Exposing GPU information for userspace processes

[Rhyland Klein]

On 7/10/2013 2:09 AM, Daniel Vetter wrote:
> On Wed, Jul 10, 2013 at 04:01:49PM +1000, Dave Airlie wrote:
>> On Wed, Jul 10, 2013 at 7:45 AM, Rhyland Klein <rklein at nvidia.com> wrote:
>>> We are currently looking into exporting some information from the linux
>>> kernel to userspace about the GPU. This information is specific per
>>> board as the data can vary depending on the fuses burnt on the board.
>>>
>>> Right now we are leaning towards adding sysfs properties to our existing
>>> nodes to share this information.
>>>
>>> Before we write up the patches and update our userspace apps to check
>>> the new locations, we wanted to make sure there wasn't a more
>>> accepted/proper way to export information already.
>>>
>>> Right now for instance we want to export a property to say how many
>>> pixel pipes there are. This could then be exported in something like:
>>>
>>> /sys/bus/platform/drivers/gr3d/54180000.gr3d/num_pixel_pipes
>>>
>>> Is this the best approach to take right now or is there another way we
>>> should export this and similar HW specific information?
>>
>> All the x86 GPU drivers use gpu specific get param ioctls for this so far.
>>
>> e.g. drivers/gpu/drm/radeon/radeon_kms.c:radeon_info_ioctl
> 
> In drm/i915 we also expose a bunch of thing to sysfs, but everything that
> the userspace drivers need for operation is exposed through the getparam
> ioctl. We add sysfs interfaces on top if there's a legit reason for the
> user to script changes (e.g. gpu error state collection or tuning the
> frequency scaling and related read-only statistics).
> -Daniel
> 

Thank you both for your helpful responses. I was considering an ioctl as
another option, and it seems that is likely the best approach to take
for this type of information.

Thanks,

-rhyland

-- 
nvpublic