Working with libdrm, repainting timeline

As the taiwins project finished with X11/Wayland backend rigging. Now I am fully on the libdrm backend development, dealing with hardwares directly. This backend requires 200 percent my energy to tame the complexity.

Like other backends, libdrm also needs to provide two resources, input and output. Backends like nested wayland backend or X11 backend, the output device are the windows created by ourselves (or by the user). Meaning we can create it or destroy it as we please. As for libdrm, those are provided by the OS. An output in libdrm’s terminology, is called connector. We query and change the connectors state by drmModeGet* functions. Then later set the hardware states by drmModeSet* or the new atomic API. It may sound simple but actually, for displaying images on the screen using the libdrm mechanism, it takes quite a few components, working together, to achieve the goal. We uses drmMode* routines for kernel-modesetting; libgbm for framebuffer allocation; OpenGL or Vulkan for rendering. If done right, you would get tear-free images on your screen constant refresh rate. But should anything goes wrong, you probably end up with a black screen.

I plan to create this series of blog of libdrm while I work on the taiwins drm-backend. Hopefully by the end of the month, I would have all the pieces together. In this Part I, we will look at the drm repainting loop.

DRM repaint timeline

Knowing vaguely the concepts is not sufficient to come up with a working setup, we want to know WHEN to do WHAT in order to get things right.

For a real world reference, here I steal an image1 of the weston repaint timeline.


We can see from the image, during every frame, the compositor does its rendering and tries to hit the vblank for displaying the result. the vblank is a timestamp here, it is used to be a period where the CRT monitors move electron gun from the bottom to the top. It is also the period for setting gamma and swapping image buffer. For now we just need to know that at vblank, there is a buffer swapping, also called page-flip.

If a client has double-buffering, it would want to

  1. Before vblank, draw on the back-buffer.
  2. After vblank, the back-buffer was used and it can draw on the new buffer.

Translating into code(here we use drmWaitBlank):

do {
	drmVBlank vbl = {
		.request.type = DRM_VBLANK_RELATIVE,
		.request.sequence = 1,
	drmWaitVBlank(fd, &vbl); 	//fd is the opened GPU
	//vblank returns, with the timestamp
	struct timespec spec = {
		.tv_sec = vbl.reply.tval_sec,
		.tv_nsec = vbl.reply.tval_usec * 1000,
	// repaint with new frame, return with new frame buffer, 
	int fb_id = repaint();
	//tell drm to use the new fb_id on present
	drmModeSetCrtc(fd, crtc_id, fb_id, 0,0, conn_id, mode);
} while (!quit);

In the snippet above, we wait for a vblank, then render and setCrtc in one shot. The obvious problem here is we have control of the duration of the rendering, it could exceed the vblank time window. Also there is a time we are waiting for nothing. This is not exactly what we want, luckly libdrm has a drmModePageFlip we can use.

static void handle_page_flip(int fd, unsigned int frame,
				             unsigned int sec, unsigned int usec,
				             void *data)
	int fb_id = repaint();
	struct timespec spec = {
		.tv_sec = sec,
		.tv_nsec = usec * 1000,
	drmModePageFlip(fd, crtc_id, fb_id, DRM_MODE_PAGE_FLIP_EVENT, data);

while(!quit) {
	drmEventContext ev = {
		.version = 2,
		.page_flip_handler = handle_page_flip,
	poll(fd); //the opend gpu.
	drmHandleEvent(fd, &ev);

This is a better example, which I modified from here2. This time we drawing on demand, using the non blocking drmModePageFlip to set buffer on next vblank. Note that the flag DRM_MODE_PAGE_FLIP_EVENT request the new page flip events for drawing, without it, we would not get notified. The potential problem is we may miss the next vblank if rendering takes too long (Triple-buffering with multi-threads could help).

  1. Weston repaint scheduling from Pekka Paalanen. ↩︎

  2. modeset-vsync from David Rheinsberg. ↩︎

comments powered by Disqus