[Intel-gfx] [PATCH v3 12/14] drm/i915/guc: Preemption! With GuC

[Chris Wilson]

Quoting Michał Winiarski (2017-10-19 19:36:17)
> @@ -686,10 +802,23 @@ static void i915_guc_irq_handler(unsigned long data)
>         struct intel_engine_cs * const engine = (struct intel_engine_cs *)data;
>         struct intel_engine_execlists * const execlists = &engine->execlists;
>         struct execlist_port *port = execlists->port;
> -       const struct execlist_port * const last_port =
> -               &execlists->port[execlists->port_mask];
>         struct drm_i915_gem_request *rq;
>  
> +       if (READ_ONCE(execlists->preempt) &&
> +           intel_read_status_page(engine, I915_GEM_HWS_PREEMPT_INDEX) ==
> +           GUC_PREEMPT_FINISHED) {
> +               execlists_cancel_port_requests(&engine->execlists);
> +
> +               spin_lock_irq(&engine->timeline->lock);
> +               execlists_unwind_incomplete_requests(execlists);
> +               spin_unlock_irq(&engine->timeline->lock);
> +
> +               wait_for_guc_preempt_report(engine);
> +
> +               WRITE_ONCE(execlists->preempt, false);
> +               intel_write_status_page(engine, I915_GEM_HWS_PREEMPT_INDEX, 0);
> +       }
> +
>         rq = port_request(&port[0]);
>         while (rq && i915_gem_request_completed(rq)) {
>                 trace_i915_gem_request_out(rq);

I don't think we have the ordering right here for
cancel_port_requests(). It should only be called for what are
effectively incomplete requests, or else the
INTEL_CONTEXT_SCHEDULE_PREEMPTED is a misnomer. It doesn't look like it
makes any great difference in the over scheme of things, but I'm tempted
to put a GEM_BUG_ON(i915_gem_request_completed(rq)) there.
-Chris