drm/vc4: Add support for binner overflow memory allocation.

We should probably hang on to some of this memory across renders,
since most renders will want some binner overflow.

Signed-off-by: Eric Anholt <eric@anholt.net>

2 files changed, 52 insertions, 6 deletions

diff --git a/drivers/gpu/drm/vc4/vc4_drv.h b/drivers/gpu/drm/vc4/vc4_drv.h
index e741627211a1..82482a3fb5f6 100644
--- a/drivers/gpu/drm/vc4/vc4_drv.h
+++ b/drivers/gpu/drm/vc4/vc4_drv.h
@@ -92,6 +92,11 @@ enum vc4_bo_mode {
 	VC4_MODE_SHADER,
 };
 
+struct vc4_bo_list_entry {
+	struct list_head head;
+	struct drm_gem_cma_object *bo;
+};
+
 struct vc4_bo_exec_state {
 	struct drm_gem_cma_object *bo;
 	enum vc4_bo_mode mode;
@@ -117,6 +122,12 @@ struct exec_info {
 	 */
 	struct drm_gem_cma_object *exec_bo;
 
+	/* List of struct vc4_list_bo_entry allocated to accomodate
+	 * binner overflow.  These will be freed when the exec is
+	 * done.
+	 */
+	struct list_head overflow_list;
+
 	/**
 	 * This tracks the per-shader-record state (packet 64) that
 	 * determines the length of the shader record and the offset
diff --git a/drivers/gpu/drm/vc4/vc4_gem.c b/drivers/gpu/drm/vc4/vc4_gem.c
index 3918d5bc8723..c449420d06c2 100644
--- a/drivers/gpu/drm/vc4/vc4_gem.c
+++ b/drivers/gpu/drm/vc4/vc4_gem.c
@@ -70,10 +70,35 @@ thread_stopped(struct drm_device *dev, uint32_t thread)
 }
 
 static int
-wait_for_bin_thread(struct drm_device *dev)
+try_adding_overflow_memory(struct drm_device *dev, struct exec_info *exec)
 {
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
-	int i;
+	struct vc4_bo_list_entry *entry = kmalloc(sizeof(*entry), GFP_KERNEL);
+
+	if (!entry)
+		return -ENOMEM;
+
+	entry->bo = drm_gem_cma_create(dev, 256 * 1024);
+	if (IS_ERR(entry->bo)) {
+		int ret = PTR_ERR(entry->bo);
+		DRM_ERROR("Couldn't allocate binner overflow mem\n");
+		kfree(entry);
+		return ret;
+	}
+
+	list_add_tail(&entry->head, &exec->overflow_list);
+
+	V3D_WRITE(V3D_BPOA, entry->bo->paddr);
+	V3D_WRITE(V3D_BPOS, entry->bo->base.size);
+
+	return 0;
+}
+
+static int
+wait_for_bin_thread(struct drm_device *dev, struct exec_info *exec)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	int i, ret;
 
 	for (i = 0; i < 1000000; i++) {
 		if (thread_stopped(dev, 0)) {
@@ -86,9 +111,9 @@ wait_for_bin_thread(struct drm_device *dev)
 		}
 
 		if (V3D_READ(V3D_PCS) & V3D_BMOOM) {
-			/* XXX */
-			DRM_ERROR("binner oom\n");
-			return -EINVAL;
+			ret = try_adding_overflow_memory(dev, exec);
+			if (ret)
+				return ret;
 		}
 	}
 
@@ -148,7 +173,7 @@ vc4_submit(struct drm_device *dev, struct exec_info *exec)
 
 	submit_cl(dev, 0, ct0ca, ct0ea);
 
-	ret = wait_for_bin_thread(dev);
+	ret = wait_for_bin_thread(dev, exec);
 	if (ret)
 		return ret;
 
@@ -366,6 +391,7 @@ vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
 
 	memset(&exec, 0, sizeof(exec));
 	exec.args = data;
+	INIT_LIST_HEAD(&exec.overflow_list);
 
 	mutex_lock(&dev->struct_mutex);
 
@@ -390,6 +416,15 @@ fail:
 		kfree(exec.bo);
 	}
 
+	while (!list_empty(&exec.overflow_list)) {
+		struct vc4_bo_list_entry *entry =
+			list_first_entry(&exec.overflow_list,
+					 struct vc4_bo_list_entry, head);
+		drm_gem_object_unreference(&entry->bo->base);
+		list_del(&entry->head);
+		kfree(entry);
+	}
+
 	drm_gem_object_unreference(&exec.exec_bo->base);
 
 	mutex_unlock(&dev->struct_mutex);