Site logo
Stories around the Genode Operating System RSS feed
Norman Feske avatar

Re-stacking the GUI stack


As I am currently right in the middle of a far-reaching rework of Genode's low-level GUI stack, I'd like to share a bit of background behind this work: the Why, the What, and the How.

Current state

The current incarnation of Genode's low-level GUI-related interfaces is almost identical to the very first version that we originally created in 2006. Genode has a very strong asymmetric notion of the relationship between client and server components. In this relationship, the client depends on the server but the server does not depend on the client.

At the lowest level, the framebuffer driver and input driver(s) were envisioned as servers that provide the Framebuffer and Input session interfaces to higher-level GUI components. From our former academic view, this felt natural because those drivers provide resources to other components. Provide... serve... server! The nitpicker GUI server uses those low-level services to implement a higher level service that ultimately allows for the multiplexing of graphics and input. Nitpicker virtualizes the low-level framebuffer and input interfaces while supplementing the notions of visible views and input focus. Such a system looks as follows:

The drivers and the nitpicker GUI server are displayed in red because they are servers. The GUI applications at the right are nitpicker clients. Nitpicker, in turn, is a client of the framebuffer and input drivers.

The client interface of the nitpicker GUI server is success story. A few evolutionary tweaks notwithstanding, the original design scales up to the windowed working environment of Sculpt OS while retaining the tiny complexity of the nitpicker GUI server.

However, the approach of modelling the framebuffer and input drivers as servers is a dead end, for the following reasons.

  • Drivers are sometimes extremely complex. When looking at the complexity of the Intel framebuffer driver (ported from the Linux kernel) for example, it feels foolish to trust that it won't fail. Because the nitpicker GUI server depends on the liveliness of the driver, however, the well-being of the driver is on the critical path for the entire GUI stack and all GUI applications.

  • It is impossible to transparently restart or replace the driver at runtime without also restarting the GUI stack and - transitively - all GUI applications. This is because the nitpicker GUI server is connected as a client to the drivers. The connections are like a life supply for the GUI server. Cutting those would ultimately kill the GUI server.

  • Adding secondary displays is rather difficult because this would require nitpicker to know what framebuffer servers are out there and to request multiple framebuffer connections. Adding such a protocol would inflate the complexity of nitpicker.

Inverting interfaces

To overcome these limitations, we must break the dependency of the GUI server from the drivers, like illustrated here:

In this scenario, the GUI server no longer uses the drivers, but the drivers use the GUI server. The relationship is reversed. Only the nitpicker GUI server remains red. Of course, the drivers cannot talk to nitpicker GUI server via nitpicker's regular client interface. Instead, two new interfaces enter the picture:

The Capture session interface

is like the inverse of the original Framebuffer session interface. It allows a client to obtain pixel data from the server. From the GUI server's perspective, a framebuffer driver is like a frame capturing device - which it is.

The Event session interface

is like the inverse of the original Input session interface. In contrast to a input session, which allows a client to obtain input events, an event session allows a client to induce input events. Nitpicker is an event server that consumes the input events from the driver(s).

The benefits of this approach are manifold:

  • The low-complexity GUI server remains as the only component that must never fail, which makes the GUI stack much more resilient compared to today.

  • Since the nitpicker GUI server no longer depends on the drivers, it can be started before the drivers are running, speeding up the boot time of graphical scenarios. In scenarios with rigid power management, we can remove the graphics driver - and switch off the graphics device - all while the GUI server and all GUI applications stay intact.

  • Drivers can be restarted anytime. From the perspective of the nitpicker GUI server, this looks like a client disconnect and connect. Drivers can also be replaced at runtime, e.g., swapping out the VESA and Intel drivers.

  • Connecting more than one driver to the nitpicker GUI server becomes conceptually simple. This clears the path towards multi-head support.

  • The new capture session interface of the nitpicker GUI server would be the right hook for implementing screenshot/screencast applications, or a remote desktop server.

  • Analogously to capture devices, input devices can enter and leave the system at any time without disruption, which will allow for the fine-grained management of USB HID devices. Think of plugging a new HID device to your Sculpt system. A new USB device would show up. The user can deploy the matching HID driver with the option of routing the event session of the driver to the system GUI server. Isn't that wonderful?

What is the downside? It is a lot of work! This leads us the following section.

The plan

Turning the low-level session interfaces upside down is an invasive operation that must be carried out in several steps, each yielding an intermediate consistent state.

  1. We start off with some low-risk cosmetic changes, namely the renaming of the "Nitpicker" session interface to "Gui" session. The strange branding of the session interface after a particular implementation is a relic from the early days. This change has already entered Genode's master branch.

  2. Consistently use 32-bit RGB as pixel format by all drivers and applications. This is a long-standing feature request that eliminates the need for color-space conversions and generally improves the output quality. It is quite labor-intensive because it requires me testing many different hardware platforms covering the drivers vesa_fb_drv, fb_sdl, intel_fb_drv, imx53_fb_drv, imx8_fb_drv, rpi_fb_drv, boot_fb_drv, omap4_fb_drv, exynos5_fb_drv. This change has already progressed well.

  3. Introducing the new Capture session interface.

  4. Implementing the capture service in the nitpicker GUI server.

  5. Introduce a nitpicker option to choose the framebuffer back end between the use of a requested framebuffer session (as done today), the use of a capture client for the pixel output, or the use of another GUI server (for stacking multiple instance of nitpicker).

  6. Turn "Framebuffer" drivers into "Capture" clients, covering all the drivers mentioned above. Shoveling code. Adjusting all existing scenarios to use nitpicker's capture service.

  7. Pruning the remaining use of the framebuffer client interface. In fact, when done, I want to remove the Framebuffer::Connection completely. This will require the reworking of the terminal, MESA back end, several tests, and the liquid_framebuffer. In the process, I hope to largely eliminate the use cases of the nit_fb (now called gui_fb) component so that this component can be removed in the end.

  8. Turning a few other framebuffer clients into "Capture" servers, in particular test-framebuffer, fb_bench, and test-driver_manager.

  9. Removal of the "Framebuffer" session interface, retaining the interface merely as a part of the GUI session.

  10. Introducing the new Event session interface and adding its implementation to the nitpicker GUI server.

  11. Adding a new event_filter component taking the place of the input_filter.

  12. With the structural changes done, it would be time for optimizations. As a prerequisite for tear-free animations, the frame buffering will be moved from the clients into the nitpicker GUI server. At first, this will be transparent to most clients. The subsequent removal of client-side buffering is second independent step. As another optimization, I plan to go though all places that employed the dithering of pixels.

Of those topics, the rework of the framebuffer drivers is certainly the most elaborate one. I'm considering to drop the Exynos5 and OMAP4 drivers. Even though I strive for replacing the input session interface by the new event session interface, the scope of the task of reworking the input drivers is not yet completely tangible to me.