output: detach buffer from renderer before commit

Right now we rely entirely on implicit sync for synchronizing
access to GPU buffers. Implicit sync works by setting
synchronization points on the buffer in writers, and letting
readers wait on these sync points before accessing the buffer.

With OpenGL, sync points are created using functions such as
eglSwapBuffers or glFlush. If none of these special functions
are called, no sync point will be created and readers will
potentially access a buffer that hasn't finished rendering yet.

In the context of wlroots, OpenGL is the writer and the backend
(KMS or parent Wayland/X11 session) is the reader. After we're
done rendering a frame, and before passing that frame to the
backend, we need to call glFlush.

glFlush is called when the buffer is detached from the renderer.
This is a task done by output_clear_back_buffer. So let's call
this function before invoking the impl->commit hook, instead of
calling it after.

All of this is maybe a little tricky to get right with the
current renderer_bind_buffer API. The new
wlr_renderer_begin_with_buffer API is much better, because glFlush
is called on wlr_renderer_end, so it's more intuitive.

Closes: https://github.com/swaywm/wlroots/issues/3020

types/wlr_output.c

@@ -736,8 +736,19 @@ bool wlr_output_commit(struct wlr_output *output) {
 	};
 	wlr_signal_emit_safe(&output->events.precommit, &pre_event);
 
-	if (!output->impl->commit(output)) {
+	// output_clear_back_buffer detaches the buffer from the renderer. This is
+	// important to do before calling impl->commit(), because this marks an
+	// implicit rendering synchronization point. The backend needs it to avoid
+	// displaying a buffer when asynchronous GPU work isn't finished.
+	struct wlr_buffer *back_buffer = NULL;
+	if ((output->pending.committed & WLR_OUTPUT_STATE_BUFFER) &&
+			output->back_buffer != NULL) {
+		back_buffer = wlr_buffer_lock(output->back_buffer);
 		output_clear_back_buffer(output);
+	}
+
+	if (!output->impl->commit(output)) {
+		wlr_buffer_unlock(back_buffer);
 		output_state_clear(&output->pending);
 		return false;
 	}
@@ -782,12 +793,11 @@ bool wlr_output_commit(struct wlr_output *output) {
 	if (output->pending.committed & WLR_OUTPUT_STATE_BUFFER) {
 		output->frame_pending = true;
 		output->needs_frame = false;
-
-		if (output->back_buffer != NULL) {
-			wlr_swapchain_set_buffer_submitted(output->swapchain,
-				output->back_buffer);
-			output_clear_back_buffer(output);
 	}
+
+	if (back_buffer != NULL) {
+		wlr_swapchain_set_buffer_submitted(output->swapchain, back_buffer);
+		wlr_buffer_unlock(back_buffer);
 	}
 
 	uint32_t committed = output->pending.committed;