoload the module for any platform with
"operating-points-v2" property in the DT ? The driver though works only if the
property is in the CPU node, while now we will try to load this driver even if a
non-CPU node has the property now.

I am not sure what's the best way forward to fix this.

Arnd, Rob, any inputs ?

>  MODULE_LICENSE("GPL");

Patch

diff --git a/drivers/cpufreq/cpufreq-dt-platdev.c b/drivers/cpufreq/cpufreq-dt-platdev.c
index 2a3e8bd317c9..7ae7c897c249 100644
--- a/drivers/cpufreq/cpufreq-dt-platdev.c
+++ b/drivers/cpufreq/cpufreq-dt-platdev.c
@@ -97,6 +97,7 @@  static const struct of_device_id allowlist[] __initconst = {
 
 	{ }
 };
+MODULE_DEVICE_TABLE(of, allowlist);
 
 /*
  * Machines for which the cpufreq device is *not* created, mostly used for
@@ -236,4 +237,16 @@  static int __init cpufreq_dt_platdev_init(void)
 }
 core_initcall(cpufreq_dt_platdev_init);
 MODULE_DESCRIPTION("Generic DT based cpufreq platdev driver");
+/*
+ * The module alias is needed because the driver automatically registers a
+ * platform device for any CPU device node that has an operating-points-v2
+ * property and is not in the block list.
+ *
+ * For this reason the MODULE_DEVICE_TABLE() macro can only export aliases
+ * of the devices in the allow list, which means that the driver will not
+ * autoload for devices whose cpufreq-dt will be registered automatically.
+ *
+ * Adding an "of:N*T*Coperating-points-v2" alias is a workaround for this.
+ */
+MODULE_ALIAS("of:N*T*Coperating-points-v2");
 MODULE_LICENSE("GPL");