Re: [PATCH 11/15] drm/xe: Convert xe_dma_buf.c for exhaustive eviction

[Matthew Brost <matthew.brost@intel.com>]

On Wed, Aug 13, 2025 at 12:51:17PM +0200, Thomas Hellström wrote:
> Convert dma-buf migration to XE_PL_TT and dma-buf import to
> support exhaustive eviction, using xe_validation_guard().
> It seems unlikely that the import would result in an -ENOMEM,
> but convert import anyway for completeness.
> 
> The dma-buf map_attachment() functionality unfortunately doesn't
> support passing a drm_exec, which means that foreign devices
> validating a dma-buf that we exported will not, unless they are
> xeKMD devices, participate in the exhaustive eviction scheme.
> 
> Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
> ---
>  drivers/gpu/drm/xe/xe_dma_buf.c | 59 +++++++++++++++++++++++----------
>  1 file changed, 42 insertions(+), 17 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_dma_buf.c b/drivers/gpu/drm/xe/xe_dma_buf.c
> index 78a827d4e726..56df1d84df21 100644
> --- a/drivers/gpu/drm/xe/xe_dma_buf.c
> +++ b/drivers/gpu/drm/xe/xe_dma_buf.c
> @@ -163,16 +163,27 @@ static int xe_dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
>  	struct xe_bo *bo = gem_to_xe_bo(obj);
>  	bool reads =  (direction == DMA_BIDIRECTIONAL ||
>  		       direction == DMA_FROM_DEVICE);
> -	struct drm_exec *exec = XE_VALIDATION_UNIMPLEMENTED;
> +	struct xe_validation_ctx ctx;
> +	struct drm_exec exec;
> +	int ret = 0;
>  
>  	if (!reads)
>  		return 0;
>  
>  	/* Can we do interruptible lock here? */
> -	xe_bo_lock(bo, false);
> -	(void)xe_bo_migrate(bo, XE_PL_TT, exec);
> -	xe_bo_unlock(bo);
> -
> +	xe_validation_guard(&ctx, &xe_bo_device(bo)->val, &exec, 0, ret, false) {
> +		ret = drm_exec_lock_obj(&exec, &bo->ttm.base);
> +		drm_exec_retry_on_contention(&exec);
> +		if (ret)
> +			goto out;
> +
> +		ret = xe_bo_migrate(bo, XE_PL_TT, &exec);
> +		drm_exec_retry_on_contention(&exec);
> +		xe_validation_retry_on_oom(&ctx, &ret);
> +	}
> +out:
> +	/* If we failed, cpu-access takes place in current placement. */
> +	(void)ret;

Do you need the above line of code? I don't see this often in kernel code.

Nit aside, patch LGTM.

Matt

>  	return 0;
>  }
>  
> @@ -211,24 +222,38 @@ xe_dma_buf_init_obj(struct drm_device *dev, struct xe_bo *storage,
>  {
>  	struct dma_resv *resv = dma_buf->resv;
>  	struct xe_device *xe = to_xe_device(dev);
> -	struct drm_exec *exec = XE_VALIDATION_UNIMPLEMENTED;
> +	struct xe_validation_ctx ctx;
> +	struct drm_gem_object *dummy_obj;
> +	struct drm_exec exec;
>  	struct xe_bo *bo;
> -	int ret;
> -
> -	dma_resv_lock(resv, NULL);
> -	bo = ___xe_bo_create_locked(xe, storage, NULL, resv, NULL, dma_buf->size,
> -				    0, /* Will require 1way or 2way for vm_bind */
> -				    ttm_bo_type_sg, XE_BO_FLAG_SYSTEM, exec);
> -	if (IS_ERR(bo)) {
> -		ret = PTR_ERR(bo);
> -		goto error;
> +	int ret = 0;
> +
> +	dummy_obj = drm_gpuvm_resv_object_alloc(&xe->drm);
> +	if (!dummy_obj)
> +		return ERR_PTR(-ENOMEM);
> +
> +	dummy_obj->resv = resv;
> +	xe_validation_guard(&ctx, &xe->val, &exec, 0, ret, false) {
> +		ret = drm_exec_lock_obj(&exec, dummy_obj);
> +		drm_exec_retry_on_contention(&exec);
> +		if (ret)
> +			goto error;
> +
> +		bo = ___xe_bo_create_locked(xe, storage, NULL, resv, NULL, dma_buf->size,
> +					    0, /* Will require 1way or 2way for vm_bind */
> +					    ttm_bo_type_sg, XE_BO_FLAG_SYSTEM, &exec);
> +		drm_exec_retry_on_contention(&exec);
> +		if (IS_ERR(bo)) {
> +			ret = PTR_ERR(bo);
> +			xe_validation_retry_on_oom(&ctx, &ret);
> +			goto error;
> +		}
>  	}
> -	dma_resv_unlock(resv);
> +	drm_gem_object_put(dummy_obj);
>  
>  	return &bo->ttm.base;
>  
>  error:
> -	dma_resv_unlock(resv);
>  	return ERR_PTR(ret);
>  }
>  
> -- 
> 2.50.1
>