[Intel-gfx] [PATCH] drm/i915: Finish i915_gem_wait_for_idle() on the global execution timeline

[Chris Wilson]

Quoting Chris Wilson (2017-12-11 21:42:22)
> In order to be sure that we catch the most recent request to be
> executed, we need to finish on the global execution timeline. This
> prevents a race where the request is being transferred between timelines
> behind our back, jumping it ahead in the list onto the global execution
> timeline that we have already processed causing us not to wait on that
> request. Document this requirement with an assertion to protect
> ourselves against future changes.
> 
> References: 73cb97010d4f ("drm/i915: Combine seqno + tracking into a global timeline struct")
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> Cc: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
> ---
>  drivers/gpu/drm/i915/i915_gem.c | 4 ++++
>  1 file changed, 4 insertions(+)
> 
> diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> index 9e957b213fdb..40db2258eb07 100644
> --- a/drivers/gpu/drm/i915/i915_gem.c
> +++ b/drivers/gpu/drm/i915/i915_gem.c
> @@ -3566,6 +3566,10 @@ int i915_gem_wait_for_idle(struct drm_i915_private *i915, unsigned int flags)
>  
>                 lockdep_assert_held(&i915->drm.struct_mutex);
>  
> +               /* Must finish by waiting on the global execution_timeline */
> +               GEM_BUG_ON(list_last_entry(&i915->gt.timelines,
> +                                          typeof(*tl), link) !=
> +                          &i915->gt.global_timeline);

Scratch all that. It's just nice to have, rather than a hard
requirement. We aren't looking at the timeline->requests list, but
timeline->last_request which will not disappear behind our backs.
-Chris