drm/i915/gvt: Refine shadow batch buffer

1) Use standard i915 GEM object sequence to access the shadow batch buffer.
2) Manage i915 vma life cycle to solve one FIXME.

v2:
- Refine code structure.
- Refine the usage of GEM APIs.
- Add the missing lock/unlock in release_shadow_batch_buffer.

Test on my SKL NuC.
Signed-off-by: Zhi Wang <zhi.a.wang@intel.com>

drivers/gpu/drm/i915/gvt/cmd_parser.c

@@ -1645,11 +1645,10 @@ static int find_bb_size(struct parser_exec_state *s, unsigned long *bb_size)
 
 static int perform_bb_shadow(struct parser_exec_state *s)
 {
-	struct intel_shadow_bb_entry *entry_obj;
 	struct intel_vgpu *vgpu = s->vgpu;
+	struct intel_vgpu_shadow_bb *bb;
 	unsigned long gma = 0;
 	unsigned long bb_size;
-	void *dst = NULL;
 	int ret = 0;
 
 	/* get the start gm address of the batch buffer */
@@ -1657,51 +1656,51 @@ static int perform_bb_shadow(struct parser_exec_state *s)
 	if (gma == INTEL_GVT_INVALID_ADDR)
 		return -EFAULT;
 
-	/* get the size of the batch buffer */
 	ret = find_bb_size(s, &bb_size);
 	if (ret)
 		return ret;
 
-	/* allocate shadow batch buffer */
-	entry_obj = kmalloc(sizeof(*entry_obj), GFP_KERNEL);
-	if (entry_obj == NULL)
+	bb = kzalloc(sizeof(*bb), GFP_KERNEL);
+	if (!bb)
 		return -ENOMEM;
 
-	entry_obj->obj =
-		i915_gem_object_create(s->vgpu->gvt->dev_priv,
-				       roundup(bb_size, PAGE_SIZE));
-	if (IS_ERR(entry_obj->obj)) {
-		ret = PTR_ERR(entry_obj->obj);
-		goto free_entry;
+	bb->obj = i915_gem_object_create(s->vgpu->gvt->dev_priv,
+					 roundup(bb_size, PAGE_SIZE));
+	if (IS_ERR(bb->obj)) {
+		ret = PTR_ERR(bb->obj);
+		goto err_free_bb;
 	}
-	entry_obj->len = bb_size;
-	INIT_LIST_HEAD(&entry_obj->list);
 
-	dst = i915_gem_object_pin_map(entry_obj->obj, I915_MAP_WB);
-	if (IS_ERR(dst)) {
-		ret = PTR_ERR(dst);
-		goto put_obj;
-	}
+	ret = i915_gem_obj_prepare_shmem_write(bb->obj, &bb->clflush);
+	if (ret)
+		goto err_free_obj;
 
-	ret = i915_gem_object_set_to_cpu_domain(entry_obj->obj, false);
-	if (ret) {
-		gvt_vgpu_err("failed to set shadow batch to CPU\n");
-		goto unmap_src;
+	bb->va = i915_gem_object_pin_map(bb->obj, I915_MAP_WB);
+	if (IS_ERR(bb->va)) {
+		ret = PTR_ERR(bb->va);
+		goto err_finish_shmem_access;
 	}
 
-	entry_obj->va = dst;
-	entry_obj->bb_start_cmd_va = s->ip_va;
+	if (bb->clflush & CLFLUSH_BEFORE) {
+		drm_clflush_virt_range(bb->va, bb->obj->base.size);
+		bb->clflush &= ~CLFLUSH_BEFORE;
+	}
 
-	/* copy batch buffer to shadow batch buffer*/
 	ret = copy_gma_to_hva(s->vgpu, s->vgpu->gtt.ggtt_mm,
 			      gma, gma + bb_size,
-			      dst);
+			      bb->va);
 	if (ret < 0) {
 		gvt_vgpu_err("fail to copy guest ring buffer\n");
-		goto unmap_src;
+		ret = -EFAULT;
+		goto err_unmap;
 	}
 
-	list_add(&entry_obj->list, &s->workload->shadow_bb);
+	INIT_LIST_HEAD(&bb->list);
+	list_add(&bb->list, &s->workload->shadow_bb);
+
+	bb->accessing = true;
+	bb->bb_start_cmd_va = s->ip_va;
+
 	/*
 	 * ip_va saves the virtual address of the shadow batch buffer, while
 	 * ip_gma saves the graphics address of the original batch buffer.
@@ -1710,17 +1709,17 @@ static int perform_bb_shadow(struct parser_exec_state *s)
 	 * buffer's gma in pair. After all, we don't want to pin the shadow
 	 * buffer here (too early).
 	 */
-	s->ip_va = dst;
+	s->ip_va = bb->va;
 	s->ip_gma = gma;
-
 	return 0;
-
-unmap_src:
-	i915_gem_object_unpin_map(entry_obj->obj);
-put_obj:
-	i915_gem_object_put(entry_obj->obj);
-free_entry:
-	kfree(entry_obj);
+err_unmap:
+	i915_gem_object_unpin_map(bb->obj);
+err_finish_shmem_access:
+	i915_gem_obj_finish_shmem_access(bb->obj);
+err_free_obj:
+	i915_gem_object_put(bb->obj);
+err_free_bb:
+	kfree(bb);
 	return ret;
 }
 
@@ -1762,7 +1761,6 @@ static int cmd_handler_mi_batch_buffer_start(struct parser_exec_state *s)
 		if (ret < 0)
 			return ret;
 	}
-
 	return ret;
 }
 

drivers/gpu/drm/i915/gvt/scheduler.c

@@ -325,31 +325,46 @@ int intel_gvt_scan_and_shadow_workload(struct intel_vgpu_workload *workload)
 	return ret;
 }
 
+static void release_shadow_batch_buffer(struct intel_vgpu_workload *workload);
+
 static int prepare_shadow_batch_buffer(struct intel_vgpu_workload *workload)
 {
 	struct intel_gvt *gvt = workload->vgpu->gvt;
 	const int gmadr_bytes = gvt->device_info.gmadr_bytes_in_cmd;
-	struct intel_shadow_bb_entry *entry_obj;
+	struct intel_vgpu_shadow_bb *bb;
+	int ret;
 
-	/* pin the gem object to ggtt */
-	list_for_each_entry(entry_obj, &workload->shadow_bb, list) {
-		struct i915_vma *vma;
+	list_for_each_entry(bb, &workload->shadow_bb, list) {
+		bb->vma = i915_gem_object_ggtt_pin(bb->obj, NULL, 0, 0, 0);
+		if (IS_ERR(bb->vma)) {
+			ret = PTR_ERR(bb->vma);
+			goto err;
+		}
 
-		vma = i915_gem_object_ggtt_pin(entry_obj->obj, NULL, 0, 4, 0);
-		if (IS_ERR(vma))
-			return PTR_ERR(vma);
+		/* relocate shadow batch buffer */
+		bb->bb_start_cmd_va[1] = i915_ggtt_offset(bb->vma);
+		if (gmadr_bytes == 8)
+			bb->bb_start_cmd_va[2] = 0;
 
-		/* FIXME: we are not tracking our pinned VMA leaving it
-		 * up to the core to fix up the stray pin_count upon
-		 * free.
-		 */
+		/* No one is going to touch shadow bb from now on. */
+		if (bb->clflush & CLFLUSH_AFTER) {
+			drm_clflush_virt_range(bb->va, bb->obj->base.size);
+			bb->clflush &= ~CLFLUSH_AFTER;
+		}
 
-		/* update the relocate gma with shadow batch buffer*/
-		entry_obj->bb_start_cmd_va[1] = i915_ggtt_offset(vma);
-		if (gmadr_bytes == 8)
-			entry_obj->bb_start_cmd_va[2] = 0;
+		ret = i915_gem_object_set_to_gtt_domain(bb->obj, false);
+		if (ret)
+			goto err;
+
+		i915_gem_obj_finish_shmem_access(bb->obj);
+		bb->accessing = false;
+
+		i915_vma_move_to_active(bb->vma, workload->req, 0);
 	}
 	return 0;
+err:
+	release_shadow_batch_buffer(workload);
+	return ret;
 }
 
 static int update_wa_ctx_2_shadow_ctx(struct intel_shadow_wa_ctx *wa_ctx)
@@ -410,22 +425,37 @@ static int prepare_shadow_wa_ctx(struct intel_shadow_wa_ctx *wa_ctx)
 
 static void release_shadow_batch_buffer(struct intel_vgpu_workload *workload)
 {
-	/* release all the shadow batch buffer */
-	if (!list_empty(&workload->shadow_bb)) {
-		struct intel_shadow_bb_entry *entry_obj =
-			list_first_entry(&workload->shadow_bb,
-					 struct intel_shadow_bb_entry,
-					 list);
-		struct intel_shadow_bb_entry *temp;
-
-		list_for_each_entry_safe(entry_obj, temp, &workload->shadow_bb,
-					 list) {
-			i915_gem_object_unpin_map(entry_obj->obj);
-			i915_gem_object_put(entry_obj->obj);
-			list_del(&entry_obj->list);
-			kfree(entry_obj);
+	struct intel_vgpu *vgpu = workload->vgpu;
+	struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
+	struct intel_vgpu_shadow_bb *bb, *pos;
+
+	if (list_empty(&workload->shadow_bb))
+		return;
+
+	bb = list_first_entry(&workload->shadow_bb,
+			struct intel_vgpu_shadow_bb, list);
+
+	mutex_lock(&dev_priv->drm.struct_mutex);
+
+	list_for_each_entry_safe(bb, pos, &workload->shadow_bb, list) {
+		if (bb->obj) {
+			if (bb->accessing)
+				i915_gem_obj_finish_shmem_access(bb->obj);
+
+			if (bb->va && !IS_ERR(bb->va))
+				i915_gem_object_unpin_map(bb->obj);
+
+			if (bb->vma && !IS_ERR(bb->vma)) {
+				i915_vma_unpin(bb->vma);
+				i915_vma_close(bb->vma);
+			}
+			__i915_gem_object_release_unless_active(bb->obj);
 		}
+		list_del(&bb->list);
+		kfree(bb);
 	}
+
+	mutex_unlock(&dev_priv->drm.struct_mutex);
 }
 
 static int prepare_workload(struct intel_vgpu_workload *workload)

drivers/gpu/drm/i915/gvt/scheduler.h

@@ -112,13 +112,14 @@ struct intel_vgpu_workload {
 	struct intel_shadow_wa_ctx wa_ctx;
 };
 
-/* Intel shadow batch buffer is a i915 gem object */
-struct intel_shadow_bb_entry {
+struct intel_vgpu_shadow_bb {
 	struct list_head list;
 	struct drm_i915_gem_object *obj;
+	struct i915_vma *vma;
 	void *va;
-	unsigned long len;
 	u32 *bb_start_cmd_va;
+	unsigned int clflush;
+	bool accessing;
 };
 
 #define workload_q_head(vgpu, ring_id) \