[Intel-gfx] [PATCH v6 7/7] drm/i915: add a sysfs entry to let users set sseu configs
Lionel Landwerlin
lionel.g.landwerlin at intel.com
Fri May 25 15:42:02 UTC 2018
On 24/05/18 11:39, Tvrtko Ursulin wrote:
>>>> --- a/drivers/gpu/drm/i915/i915_gem_context.c
>>>> +++ b/drivers/gpu/drm/i915/i915_gem_context.c
>>>> @@ -981,7 +981,8 @@ int i915_gem_context_setparam_ioctl(struct
>>>> drm_device *dev, void *data,
>>>> break;
>>>> }
>>>> - if (!capable(CAP_SYS_ADMIN)) {
>>>> + if (!dev_priv->contexts.allow_sseu &&
>>>> + !capable(CAP_SYS_ADMIN)) {
>>>
>>> So the thing I mentioned in the previous patch. My thinking was not
>>> to fail it if sysfs toggle is disabled, but just don't do dynamic
>>> switching. That way the software stack works in all cases but user
>>> has the option of tuning his system.
>>>
>>> I mean it can still be done but in userspace. Set param fails,
>>> userspace goes for a fall back.
>>>
>>> Only difference is I guess whether it is useful to allow switching
>>> at runtime. I am imagining running some 3d and media, and then
>>> toggling the knob to see what happens to power and performance on
>>> the fly. But maybe that is not so interesting.
>>>
>>> Along the same lines I was thinking CAP_SYS_ADMIN limitation could
>>> be dropped.
>>>
>>> Both points are for a wider discussion I guess.
>>
>> Okay, let's get Dmitry input on this.
>
> To expand, my thinking is to let media stack configure its contexts
> for fewer slices, but let the user/sysadmin decide whether 3d/compute
> performance is more important, or media.
>
> Downside is that with this approach you would have to re-configure all
> contexts when sysfs toggle is changed from zero to one.
Right, I've updated this series locally with Tvrkto's latest comments on v7.
The issue discussed above is the last remaining detail on which some
kind of decision needs to be taken.
Ccing Joonas who might have an input.
Cheers,
-
Lionel
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/intel-gfx/attachments/20180525/55128e27/attachment-0001.html>
More information about the Intel-gfx
mailing list