[PATCH 4/4] drm/ttm: Switch to using the new intersect callback

[Christian König]

Am 20.07.22 um 09:36 schrieb Arunpravin Paneer Selvam:
> Use new intersect callback instead of having a generic
> placement range verification.
>
> Signed-off-by: Christian König <christian.koenig at amd.com>
> Signed-off-by: Arunpravin Paneer Selvam <Arunpravin.PaneerSelvam at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 45 +++++++------------------
>   drivers/gpu/drm/ttm/ttm_bo.c            |  9 +++--
>   drivers/gpu/drm/ttm/ttm_resource.c      |  5 +--
>   3 files changed, 20 insertions(+), 39 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> index 170935c294f5..7d25a10395c0 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> @@ -1328,11 +1328,12 @@ uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
>   static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
>   					    const struct ttm_place *place)
>   {
> -	unsigned long num_pages = bo->resource->num_pages;
>   	struct dma_resv_iter resv_cursor;
> -	struct amdgpu_res_cursor cursor;
>   	struct dma_fence *f;
>   
> +	if (!amdgpu_bo_is_amdgpu_bo(bo))
> +		return ttm_bo_eviction_valuable(bo, place);
> +
>   	/* Swapout? */
>   	if (bo->resource->mem_type == TTM_PL_SYSTEM)
>   		return true;
> @@ -1351,40 +1352,20 @@ static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
>   			return false;
>   	}
>   
> -	switch (bo->resource->mem_type) {
> -	case AMDGPU_PL_PREEMPT:
> -		/* Preemptible BOs don't own system resources managed by the
> -		 * driver (pages, VRAM, GART space). They point to resources
> -		 * owned by someone else (e.g. pageable memory in user mode
> -		 * or a DMABuf). They are used in a preemptible context so we
> -		 * can guarantee no deadlocks and good QoS in case of MMU
> -		 * notifiers or DMABuf move notifiers from the resource owner.
> -		 */
> +	/* Preemptible BOs don't own system resources managed by the
> +	 * driver (pages, VRAM, GART space). They point to resources
> +	 * owned by someone else (e.g. pageable memory in user mode
> +	 * or a DMABuf). They are used in a preemptible context so we
> +	 * can guarantee no deadlocks and good QoS in case of MMU
> +	 * notifiers or DMABuf move notifiers from the resource owner.
> +	 */
> +	if (bo->resource->mem_type == AMDGPU_PL_PREEMPT)
>   		return false;
> -	case TTM_PL_TT:
> -		if (amdgpu_bo_is_amdgpu_bo(bo) &&
> -		    amdgpu_bo_encrypted(ttm_to_amdgpu_bo(bo)))
> -			return false;
> -		return true;
>   
> -	case TTM_PL_VRAM:
> -		/* Check each drm MM node individually */
> -		amdgpu_res_first(bo->resource, 0, (u64)num_pages << PAGE_SHIFT,
> -				 &cursor);
> -		while (cursor.remaining) {
> -			if (place->fpfn < PFN_DOWN(cursor.start + cursor.size)
> -			    && !(place->lpfn &&
> -				 place->lpfn <= PFN_DOWN(cursor.start)))
> -				return true;
> -
> -			amdgpu_res_next(&cursor, cursor.size);
> -		}
> +	if (bo->resource->mem_type == TTM_PL_TT &&
> +	    amdgpu_bo_encrypted(ttm_to_amdgpu_bo(bo)))
>   		return false;
>   
> -	default:
> -		break;
> -	}
> -
>   	return ttm_bo_eviction_valuable(bo, place);
>   }
>   
> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
> index c1bd006a5525..03409409e43e 100644
> --- a/drivers/gpu/drm/ttm/ttm_bo.c
> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
> @@ -518,6 +518,9 @@ static int ttm_bo_evict(struct ttm_buffer_object *bo,
>   bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
>   			      const struct ttm_place *place)
>   {
> +	struct ttm_resource *res = bo->resource;
> +	struct ttm_device *bdev = bo->bdev;
> +
>   	dma_resv_assert_held(bo->base.resv);
>   	if (bo->resource->mem_type == TTM_PL_SYSTEM)
>   		return true;
> @@ -525,11 +528,7 @@ bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
>   	/* Don't evict this BO if it's outside of the
>   	 * requested placement range
>   	 */
> -	if (place->fpfn >= (bo->resource->start + bo->resource->num_pages) ||
> -	    (place->lpfn && place->lpfn <= bo->resource->start))
> -		return false;
> -
> -	return true;
> +	return ttm_resource_intersect(bdev, res, place, bo->base.size);
>   }
>   EXPORT_SYMBOL(ttm_bo_eviction_valuable);
>   
> diff --git a/drivers/gpu/drm/ttm/ttm_resource.c b/drivers/gpu/drm/ttm/ttm_resource.c
> index 84c21f92b422..ff3e9058943c 100644
> --- a/drivers/gpu/drm/ttm/ttm_resource.c
> +++ b/drivers/gpu/drm/ttm/ttm_resource.c
> @@ -285,6 +285,8 @@ static bool ttm_resource_places_compat(struct ttm_resource *res,
>   				       const struct ttm_place *places,
>   				       unsigned num_placement)
>   {
> +	struct ttm_buffer_object *bo = res->bo;
> +	struct ttm_device *bdev = bo->bdev;
>   	unsigned i;
>   
>   	if (res->placement & TTM_PL_FLAG_TEMPORARY)
> @@ -293,8 +295,7 @@ static bool ttm_resource_places_compat(struct ttm_resource *res,
>   	for (i = 0; i < num_placement; i++) {
>   		const struct ttm_place *heap = &places[i];
>   
> -		if (res->start < heap->fpfn || (heap->lpfn &&
> -		    (res->start + res->num_pages) > heap->lpfn))
> +		if (!ttm_resource_intersect(bdev, res, heap, bo->base.size))

Thinking more about it that probably won't work like this.

Intersection doesn't mean the placement is compatible (e.g. it can be 
that we only partial intersect).

We need to add a separate callback for that I think.

Regards,
Christian.

>   			continue;
>   
>   		if ((res->mem_type == heap->mem_type) &&