drm/amdgpu: use dma_resv_for_each_fence for CS workaround v2

Get the write fence using dma_resv_for_each_fence instead of accessing
it manually.

v2: add TODO comment
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: amd-gfx@lists.freedesktop.org
Link: https://patchwork.freedesktop.org/patch/msgid/20220321135856.1331-9-christian.koenig@amd.com

drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c

@@ -1267,18 +1267,23 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
 	amdgpu_bo_list_for_each_entry(e, p->bo_list) {
 		struct dma_resv *resv = e->tv.bo->base.resv;
 		struct dma_fence_chain *chain = e->chain;
+		struct dma_resv_iter cursor;
+		struct dma_fence *fence;
 
 		if (!chain)
 			continue;
 
 		/*
-		 * Work around dma_resv shortcommings by wrapping up the
-		 * submission in a dma_fence_chain and add it as exclusive
+		 * Temporary workaround dma_resv shortcommings by wrapping up
+		 * the submission in a dma_fence_chain and add it as exclusive
 		 * fence.
+		 *
+		 * TODO: Remove together with dma_resv rework.
 		 */
-		dma_fence_chain_init(chain, dma_resv_excl_fence(resv),
-				     dma_fence_get(p->fence), 1);
-
+		dma_resv_for_each_fence(&cursor, resv, false, fence) {
+			break;
+		}
+		dma_fence_chain_init(chain, fence, dma_fence_get(p->fence), 1);
 		rcu_assign_pointer(resv->fence_excl, &chain->base);
 		e->chain = NULL;
 	}