drm/amdgpu: rework dma_resv handling v3

Drop the workaround and instead implement a better solution.

Basically we are now chaining all submissions using a dma_fence_chain
container and adding them as exclusive fence to the dma_resv object.

This way other drivers can still sync to the single exclusive fence
while amdgpu only sync to fences from different processes.

v3: add the shared fence first before the exclusive one

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20210614174536.5188-2-christian.koenig@amd.com

1 files changed, 51 insertions, 11 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
index 9ce649a1a8d3..25655414e9c0 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
@@ -572,6 +572,20 @@ static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
 		goto out;
 	}
 
+	amdgpu_bo_list_for_each_entry(e, p->bo_list) {
+		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
+
+		e->bo_va = amdgpu_vm_bo_find(vm, bo);
+
+		if (bo->tbo.base.dma_buf && !amdgpu_bo_explicit_sync(bo)) {
+			e->chain = dma_fence_chain_alloc();
+			if (!e->chain) {
+				r = -ENOMEM;
+				goto error_validate;
+			}
+		}
+	}
+
 	amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
 					  &p->bytes_moved_vis_threshold);
 	p->bytes_moved = 0;
@@ -599,15 +613,6 @@ static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
 	gws = p->bo_list->gws_obj;
 	oa = p->bo_list->oa_obj;
 
-	amdgpu_bo_list_for_each_entry(e, p->bo_list) {
-		struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
-
-		/* Make sure we use the exclusive slot for shared BOs */
-		if (bo->prime_shared_count)
-			e->tv.num_shared = 0;
-		e->bo_va = amdgpu_vm_bo_find(vm, bo);
-	}
-
 	if (gds) {
 		p->job->gds_base = amdgpu_bo_gpu_offset(gds) >> PAGE_SHIFT;
 		p->job->gds_size = amdgpu_bo_size(gds) >> PAGE_SHIFT;
@@ -629,8 +634,13 @@ static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
 	}
 
 error_validate:
-	if (r)
+	if (r) {
+		amdgpu_bo_list_for_each_entry(e, p->bo_list) {
+			dma_fence_chain_free(e->chain);
+			e->chain = NULL;
+		}
 		ttm_eu_backoff_reservation(&p->ticket, &p->validated);
+	}
 out:
 	return r;
 }
@@ -670,9 +680,17 @@ static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error,
 {
 	unsigned i;
 
-	if (error && backoff)
+	if (error && backoff) {
+		struct amdgpu_bo_list_entry *e;
+
+		amdgpu_bo_list_for_each_entry(e, parser->bo_list) {
+			dma_fence_chain_free(e->chain);
+			e->chain = NULL;
+		}
+
 		ttm_eu_backoff_reservation(&parser->ticket,
 					   &parser->validated);
+	}
 
 	for (i = 0; i < parser->num_post_deps; i++) {
 		drm_syncobj_put(parser->post_deps[i].syncobj);
@@ -1245,6 +1263,28 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
 
 	amdgpu_vm_move_to_lru_tail(p->adev, &fpriv->vm);
 
+	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;
+
+		if (!chain)
+			continue;
+
+		/*
+		 * Work around dma_resv shortcommings by wrapping up the
+		 * submission in a dma_fence_chain and add it as exclusive
+		 * fence, but first add the submission as shared fence to make
+		 * sure that shared fences never signal before the exclusive
+		 * one.
+		 */
+		dma_fence_chain_init(chain, dma_resv_excl_fence(resv),
+				     dma_fence_get(p->fence), 1);
+
+		dma_resv_add_shared_fence(resv, p->fence);
+		rcu_assign_pointer(resv->fence_excl, &chain->base);
+		e->chain = NULL;
+	}
+
 	ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
 	mutex_unlock(&p->adev->notifier_lock);