[Intel-gfx] [PATCH] drm/i915: Always load guc by default.

[Chris Wilson]

On Fri, Dec 02, 2016 at 06:11:12PM +0000, Srivatsa, Anusha wrote:
> 
> 
> >-----Original Message-----
> >From: Chris Wilson [mailto:chris at chris-wilson.co.uk]
> >Sent: Thursday, November 24, 2016 12:41 AM
> >To: Tvrtko Ursulin <tvrtko.ursulin at linux.intel.com>
> >Cc: Srivatsa, Anusha <anusha.srivatsa at intel.com>; intel-
> >gfx at lists.freedesktop.org; Mcgee, Jeff <jeff.mcgee at intel.com>
> >Subject: Re: [Intel-gfx] [PATCH] drm/i915: Always load guc by default.
> >
> >On Thu, Nov 24, 2016 at 08:31:22AM +0000, Tvrtko Ursulin wrote:
> >>
> >> On 24/11/2016 08:21, Chris Wilson wrote:
> >> >On Thu, Nov 24, 2016 at 08:15:31AM +0000, Tvrtko Ursulin wrote:
> >> >>
> >> >>On 24/11/2016 07:13, Chris Wilson wrote:
> >> >>>On Wed, Nov 23, 2016 at 04:52:38PM -0800, Anusha Srivatsa wrote:
> >> >>>>Remove the enable_guc_loading parameter. Load the GuC on plaforms
> >> >>>>that have GuC. All issues we found so far are related to GuC
> >> >>>>features like the command submission, but no bug is related to the
> >> >>>>guc loading itself.
> >> >>>>
> >> >>>>This addresses the case when we need GuC loaded even with no GuC
> >> >>>>feature in use, like - GuC  authenticating HuC loading.
> >> >>>
> >> >>>Why not just load the firmware if it may be used?
> >> >>
> >> >>It was discussed briefly in the other thread, but I suppose as soon
> >> >>as the HuC patches go in that would be always so it may not be that
> >> >>useful.
> >> >>
> >> >>Unless there is a reason to add a HuC enable/disable parameter in
> >> >>general. I have no idea on that one.
> >> >
> >> >In hindsight, we should have had i915.enable_dmc to easily protect
> >> >users against failures. History says we will regret enabling a new
> >> >piece of hw/fw without a feature option.
> >>
> >> So..
> >>
> >>  1. Add i915.enable_huc, default to enabled  2. Unexport
> >> i915.enable_guc_loading  3. Gate enable_guc_loading by i915.enable_huc
> >> and i915.enable_guc_submission
> >
> >Aye, that would be my preference.
> >-Chris
> 
> So, basically control the guc loads depending on huc_enable or guc_submission parameter. If either or both are enabled then load guc.
> No need to load guc if a platform has one? 
> Any thought on having 0,-1,1 and 2 values for submission? 1 is load if guc is available (but do not error out if not present) and 2 is load and fail if not present. I am thinking what if we need that differentiation.....

I have no idea what value 2 has for anybody. Userspace can check whether
or not guc submission is enabled and fail validation without resorting to
the kernel wedging the machine.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre