drm/vc4: Add support for async pageflips.

An async pageflip stores the modeset to be done and executes it once
the BOs are ready to be displayed.  This gets us about 3x performance
in full screen rendering with pageflipping.

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

1 files changed, 148 insertions, 1 deletions

diff --git a/drivers/gpu/drm/vc4/vc4_kms.c b/drivers/gpu/drm/vc4/vc4_kms.c
index 2e5597d..f95f2df 100644
--- a/drivers/gpu/drm/vc4/vc4_kms.c
+++ b/drivers/gpu/drm/vc4/vc4_kms.c
@@ -15,6 +15,7 @@
  */
 
 #include "drm_crtc.h"
+#include "drm_atomic.h"
 #include "drm_atomic_helper.h"
 #include "drm_crtc_helper.h"
 #include "drm_plane_helper.h"
@@ -29,10 +30,152 @@ static void vc4_output_poll_changed(struct drm_device *dev)
 		drm_fbdev_cma_hotplug_event(vc4->fbdev);
 }
 
+struct vc4_commit {
+	struct drm_device *dev;
+	struct drm_atomic_state *state;
+	struct vc4_seqno_cb cb;
+};
+
+static void
+vc4_atomic_complete_commit(struct vc4_commit *c)
+{
+	struct drm_atomic_state *state = c->state;
+	struct drm_device *dev = state->dev;
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+	drm_atomic_helper_commit_modeset_disables(dev, state);
+
+	drm_atomic_helper_commit_planes(dev, state, false);
+
+	drm_atomic_helper_commit_modeset_enables(dev, state);
+
+	drm_atomic_helper_wait_for_vblanks(dev, state);
+
+	drm_atomic_helper_cleanup_planes(dev, state);
+
+	drm_atomic_state_free(state);
+
+	up(&vc4->async_modeset);
+
+	kfree(c);
+}
+
+static void
+vc4_atomic_complete_commit_seqno_cb(struct vc4_seqno_cb *cb)
+{
+	struct vc4_commit *c = container_of(cb, struct vc4_commit, cb);
+
+	vc4_atomic_complete_commit(c);
+}
+
+static struct vc4_commit *commit_init(struct drm_atomic_state *state)
+{
+	struct vc4_commit *c = kzalloc(sizeof(*c), GFP_KERNEL);
+
+	if (!c)
+		return NULL;
+	c->dev = state->dev;
+	c->state = state;
+
+	return c;
+}
+
+/**
+ * vc4_atomic_commit - commit validated state object
+ * @dev: DRM device
+ * @state: the driver state object
+ * @async: asynchronous commit
+ *
+ * This function commits a with drm_atomic_helper_check() pre-validated state
+ * object. This can still fail when e.g. the framebuffer reservation fails. For
+ * now this doesn't implement asynchronous commits.
+ *
+ * RETURNS
+ * Zero for success or -errno.
+ */
+static int vc4_atomic_commit(struct drm_device *dev,
+			     struct drm_atomic_state *state,
+			     bool async)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	int ret;
+	int i;
+	uint64_t wait_seqno = 0;
+	struct vc4_commit *c;
+
+	c = commit_init(state);
+	if (!c)
+		return -ENOMEM;
+
+	/* Make sure that any outstanding modesets have finished. */
+	ret = down_interruptible(&vc4->async_modeset);
+	if (ret) {
+		kfree(c);
+		return ret;
+	}
+
+	ret = drm_atomic_helper_prepare_planes(dev, state);
+	if (ret) {
+		kfree(c);
+		up(&vc4->async_modeset);
+		return ret;
+	}
+
+	for (i = 0; i < dev->mode_config.num_total_plane; i++) {
+		struct drm_plane *plane = state->planes[i];
+		struct drm_plane_state *new_state = state->plane_states[i];
+
+		if (!plane)
+			continue;
+
+		if ((plane->state->fb != new_state->fb) && new_state->fb) {
+			struct drm_gem_cma_object *cma_bo =
+				drm_fb_cma_get_gem_obj(new_state->fb, 0);
+			struct vc4_bo *bo = to_vc4_bo(&cma_bo->base);
+
+			wait_seqno = max(bo->seqno, wait_seqno);
+		}
+	}
+
+	/*
+	 * This is the point of no return - everything below never fails except
+	 * when the hw goes bonghits. Which means we can commit the new state on
+	 * the software side now.
+	 */
+
+	drm_atomic_helper_swap_state(dev, state);
+
+	/*
+	 * Everything below can be run asynchronously without the need to grab
+	 * any modeset locks at all under one condition: It must be guaranteed
+	 * that the asynchronous work has either been cancelled (if the driver
+	 * supports it, which at least requires that the framebuffers get
+	 * cleaned up with drm_atomic_helper_cleanup_planes()) or completed
+	 * before the new state gets committed on the software side with
+	 * drm_atomic_helper_swap_state().
+	 *
+	 * This scheme allows new atomic state updates to be prepared and
+	 * checked in parallel to the asynchronous completion of the previous
+	 * update. Which is important since compositors need to figure out the
+	 * composition of the next frame right after having submitted the
+	 * current layout.
+	 */
+
+	if (async) {
+		vc4_queue_seqno_cb(dev, &c->cb, wait_seqno,
+				   vc4_atomic_complete_commit_seqno_cb);
+	} else {
+		vc4_wait_for_seqno(dev, wait_seqno, ~0ull, false);
+		vc4_atomic_complete_commit(c);
+	}
+
+	return 0;
+}
+
 static const struct drm_mode_config_funcs vc4_mode_funcs = {
 	.output_poll_changed = vc4_output_poll_changed,
 	.atomic_check = drm_atomic_helper_check,
-	.atomic_commit = drm_atomic_helper_commit,
+	.atomic_commit = vc4_atomic_commit,
 	.fb_create = drm_fb_cma_create,
 };
 
@@ -41,6 +184,8 @@ int vc4_kms_load(struct drm_device *dev)
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
 	int ret;
 
+	sema_init(&vc4->async_modeset, 1);
+
 	ret = drm_vblank_init(dev, dev->mode_config.num_crtc);
 	if (ret < 0) {
 		dev_err(dev->dev, "failed to initialize vblank\n");
@@ -51,6 +196,8 @@ int vc4_kms_load(struct drm_device *dev)
 	dev->mode_config.max_height = 2048;
 	dev->mode_config.funcs = &vc4_mode_funcs;
 	dev->mode_config.preferred_depth = 24;
+	dev->mode_config.async_page_flip = true;
+
 	dev->vblank_disable_allowed = true;
 
 	drm_mode_config_reset(dev);