Re: [Intel-gfx] [PATCH v6 5/8] drm/i915/pxp: Add ARB session creation and cleanup

[Rodrigo Vivi <rodrigo.vivi@intel.com>]

On Tue, Mar 28, 2023 at 05:01:36PM +0000, Teres Alexis, Alan Previn wrote:
> On Mon, 2023-03-27 at 17:15 +0100, Tvrtko Ursulin wrote:
> > These two:
> > e6177ec586d1 ("drm/i915/huc: stall media submission until HuC is loaded")
> > b76c14c8fb2a ("drm/i915/huc: better define HuC status getparam possible return values.")
> > They do not help here? It is not possible to use or extend the refined I915_PARAM_HUC_STATUS return values combined with huc load fence for this all to keep working?
> Checking is-huc-loaded won't reflect is-pxp-available (in case fw/fusing isn't allowing it). But the connection between them is hw-internal so i915 asking PXP-fw to attempt a PXP
> session depends on HuC (and the 3 other things i mentioned). However, Tvrtko's point on using fences-or-equivalent is the same thing John Harrison brought up offline with Daniele
> as the proper kernel way to do this type of dependency checking. However, any form of dependency-checking won't improve UMD's experience. We still need to decide if i915-PXP should
> wait-in-kernel or return some-new-spec-error. A useful data point: we are debugging a related issue on actual customer stack. The compositor using mesa is hitting this code path
> very early.. even before gsc-proxy component is loaded and we are trying to figure out why delaying inside intel_pxp_start is not helping (more delays causes the gsc-proxy
> component to also get delayed) but that is a different conversation. I'm only mentioning this because we have a strict requirement to get the desktop login window up within 1-2
> seconds of bootloader->kernel handoff. That said, if use -EAGAIN, I'm not sure if that would work as it would delay the compositor startup beyond the typical end user experience
> unless MESA has a timeout to give up on a cap-testing when seeing repeated -EAGAIN (doubt mesa folks like this?). Perhaps we could just immediately return with a different error
> (instead of current PXP-UAPI spec of -EINVAL or -ENODEV)... perhaps use -ENXIO which apparently is already part of the original pxp code but was never mentioned in UAPI - but we
> return this immediately and document it in UAPI as "hw/fw insfratructure is not yet ready to create pxp arb session, user space can retry but may need delays of up to x-seconds on
> ADl/TGL or y-seconds on MTL, before getting a SUCCESS or one of the other errors).

fair enough. It looks like we need a new get_param! :)