Re: [PATCH v15 2/3] virt: Add TDX guest driver

[Sathyanarayanan Kuppuswamy]

On 10/21/22 11:05 PM, Greg Kroah-Hartman wrote:
> On Fri, Oct 21, 2022 at 04:51:34PM -0700, Sathyanarayanan Kuppuswamy wrote:
>> Hi Greg,
>>
>> On 10/20/22 9:39 PM, Greg Kroah-Hartman wrote:
>>>>>> +#ifdef MODULE
>>>>>> +static const struct x86_cpu_id tdx_guest_ids[] = {
>>>>>> +	X86_MATCH_FEATURE(X86_FEATURE_TDX_GUEST, NULL),
>>>>>> +	{}
>>>>>> +};
>>>>>> +MODULE_DEVICE_TABLE(x86cpu, tdx_guest_ids);
>>>>>> +#endif
>>>>> Why the #ifdef?  Should not be needed, right?
>>>> I have added it to fix the following warning reported by 0-day.
>>>>
>>>> https://lore.kernel.org/lkml/202209211607.tCtTWKbV-lkp@xxxxxxxxx/
>>>>
>>>> It is related to nullifying the MODULE_DEVICE_TABLE in #ifndef MODULE
>>>> case in linux/module.h.
>>> Then fix it properly, by correctly using that structure no matter what.
>>> You don't do that here...
>>
>> I think we can use __maybe_unused attribute to fix this warning like
>> mentioned below. Are you fine with it?
>>
>> --- a/drivers/virt/coco/tdx-guest/tdx-guest.c
>> +++ b/drivers/virt/coco/tdx-guest/tdx-guest.c
>> @@ -118,13 +118,11 @@ static void __exit tdx_guest_exit(void)
>>  }
>>  module_exit(tdx_guest_exit);
>>  
>> -#ifdef MODULE
>> -static const struct x86_cpu_id tdx_guest_ids[] = {
>> +static const struct x86_cpu_id __maybe_unused tdx_guest_ids[] = {
>>         X86_MATCH_FEATURE(X86_FEATURE_TDX_GUEST, NULL),
>>         {}
>>  };
>>  MODULE_DEVICE_TABLE(x86cpu, tdx_guest_ids);
>> -#endif
>>
>> Solution 2:
>> -----------
>>
>> We can also modify the code to use this structure in all cases like
>> below. But it requires me to use slower x86_match_cpu() in place of 
>> cpu_feature_enabled() which I think is unnecessary.
>>
>> --- a/drivers/virt/coco/tdx-guest/tdx-guest.c
>> +++ b/drivers/virt/coco/tdx-guest/tdx-guest.c
>> @@ -103,9 +103,15 @@ static struct miscdevice tdx_misc_dev = {
>>         .fops = &tdx_guest_fops,
>>  };
>>  
>> +static const struct x86_cpu_id tdx_guest_ids[] = {
>> +       X86_MATCH_FEATURE(X86_FEATURE_TDX_GUEST, NULL),
>> +       {}
>> +};
>> +MODULE_DEVICE_TABLE(x86cpu, tdx_guest_ids);
>> +
>>  static int __init tdx_guest_init(void)
>>  {
>> -       if (!cpu_feature_enabled(X86_FEATURE_TDX_GUEST))
>> +       if (!x86_match_cpu(tdx_guest_ids))
> 
> Please use this as it's what all other users of the x86cpu module device

Ok. I will use it.

> table code uses, right?

Not all, but most of them use the above model. 

Following two drivers seems to use __maybe_unused method.

./drivers/cpufreq/acpi-cpufreq.c
./drivers/char/hw_random/via-rng.c

and following two drivers uses #ifdef MODULE method.

./arch/x86/kvm/vmx/vmx.c
./arch/x86/kvm/svm/svm.c


> 
> And what is the "speed" difference here?  Is is measurable and where

> does it matter?

Speed difference does not really matter in init code. So I am fine
with using this approach.

> 
> thanks,
> 
> greg k-h

-- 
Sathyanarayanan Kuppuswamy
Linux Kernel Developer