Re: [PATCH v9 3/5] EINJ: Migrate to a platform driver

[Ben Cheatham <benjamin.cheatham@amd.com>]

Hi Dan, thanks for the quick response! Comments inline.

On 1/16/24 5:47 PM, Dan Williams wrote:
> Ben Cheatham wrote:
>> Change the EINJ module to install a platform device/driver on module
>> init and move the module init() and exit() functions to driver probe and
>> remove. This change allows the EINJ module to load regardless of whether
>> setting up EINJ succeeds, which allows dependent modules to still load
>> (i.e. the CXL core).
>>
>> Since EINJ may no longer be initialized when the module loads, any
>> functions that are called from dependent/external modules should check
>> the einj_initialized variable before calling any EINJ functions.
>>
>> Signed-off-by: Ben Cheatham <Benjamin.Cheatham@amd.com>
>> ---
>>  drivers/acpi/apei/einj.c | 43 ++++++++++++++++++++++++++++++++++++++--
>>  1 file changed, 41 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/acpi/apei/einj.c b/drivers/acpi/apei/einj.c
>> index 013eb621dc92..10d51cd73fa4 100644
>> --- a/drivers/acpi/apei/einj.c
>> +++ b/drivers/acpi/apei/einj.c
>> @@ -21,6 +21,7 @@
>>  #include <linux/nmi.h>
>>  #include <linux/delay.h>
>>  #include <linux/mm.h>
>> +#include <linux/platform_device.h>
>>  #include <asm/unaligned.h>
>>  
>>  #include "apei-internal.h"
>> @@ -136,6 +137,12 @@ static struct apei_exec_ins_type einj_ins_type[] = {
>>   */
>>  static DEFINE_MUTEX(einj_mutex);
>>  
>> +/*
>> + * Functions called from dependent modules should check this variable
>> + * before using any EINJ functionality.
>> + */
> 
> This reads slightly odd to me, is this clearer?
> 
> "Exported APIs use this flag to exit early if einj_probe() failed."
> 

That is clearer, I'll change it.

>> +static bool einj_initialized;
> 
> This can be marked __ro_after_init to make it clear that it is static
> for the lifetime of the module.
> 

Will do.

>> +
>>  static void *einj_param;
>>  
>>  static void einj_exec_ctx_init(struct apei_exec_context *ctx)
>> @@ -684,7 +691,7 @@ static int einj_check_table(struct acpi_table_einj *einj_tab)
>>  	return 0;
>>  }
>>  
>> -static int __init einj_init(void)
>> +static int einj_probe(struct platform_device *pdev)
> 
> This can remain __init since nothing will call this function port
> init().
> 
>>  {
>>  	int rc;
>>  	acpi_status status;
>> @@ -782,7 +789,7 @@ static int __init einj_init(void)
>>  	return rc;
>>  }
>>  
>> -static void __exit einj_exit(void)
>> +static void einj_remove(struct platform_device *pdev)
> 
> Similarly this can remain __exit.
>  
>>  {
>>  	struct apei_exec_context ctx;
>>  
>> @@ -801,6 +808,38 @@ static void __exit einj_exit(void)
>>  	acpi_put_table((struct acpi_table_header *)einj_tab);
>>  }
>>  
>> +static struct platform_device *einj_dev;
>> +struct platform_driver einj_driver = {
>> +	.remove_new = einj_remove,
>> +	.driver = {
>> +		.name = "einj",
> 
> Perhaps call this acpi-einj just to preserve the namespace in case a
> cross-platform generic "einj" comes along.
> 

Gotcha, I'll make that change.

>> +	},
>> +};
>> +
>> +static int __init einj_init(void)
>> +{
>> +	struct platform_device_info einj_dev_info = {
>> +		.name = "einj",
> 
> Ditt "acpi-einj"
> 
>> +		.id = -1,
>> +	};
>> +
>> +	einj_dev = platform_device_register_full(&einj_dev_info);
> 
> Just return early here if this failed.
> 

Will do.

>> +	einj_initialized = !platform_driver_probe(&einj_driver, einj_probe);
> 
> Nit, but since platform_driver_probe() does not return bool, I would
> prefer this to be more explicit:
> 
> 	err = platform_driver_probe();
> 	einj_initialized = err == 0;
> 

I agree. I was trying to not have to use an extra variable for only one line
but
	einj_initialized = platform_driver_probe() == 0;

went over the column limit :/.

> I think it is ok for the platform-device to stick around if einj_probe()
> failures as userspace can see that the module is loaded but driver-init
> failed.
> 

My reasoning for this was since this is really a dummy device I
didn't want to pollute the platform device in the case the driver failed, but I
see the reasoning here and agree with you.

> Similarly it's probably also ok to fail the module load if
> platform_device_register_full() fails since something deeper is wrong
> with the system if it is starting to fail something so basic.
> 

Alright, will do.

>> +	if (!(einj_initialized || IS_ERR_OR_NULL(einj_dev)))
>> +		platform_device_del(einj_dev);
>> +
>> +	return 0;
>> +}
>> +
>> +static void __exit einj_exit(void)
>> +{
>> +	if (einj_initialized) {
>> +		platform_driver_unregister(&einj_driver);
>> +		platform_device_del(einj_dev);
> 
> Per above, this device_del can move outside the conditional.