Merge branch 'esync_cursor' into 'master'

Draft: cursor: explicit sync without `signal_release_with_buffer`

See merge request wlroots/wlroots!5302

examples/pointer.c

@@ -111,7 +111,7 @@ static void output_frame_notify(struct wl_listener *listener, void *data) {
 			state->clear_color[3],
 		},
 	});
-	wlr_output_add_software_cursors_to_render_pass(wlr_output, pass, NULL);
+	wlr_output_add_software_cursors_to_render_pass(wlr_output, pass, NULL, NULL, 0);
 	wlr_render_pass_submit(pass);
 	wlr_output_commit_state(wlr_output, &output_state);
 	wlr_output_state_finish(&output_state);

include/render/drm_syncobj_merger.h

@@ -3,6 +3,8 @@
 
 #include <wayland-server-core.h>
 
+struct wlr_buffer;
+
 /**
  * Accumulate timeline points, to have a destination timeline point be
  * signalled when all inputs are
@@ -41,4 +43,22 @@ bool wlr_drm_syncobj_merger_add(struct wlr_drm_syncobj_merger *merger,
 	struct wlr_drm_syncobj_timeline *dst_timeline, uint64_t dst_point,
 	struct wl_event_loop *loop);
 
+/**
+ * Add a new sync file to wait for.
+ *
+ * Ownership of fd is transferred to the merger.
+ */
+bool wlr_drm_syncobj_merger_add_sync_file(struct wlr_drm_syncobj_merger *merger,
+	int fd);
+
+/**
+ * Add a new DMA-BUF release to wait for.
+ *
+ * Waits for write access.
+ * If the platform does not support DMA-BUF<->sync file interop, the supplied
+ * event_loop is used to schedule a wait.
+ */
+bool wlr_drm_syncobj_merger_add_dmabuf(struct wlr_drm_syncobj_merger *merger,
+	struct wlr_buffer *buffer, struct wl_event_loop *event_loop);
+
 #endif

include/wlr/types/wlr_linux_drm_syncobj_v1.h

@@ -74,4 +74,21 @@ bool wlr_linux_drm_syncobj_v1_state_add_release_point(
 	struct wlr_drm_syncobj_timeline *release_timeline, uint64_t release_point,
 	struct wl_event_loop *event_loop);
 
+/**
+ * Register the DMA-BUF release of a buffer for buffer usage.
+ * Non-dmabuf buffers are considered to be immediately available (no wait).
+ *
+ * This function may be called multiple times for the same commit. The client's
+ * release point will be signalled when all registered points are signalled, and
+ * a new buffer has been committed.
+ *
+ * Because the platform may not support DMA-BUF fence merges, a wl_event_loop
+ * must be supplied to schedule a wait internally, if needed
+ *
+ * Waits for write access
+ */
+bool wlr_linux_drm_syncobj_v1_state_add_release_from_implicit_sync(
+	struct wlr_linux_drm_syncobj_surface_v1_state *state,
+	struct wlr_buffer *buffer, struct wl_event_loop *event_loop);
+
 #endif

include/wlr/types/wlr_output.h

@@ -36,6 +36,11 @@ struct wlr_output_mode {
 	struct wl_list link;
 };
 
+struct wlr_output_cursor_texture_sample_event {
+	struct wlr_drm_syncobj_timeline *release_timeline;
+	uint64_t release_point;
+};
+
 struct wlr_output_cursor {
 	struct wlr_output *output;
 	double x, y;
@@ -51,6 +56,10 @@ struct wlr_output_cursor {
 	uint64_t wait_point;
 	struct wl_list link;
 
+	struct {
+		struct wl_signal texture_sample; // struct wlr_output_cursor_texture_sample_event
+	} events;
+
 	struct {
 		struct wl_listener renderer_destroy;
 		struct wlr_color_transform *color_transform;
@@ -442,7 +451,8 @@ void wlr_output_lock_software_cursors(struct wlr_output *output, bool lock);
  * This is a utility function that can be called when compositors render.
  */
 void wlr_output_add_software_cursors_to_render_pass(struct wlr_output *output,
-	struct wlr_render_pass *render_pass, const pixman_region32_t *damage);
+	struct wlr_render_pass *render_pass, const pixman_region32_t *damage,
+	struct wlr_drm_syncobj_timeline *release_timeline, uint64_t release_point);
 /**
  * Get the set of DRM formats suitable for the primary buffer, assuming a
  * buffer with the specified capabilities.

render/drm_syncobj_merger.c

@@ -4,8 +4,10 @@
 #include <unistd.h>
 #include <wayland-util.h>
 #include <wlr/render/drm_syncobj.h>
+#include <wlr/types/wlr_buffer.h>
 #include <wlr/util/log.h>
 #include <xf86drm.h>
+#include "render/dmabuf.h"
 #include "render/drm_syncobj_merger.h"
 
 #include "config.h"
@@ -79,14 +81,7 @@ void wlr_drm_syncobj_merger_unref(struct wlr_drm_syncobj_merger *merger) {
 static bool merger_add_exportable(struct wlr_drm_syncobj_merger *merger,
 		struct wlr_drm_syncobj_timeline *src_timeline, uint64_t src_point) {
 	int new_sync = wlr_drm_syncobj_timeline_export_sync_file(src_timeline, src_point);
-	if (merger->sync_fd != -1) {
-		int fd2 = new_sync;
-		new_sync = sync_file_merge(merger->sync_fd, fd2);
-		close(fd2);
-		close(merger->sync_fd);
-	}
-	merger->sync_fd = new_sync;
-	return true;
+	return wlr_drm_syncobj_merger_add_sync_file(merger, new_sync);
 }
 
 struct export_waiter {
@@ -131,3 +126,69 @@ bool wlr_drm_syncobj_merger_add(struct wlr_drm_syncobj_merger *merger,
 	merger->n_ref++;
 	return true;
 }
+
+bool wlr_drm_syncobj_merger_add_sync_file(struct wlr_drm_syncobj_merger *merger,
+		int fd) {
+	int new_sync = fd;
+	if (merger->sync_fd != -1) {
+		new_sync = sync_file_merge(merger->sync_fd, fd);
+		close(fd);
+		close(merger->sync_fd);
+	}
+	merger->sync_fd = new_sync;
+	return merger->sync_fd != -1;
+}
+
+struct poll_waiter {
+	struct wl_event_source *event_source;
+	struct wlr_drm_syncobj_merger *merger;
+};
+
+static int poll_waiter_handle_done(int fd, uint32_t mask, void *data) {
+	struct poll_waiter *waiter = data;
+	wlr_drm_syncobj_merger_unref(waiter->merger);
+	wl_event_source_remove(waiter->event_source);
+	free(waiter);
+	return 0;
+}
+
+bool wlr_drm_syncobj_merger_add_dmabuf(struct wlr_drm_syncobj_merger *merger,
+		struct wlr_buffer *buffer, struct wl_event_loop *event_loop) {
+	struct wlr_dmabuf_attributes dmabuf_attributes;
+	if (!wlr_buffer_get_dmabuf(buffer, &dmabuf_attributes)) {
+		return true;
+	}
+
+	bool res = true;
+	for (int i = 0; i < dmabuf_attributes.n_planes; ++i) {
+		int sync_fd = dmabuf_export_sync_file(dmabuf_attributes.fd[i], DMA_BUF_SYNC_WRITE);
+		if (sync_fd == -1) {
+			res = false;
+			break;
+		}
+		if (!wlr_drm_syncobj_merger_add_sync_file(merger, sync_fd)) {
+			return false;
+		}
+	}
+
+	if (res) {
+		return true;
+	}
+
+	uint32_t mask = WL_EVENT_ERROR | WL_EVENT_HANGUP | WL_EVENT_WRITABLE;
+	for (int i = 0; i < dmabuf_attributes.n_planes; ++i) {
+		struct poll_waiter *waiter = calloc(1, sizeof(*waiter));
+		if (waiter == NULL) {
+			return false;
+		}
+		waiter->merger = wlr_drm_syncobj_merger_ref(merger);
+		waiter->event_source = wl_event_loop_add_fd(event_loop,
+			dmabuf_attributes.fd[i], mask, poll_waiter_handle_done, waiter);
+		if (waiter->event_source == NULL) {
+			wlr_drm_syncobj_merger_unref(waiter->merger);
+			free(waiter);
+			return false;
+		}
+	}
+	return true;
+}

types/output/cursor.c

@@ -89,7 +89,8 @@ static void output_cursor_get_box(struct wlr_output_cursor *cursor,
 }
 
 void wlr_output_add_software_cursors_to_render_pass(struct wlr_output *output,
-		struct wlr_render_pass *render_pass, const pixman_region32_t *damage) {
+		struct wlr_render_pass *render_pass, const pixman_region32_t *damage,
+		struct wlr_drm_syncobj_timeline *release_timeline, uint64_t release_point) {
 	int width, height;
 	wlr_output_transformed_resolution(output, &width, &height);
 
@@ -128,7 +129,14 @@ void wlr_output_add_software_cursors_to_render_pass(struct wlr_output *output,
 			.dst_box = box,
 			.clip = &cursor_damage,
 			.transform = output->transform,
+			.wait_timeline = cursor->wait_timeline,
+			.wait_point = cursor->wait_point,
 		});
+		struct wlr_output_cursor_texture_sample_event event = {
+			.release_timeline = release_timeline,
+			.release_point = release_point,
+		};
+		wl_signal_emit_mutable(&cursor->events.texture_sample, &event);
 
 		pixman_region32_fini(&cursor_damage);
 	}
@@ -285,6 +293,9 @@ static struct wlr_buffer *render_cursor_buffer(struct wlr_output_cursor *cursor)
 		return NULL;
 	}
 
+	struct wlr_output_cursor_texture_sample_event event = {0};
+	wl_signal_emit_mutable(&cursor->events.texture_sample, &event);
+
 	return buffer;
 }
 
@@ -472,6 +483,7 @@ struct wlr_output_cursor *wlr_output_cursor_create(struct wlr_output *output) {
 	cursor->output = output;
 	wl_list_insert(&output->cursors, &cursor->link);
 	cursor->visible = true; // default position is at (0, 0)
+	wl_signal_init(&cursor->events.texture_sample);
 	wl_list_init(&cursor->renderer_destroy.link);
 	output_cursor_refresh_color_transform(cursor, output->image_description);
 	return cursor;
@@ -493,6 +505,7 @@ void wlr_output_cursor_destroy(struct wlr_output_cursor *cursor) {
 	}
 	wlr_drm_syncobj_timeline_unref(cursor->wait_timeline);
 	wl_list_remove(&cursor->link);
+	assert(wl_list_empty(&cursor->events.texture_sample.listener_list));
 	wlr_color_transform_unref(cursor->color_transform);
 	free(cursor);
 }

types/scene/wlr_scene.c

@@ -2581,7 +2581,8 @@ bool wlr_scene_output_build_state(struct wlr_scene_output *scene_output,
 		}
 	}
 
-	wlr_output_add_software_cursors_to_render_pass(output, render_pass, &render_data.damage);
+	wlr_output_add_software_cursors_to_render_pass(output, render_pass,
+		&render_data.damage, scene_output->in_timeline, scene_output->in_point);
 	pixman_region32_fini(&render_data.damage);
 
 	if (!wlr_render_pass_submit(render_pass)) {

types/wlr_cursor.c

@@ -63,6 +63,7 @@ struct wlr_cursor_output_cursor {
 
 	// only when using a surface as the cursor image
 	struct wl_listener output_commit;
+	struct wl_listener texture_sample;
 
 	// only when using an XCursor as the cursor image
 	struct wlr_xcursor *xcursor;
@@ -160,6 +161,7 @@ static void output_cursor_destroy(struct wlr_cursor_output_cursor *output_cursor
 	wl_list_remove(&output_cursor->layout_output_destroy.link);
 	wl_list_remove(&output_cursor->link);
 	wl_list_remove(&output_cursor->output_commit.link);
+	wl_list_remove(&output_cursor->texture_sample.link);
 	wlr_output_cursor_destroy(output_cursor->output_cursor);
 	free(output_cursor);
 }
@@ -579,13 +581,6 @@ static void cursor_output_cursor_update(struct wlr_cursor_output_cursor *output_
 			&src_box, dst_width, dst_height, surface->current.transform,
 			hotspot_x, hotspot_y, wait_timeline, wait_point);
 
-		if (syncobj_surface_state != NULL &&
-				surface->buffer != NULL && surface->buffer->source != NULL &&
-				(surface->current.committed & WLR_SURFACE_STATE_BUFFER)) {
-			wlr_linux_drm_syncobj_v1_state_signal_release_with_buffer(syncobj_surface_state,
-				surface->buffer->source);
-		}
-
 		if (output_cursor->output_cursor->visible) {
 			wlr_surface_send_enter(surface, output);
 		} else {
@@ -645,6 +640,29 @@ static void output_cursor_output_handle_output_commit(
 	}
 }
 
+static void output_cursor_output_handle_texture_sample(
+		struct wl_listener *listener, void *data) {
+	struct wlr_cursor_output_cursor *output_cursor =
+		wl_container_of(listener, output_cursor, texture_sample);
+	const struct wlr_output_cursor_texture_sample_event *event = data;
+
+	struct wlr_surface *surface = output_cursor->cursor->state->surface;
+	struct wlr_linux_drm_syncobj_surface_v1_state *syncobj_state = NULL;
+	if (surface != NULL) {
+		syncobj_state = wlr_linux_drm_syncobj_v1_get_surface_state(surface);
+	}
+	if (syncobj_state != NULL) {
+		struct wl_event_loop *event_loop = output_cursor->output_cursor->output->event_loop;
+		if (event->release_timeline != NULL) {
+			wlr_linux_drm_syncobj_v1_state_add_release_point(syncobj_state,
+				event->release_timeline, event->release_point, event_loop);
+		} else if (surface->buffer != NULL && surface->buffer->source != NULL) {
+			wlr_linux_drm_syncobj_v1_state_add_release_from_implicit_sync(syncobj_state,
+				surface->buffer->source, event_loop);
+		}
+	}
+}
+
 static void cursor_update_outputs(struct wlr_cursor *cur) {
 	struct wlr_cursor_output_cursor *output_cursor;
 	wl_list_for_each(output_cursor, &cur->state->output_cursors, link) {
@@ -1167,6 +1185,9 @@ static void layout_add(struct wlr_cursor_state *state,
 	wl_signal_add(&output_cursor->output_cursor->output->events.commit,
 		&output_cursor->output_commit);
 	output_cursor->output_commit.notify = output_cursor_output_handle_output_commit;
+	wl_signal_add(&output_cursor->output_cursor->events.texture_sample,
+		&output_cursor->texture_sample);
+	output_cursor->texture_sample.notify = output_cursor_output_handle_texture_sample;
 
 	output_cursor_move(output_cursor);
 	cursor_output_cursor_update(output_cursor);

types/wlr_linux_drm_syncobj_v1.c

@@ -12,6 +12,7 @@
 #include <xf86drm.h>
 #include "config.h"
 #include "linux-drm-syncobj-v1-protocol.h"
+#include "render/dmabuf.h"
 #include "render/drm_syncobj_merger.h"
 
 #define LINUX_DRM_SYNCOBJ_V1_VERSION 1
@@ -540,3 +541,12 @@ bool wlr_linux_drm_syncobj_v1_state_add_release_point(
 	return wlr_drm_syncobj_merger_add(state->release_merger,
 		release_timeline, release_point, event_loop);
 }
+
+bool wlr_linux_drm_syncobj_v1_state_add_release_from_implicit_sync(
+		struct wlr_linux_drm_syncobj_surface_v1_state *state,
+		struct wlr_buffer *buffer, struct wl_event_loop *event_loop) {
+	if (state->release_merger != NULL) {
+		return true;
+	}
+	return wlr_drm_syncobj_merger_add_dmabuf(state->release_merger, buffer, event_loop);
+}