Re: [PATCH V1] accel/amdxdna: Add ioctl DRM_IOCTL_AMDXDNA_GET_ARRAY

[Mario Limonciello <mario.limonciello@amd.com>]

On 8/27/2025 11:41 AM, Lizhi Hou wrote:
> 
> On 8/26/25 17:31, Mario Limonciello wrote:
>> On 8/26/2025 1:10 PM, Lizhi Hou wrote:
>>>
>>> On 8/26/25 10:58, Mario Limonciello wrote:
>>>> On 8/26/2025 12:55 PM, Lizhi Hou wrote:
>>>>>
>>>>> On 8/26/25 10:18, Mario Limonciello wrote:
>>>>>> On 8/25/2025 11:48 PM, Lizhi Hou wrote:
>>>>>>>
>>>>>>> On 8/25/25 14:28, Mario Limonciello wrote:
>>>>>>>> On 8/22/2025 12:23 PM, Lizhi Hou wrote:
>>>>>>>>> Add interface for applications to get information array. The 
>>>>>>>>> application
>>>>>>>>> provides a buffer pointer along with information type, maximum 
>>>>>>>>> number of
>>>>>>>>> entries and maximum size of each entry. The buffer may also 
>>>>>>>>> contain match
>>>>>>>>> conditions based on the information type. After the ioctl 
>>>>>>>>> completes, the
>>>>>>>>> actual number of entries and entry size are returned.
>>>>>>>>>
>>>>>>>>> Signed-off-by: Lizhi Hou <lizhi.hou@amd.com>
>>>>>>>>
>>>>>>>> How does userspace discover whether or not the new IOCTL call is 
>>>>>>>> supported?  Just a test call?
>>>>>>> The kernel header version will be used to determine whether the 
>>>>>>> application which uses new IOCTL will be compiled or not.
>>>>>>>
>>>>>>
>>>>>> But it's not actually an application compile time decision, it's a 
>>>>>> runtime decision.  IE I can compile an application with the 
>>>>>> headers on kernel 6.18 that has this, but if I try to run it on 
>>>>>> 6.15 it's going to barf.
>>>>>>
>>>>>> To some extent that comes with the territory, but I'm wondering if 
>>>>>> a better solution going forward would be for there to be a 
>>>>>> dedicated version command that you bump.
>>>>>
>>>>> For in-tree driver, I did not aware a common way for this other 
>>>>> than checking the kernel version.
>>>>>
>>>>> And here is qaic patch of adding a new IOCTL.
>>>>>
>>>>> https://github.com/torvalds/linux/ 
>>>>> commit/217b812364d360e1933d8485f063400e5dda7d66
>>>>>
>>>>>
>>>>> I know there is major, minor, patchlevel in struct drm_driver. And 
>>>>> I think that is not required for in-tree driver.
>>>>>
>>>>> Please let me know if I missed anything.
>>>>>
>>>>> Thanks,
>>>>
>>>> Right; so bump up one of those so that userspace can check it. Maybe 
>>>> "minor"?
>>>
>>> I meant for in-tree driver, is it good enough for userspace to just 
>>> check kernel version?  E.g. The drm driver versions are not used by 
>>> ivpu or qaic.
>>>
>>
>> Just because they don't doesn't mean you shouldn't.
> Ok. :) It does not sound amdxdna specific. Just wondering how the other 
> driver/application under accel subsystem handle this.
>>
>> Take a look at what amdgpu does for user queues earlier this year for 
>> example: 100b6010d7540e
>>
>> This means that a userspace application can look for that minor bump 
>> or newer to know the ioctl supports user queues.
> 
> As in-tree driver is part of kernel, the userspace application can check 
> kernel version to determine whether a feature is supported or not. Could 
> you share the idea why would user application to check drm driver 
> version for this?
> 
> And amdxdna driver is new added driver which never bumped drm major/ 
> minor before. Thus there is not any application use drm versions. Maybe 
> using kernel version directly is good enough in this case?
> 
> I am fine to bump minor if it provides better support to user applications.
> 
> 
> Thanks,
> 
> Lizhi
> 

If you're running mainline kernel I totally agree with you that you can 
make a runtime call based upon major/minor kernel version.  To me the 
problem ends up being cases that distros do a backport of a driver or 
subsystem that this falls apart.

For example RHEL and CentOS stream both do this, and then such 
comparisons can no longer be made accurately.