backend/drm: add explicit sync support to multi-GPU blits

backend/drm/atomic.c

@@ -275,9 +275,9 @@ bool drm_atomic_connector_prepare(struct wlr_drm_connector_state *state, bool mo
 	}
 
 	int in_fence_fd = -1;
-	if (state->base->committed & WLR_OUTPUT_STATE_WAIT_TIMELINE) {
+	if (state->wait_timeline != NULL) {
-		in_fence_fd = wlr_drm_syncobj_timeline_export_sync_file(state->base->wait_timeline,
+		in_fence_fd = wlr_drm_syncobj_timeline_export_sync_file(state->wait_timeline,
-			state->base->wait_point);
+			state->wait_point);
 		if (in_fence_fd < 0) {
 			return false;
 		}

backend/drm/drm.c

@@ -14,6 +14,7 @@
 #include <wayland-util.h>
 #include <wlr/backend/interface.h>
 #include <wlr/interfaces/wlr_output.h>
+#include <wlr/render/drm_syncobj.h>
 #include <wlr/render/wlr_renderer.h>
 #include <wlr/util/box.h>
 #include <wlr/util/log.h>
@@ -693,6 +694,7 @@ static void drm_connector_state_init(struct wlr_drm_connector_state *state,
 static void drm_connector_state_finish(struct wlr_drm_connector_state *state) {
 	drm_fb_clear(&state->primary_fb);
 	drm_fb_clear(&state->cursor_fb);
+	wlr_drm_syncobj_timeline_unref(state->wait_timeline);
 }
 
 static bool drm_connector_state_update_primary_fb(struct wlr_drm_connector *conn,
@@ -707,6 +709,14 @@ static bool drm_connector_state_update_primary_fb(struct wlr_drm_connector *conn
 	struct wlr_drm_plane *plane = crtc->primary;
 	struct wlr_buffer *source_buf = state->base->buffer;
 
+	struct wlr_drm_syncobj_timeline *wait_timeline = NULL;
+	uint64_t wait_point = 0;
+	if (state->base->committed & WLR_OUTPUT_STATE_WAIT_TIMELINE) {
+		wait_timeline = state->base->wait_timeline;
+		wait_point = state->base->wait_point;
+	}
+	assert(state->wait_timeline == NULL);
+
 	struct wlr_buffer *local_buf;
 	if (drm->parent) {
 		struct wlr_drm_format format = {0};
@@ -723,12 +733,23 @@ static bool drm_connector_state_update_primary_fb(struct wlr_drm_connector *conn
 			return false;
 		}
 
-		local_buf = drm_surface_blit(&plane->mgpu_surf, source_buf);
+		local_buf = drm_surface_blit(&plane->mgpu_surf, source_buf,
+			wait_timeline, wait_point);
 		if (local_buf == NULL) {
 			return false;
 		}
+
+		if (plane->mgpu_surf.timeline != NULL) {
+			state->wait_timeline = wlr_drm_syncobj_timeline_ref(plane->mgpu_surf.timeline);
+			state->wait_point = plane->mgpu_surf.point;
+		}
 	} else {
 		local_buf = wlr_buffer_lock(source_buf);
+
+		if (wait_timeline != NULL) {
+			state->wait_timeline = wlr_drm_syncobj_timeline_ref(wait_timeline);
+			state->wait_point = wait_point;
+		}
 	}
 
 	bool ok = drm_fb_import(&state->primary_fb, drm, local_buf,
@@ -1088,7 +1109,7 @@ static bool drm_connector_set_cursor(struct wlr_output *output,
 				return false;
 			}
 
-			local_buf = drm_surface_blit(&plane->mgpu_surf, buffer);
+			local_buf = drm_surface_blit(&plane->mgpu_surf, buffer, NULL, 0);
 			if (local_buf == NULL) {
 				return false;
 			}
@@ -1625,9 +1646,11 @@ static bool connect_drm_connector(struct wlr_drm_connector *wlr_conn,
 		output->non_desktop = non_desktop;
 	}
 
-	// TODO: support sync timelines in multi-GPU mode
 	// TODO: support sync timelines with libliftoff
-	output->timeline = drm->parent == NULL && drm->iface == &atomic_iface;
+	output->timeline = drm->iface == &atomic_iface;
+	if (drm->parent) {
+		output->timeline = output->timeline && drm->mgpu_renderer.wlr_rend->features.timeline;
+	}
 
 	memset(wlr_conn->max_bpc_bounds, 0, sizeof(wlr_conn->max_bpc_bounds));
 	if (wlr_conn->props.max_bpc != 0) {

backend/drm/renderer.c

@@ -1,5 +1,6 @@
 #include <assert.h>
 #include <drm_fourcc.h>
+#include <wlr/render/drm_syncobj.h>
 #include <wlr/render/swapchain.h>
 #include <wlr/render/wlr_renderer.h>
 #include <wlr/util/log.h>
@@ -46,6 +47,7 @@ void finish_drm_surface(struct wlr_drm_surface *surf) {
 		return;
 	}
 
+	wlr_drm_syncobj_timeline_unref(surf->timeline);
 	wlr_swapchain_destroy(surf->swapchain);
 
 	*surf = (struct wlr_drm_surface){0};
@@ -68,13 +70,24 @@ bool init_drm_surface(struct wlr_drm_surface *surf,
 		return false;
 	}
 
+	int drm_fd = wlr_renderer_get_drm_fd(renderer->wlr_rend);
+	if (renderer->wlr_rend->features.timeline && drm_fd >= 0) {
+		surf->timeline = wlr_drm_syncobj_timeline_create(drm_fd);
+		if (surf->timeline == NULL) {
+			finish_drm_surface(surf);
+			wlr_log(WLR_ERROR, "Failed to create DRM syncobj timeline");
+			return false;
+		}
+	}
+
 	surf->renderer = renderer;
 
 	return true;
 }
 
 struct wlr_buffer *drm_surface_blit(struct wlr_drm_surface *surf,
-		struct wlr_buffer *buffer) {
+		struct wlr_buffer *buffer,
+		struct wlr_drm_syncobj_timeline *wait_timeline, uint64_t wait_point) {
 	struct wlr_renderer *renderer = surf->renderer->wlr_rend;
 
 	if (surf->swapchain->width != buffer->width ||
@@ -95,7 +108,12 @@ struct wlr_buffer *drm_surface_blit(struct wlr_drm_surface *surf,
 		goto error_tex;
 	}
 
-	struct wlr_render_pass *pass = wlr_renderer_begin_buffer_pass(renderer, dst, NULL);
+	surf->point++;
+	const struct wlr_buffer_pass_options pass_options = {
+		.signal_timeline = surf->timeline,
+		.signal_point = surf->point,
+	};
+	struct wlr_render_pass *pass = wlr_renderer_begin_buffer_pass(renderer, dst, &pass_options);
 	if (pass == NULL) {
 		wlr_log(WLR_ERROR, "Failed to begin render pass with multi-GPU destination buffer");
 		goto error_dst;
@@ -104,6 +122,8 @@ struct wlr_buffer *drm_surface_blit(struct wlr_drm_surface *surf,
 	wlr_render_pass_add_texture(pass, &(struct wlr_render_texture_options){
 		.texture = tex,
 		.blend_mode = WLR_RENDER_BLEND_MODE_NONE,
+		.wait_timeline = wait_timeline,
+		.wait_point = wait_point,
 	});
 	if (!wlr_render_pass_submit(pass)) {
 		wlr_log(WLR_ERROR, "Failed to submit multi-GPU render pass");

include/backend/drm/drm.h

@@ -141,6 +141,9 @@ struct wlr_drm_connector_state {
 	struct wlr_drm_fb *primary_fb;
 	struct wlr_drm_fb *cursor_fb;
 
+	struct wlr_drm_syncobj_timeline *wait_timeline;
+	uint64_t wait_point;
+
 	// used by atomic
 	uint32_t mode_id;
 	uint32_t gamma_lut;

include/backend/drm/renderer.h

@@ -20,6 +20,9 @@ struct wlr_drm_renderer {
 struct wlr_drm_surface {
 	struct wlr_drm_renderer *renderer;
 	struct wlr_swapchain *swapchain;
+
+	struct wlr_drm_syncobj_timeline *timeline;
+	uint64_t point;
 };
 
 bool init_drm_renderer(struct wlr_drm_backend *drm,
@@ -32,7 +35,8 @@ bool init_drm_surface(struct wlr_drm_surface *surf,
 void finish_drm_surface(struct wlr_drm_surface *surf);
 
 struct wlr_buffer *drm_surface_blit(struct wlr_drm_surface *surf,
-	struct wlr_buffer *buffer);
+	struct wlr_buffer *buffer,
+	struct wlr_drm_syncobj_timeline *wait_timeline, uint64_t wait_point);
 
 bool drm_plane_pick_render_format(struct wlr_drm_plane *plane,
 	struct wlr_drm_format *fmt, struct wlr_drm_renderer *renderer);