Re: [PATCH 1/2] drm/xe: Skip CCS clear for WB type BOs

["Thomas Hellström" <thomas.hellstrom@linux.intel.com>]

Hi,

On Tue, 2024-08-27 at 17:49 +0200, Nirmoy Das wrote:
> HW treats any access to 1-way or 2-way coherent memory as compression
> disabled memory. So for such BOs there is no need to do CCS clearing.
> 
> Cc: Matthew Auld <matthew.auld@intel.com>
> Cc: Matthew Brost <matthew.brost@intel.com>
> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
> Signed-off-by: Nirmoy Das <nirmoy.das@intel.com>
> ---
>  drivers/gpu/drm/xe/xe_bo.c | 8 +++++++-
>  1 file changed, 7 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c
> index cbe7bf098970..24701272e3af 100644
> --- a/drivers/gpu/drm/xe/xe_bo.c
> +++ b/drivers/gpu/drm/xe/xe_bo.c
> @@ -283,6 +283,7 @@ struct xe_ttm_tt {
>  	struct device *dev;
>  	struct sg_table sgt;
>  	struct sg_table *sg;
> +	bool skip_ccs_clear:1;
>  };
>  
>  static int xe_tt_map_sg(struct ttm_tt *tt)
> @@ -404,6 +405,8 @@ static struct ttm_tt *xe_ttm_tt_create(struct
> ttm_buffer_object *ttm_bo,
>  	if (ttm_bo->type == ttm_bo_type_device && xe-
> >mem.gpu_page_clear_sys)
>  		page_flags |= TTM_TT_FLAG_CLEARED_ON_FREE;
>  
> +	/* compression is not allowed for cached BO so ccs clear can
> be skipped. */
> +	tt->skip_ccs_clear = caching == ttm_cached;

In theory, BOs that are promoted to fb (not created with the SCANOUT
flag) can AFAICT have caching remaining at ttm_cached, yet still sent
to the display engine, reading uninitialized ccs.

Also I think LNL will be the only HW having the "feature" that clean
cache-lines are written back so in the future we might allow 0-coherent
with ttm_cached.

So IMO we need to improve the detection of "skip_ccs_clear" here.
Otherwise, I'm all for the optimizaion.

/Thomas


>  	err = ttm_tt_init(&tt->ttm, &bo->ttm, page_flags, caching,
> extra_pages);
>  	if (err) {
>  		kfree(tt);
> @@ -664,13 +667,16 @@ static int xe_bo_move(struct ttm_buffer_object
> *ttm_bo, bool evict,
>  	struct ttm_resource *old_mem = ttm_bo->resource;
>  	u32 old_mem_type = old_mem ? old_mem->mem_type :
> XE_PL_SYSTEM;
>  	struct ttm_tt *ttm = ttm_bo->ttm;
> +	struct xe_ttm_tt *xe_tt = container_of(ttm_bo->ttm, struct
> xe_ttm_tt,
> +					       ttm);
>  	struct xe_migrate *migrate = NULL;
>  	struct dma_fence *fence;
>  	bool move_lacks_source;
>  	bool tt_has_data;
>  	bool needs_clear;
>  	bool handle_system_ccs = (!IS_DGFX(xe) &&
> xe_bo_needs_ccs_pages(bo) &&
> -				  ttm && ttm_tt_is_populated(ttm)) ?
> true : false;
> +				  ttm && ttm_tt_is_populated(ttm) &&
> +				  !xe_tt->skip_ccs_clear) ? true :
> false;
>  	bool clear_system_pages;
>  	int ret = 0;
>