[Intel-xe] [PATCH v2 1/2] drm/xe: Always invalidate GGTT on map during driver load

[Matthew Auld]

On Wed, 5 Apr 2023 at 03:40, Matthew Brost <matthew.brost at intel.com> wrote:
>
> Coalescing GGTT invalidations should work but it creating an issue where
> the driver sporadically fails to load. For now just disable coalescing
> during load.
>
> Should fix:
> https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/232
>
> Signed-off-by: Matthew Brost <matthew.brost at intel.com>

Any idea why we don't currently add any 4K padding for
insert_special_node() users? Don't we need padding for all GGTT
allocations due to overfetch + coalescing?

> ---
>  drivers/gpu/drm/xe/xe_ggtt.c | 9 +++++++--
>  1 file changed, 7 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/xe/xe_ggtt.c b/drivers/gpu/drm/xe/xe_ggtt.c
> index a430d1568890..55d0c2554f7b 100644
> --- a/drivers/gpu/drm/xe/xe_ggtt.c
> +++ b/drivers/gpu/drm/xe/xe_ggtt.c
> @@ -263,6 +263,7 @@ int xe_ggtt_insert_special_node(struct xe_ggtt *ggtt, struct drm_mm_node *node,
>
>  void xe_ggtt_map_bo(struct xe_ggtt *ggtt, struct xe_bo *bo)
>  {
> +       struct xe_gt *gt = ggtt->gt;
>         u64 start = bo->ggtt_node.start;
>         u64 offset, pte;
>
> @@ -278,8 +279,12 @@ void xe_ggtt_map_bo(struct xe_ggtt *ggtt, struct xe_bo *bo)
>                 xe_ggtt_set_pte(ggtt, start + bo->size, pte);
>         }
>
> -       if (ggtt->invalidate) {
> -               xe_ggtt_invalidate(ggtt->gt);
> +       /*
> +        * XXX: Unsure why without invalidation on every map during driver load
> +        * we can fail the driver load.
> +        */
> +       if (ggtt->invalidate || !gt->uc.guc.submission_state.enabled) {
> +               xe_ggtt_invalidate(gt);
>                 ggtt->invalidate = false;
>         }
>  }
> --
> 2.34.1
>