From time to time, we receive regular complaints about frame scheduling. In particular, compositing not being synchronized to vblanks, missed frames, repainting monitors with different refresh rates, etc. This blog post will (hopefully) explain why these issues are present and how we plan to fix them.
Past & Present
With the current scheduling algorithm, compositing /should/ start immediately right after a vblank. A vblank is the time between the vertical front porch and the vertical back porch, or simply put, it’s the time when the display starts scanning out the contents of the next frame.
One thing that’s worth point out is that buffers are not swapped after finishing a compositing cycle, they are swapped at the start of the next compositing cycle, in other words, at the next vblank
KWin assumes that glXSwapBuffers() and eglSwapBuffers() will always block until the next vblank. By delaying the buffer swap, we have more time to process input events, do some window manager things, etc. But, this assumption is outdated, nowadays, it’s rare to see a GLX or an EGL implementation where a buffer swap operation blocks when rendering double buffered.
In case the buffer swap operation doesn’t block, which is typically the case with Mesa drivers, glXSwapBuffers() or eglSwapBuffers() will be called at the end of a compositing cycle. There is a catch though. Compositing won’t be synchronized to vblanks.
Since compositing is not synchronized with vblanks anymore, you may notice that animations in some application don’t look butter smooth as they should. This issue can be easily verified using the black frame insertion test .
Another problem with our compositing scheduling algorithm is latency. Ideally, if you press a key, the corresponding symbol should show up on the screen as soon as possible. In practice, things are slightly different
With the current compositing timing, if you press a key on the keyboard, it may take up to two frame before the corresponding symbol shows up on the screen. Same thing with videos, the audio might be playing two frames ahead of what is on the screen.
Monitors With Different Refresh Rates
Things get trickier if you have several monitors and they have different refresh rates. On X11, compositing is throttled to the lowest common refresh rate, in other words if you have two monitors with a refresh rate of 60Hz and one with a refresh rate of 120Hz, compositing will be performed at a rate of 60Hz. There is probably nothing that we can do about it.
On Wayland, it’s a completely different situation. From the technical point of view, we don’t have anything that prevents compositing being performed separately per screen at different refresh rates. But due to historical reasons, compositing on Wayland is throttled similar to the X11 case.
Our main goals are to unlock true per screen rendering on Wayland and reduce latency caused by compositing (both on X11 and Wayland). Some work  has already been started to fix compositing timing and if things go smoothly, you should be able to enjoy improved frame timings in KDE Plasma 5.21.
If we start compositing as close as possible to the next vblank, then applications, such as video players, will be able to get their contents on the screen in the shortest amount of time without inducing any screen tearing.
The main drawback of this approach is that the compositor has to know how much time exactly it will take to render the next frame. In other words, we need a reliable way to predict the future, easy, no problem!
The main idea behind the compositing timing rework is to introduce a new class, called RenderLoop, that notifies the compositor when it’s a good time to start painting the next frame. On X11, there is going to be only one RenderLoop. On Wayland, every output is going to have its own RenderLoop.
As it was mentioned previously, the compositor needs to predict how long it will take to render the next frame. We solve this inconvenient problem by making two guesses:
The first guess is based on a desired latency level that comes from a config. If the desired latency level is high, the predicted render time will be longer; on the other hand, if the desired latency level is low, the predicted render time will be shorter;
The second guess is based on the duration of previous compositing cycles.
The RenderLoop makes two guesses and the one with the longest render time is used for scheduling compositing for the next frame. By making two estimates rather than one, hopefully, animations will be more or less stable.
There is no “silver bullet” solution for the render time prediction problem, unfortunately. In the end, it all comes down to making a trade-off between latency and stability. The config option lets the user decide what matters the most. It’s worth noting that with the default latency level, the compositor will make a compromise between frame latency and animation stability that should be good enough for most of users.
The introduction of the RenderLoop helper is only half of the battle. At the moment, all compositing is done on the main thread and it can get crowded. For example, if you have several outputs with different refresh rates, some of them will have to wait until it’s their turn to get repainted. This may result in missed vblanks, and thus laggy frames. In order to address this issue, we need to put compositing on different threads. That way, monitors will be repainted independently of each other. There is no concrete milestone for compositing on different threads, but most likely, it’s going to be KDE Plasma 5.22.
Currently, compositing infrastructure in KWin is heavily influenced by the X11 requirements, e.g. there is only one compositing clock, compositing is throttled to the lowest refresh rate, etc. Besides that, incorrect assumptions were made about the behavior of glXSwapBuffers() and eglSwapBuffers(), unfortunately, which result in frame drops and other related issues. With the ongoing Wayland improvements, we hope to fix the aforementioned issues.
If you are a long time Plasma user, you probably remember the times when most GTK applications in KDE Plasma prior to 5.18 were practically unusable due to the lacking support for _GTK_FRAME_EXTENTS. In this blog post, I would like to get a bit technical and walk you through some changes that happened during the 5.18 time frame that made it possible to support _GTK_FRAME_EXTENTS in KWin. I also would like to explain why after so many years of resistance, we had finally added support for client-side decorations. So, buckle up your seat belt!
What is _GTK_FRAME_EXTENTS, anyway?
A window can be viewed as a thing that has some contents and a frame around it with a close button and so on. The “server-side decoration” term is used to refer to a window frame drawn by the window manager (KWin). If the application draws the window frame by itself, then we refer to that window frame as a “client-side decoration.”
A cool thing about client-side decorations is that they are very eye candy, but there is also a lot of drawbacks about them, for example if the application hangs, the user won’t be able to close the window by clicking the close button in the window frame. But the biggest issue with client-side decorations is that the window manager has to know the extents of the client-side drop shadow otherwise things such as window snapping and so on won’t work as desired.
_GTK_FRAME_EXTENTS is a proprietary GTK extension that describes the extents of the client-side decoration on each side (left, right, top, and bottom). From the word “proprietary” you have probably already guessed that _GTK_FRAME_EXTENTS is a thing that is not in any spec. We can’t afford implementing a proprietary extension simply because we don’t know whether re-designing KWin will pay off in the end. What if GTK ditches _GTK_FRAME_EXTENTS for something else and our hard work will be for nothing? There were some suggestions to standardize _GTK_FRAME_EXTENTS in the NETWM spec, but it didn’t go well.
So, what did change our minds?
It might come as a surprise, but the reason why we decided to add CSD support after so many years of being reluctant was Wayland. In order to fully implement the xdg-shell protocol (the de-facto protocol for creating desktop-style surfaces), we must support client-side decorated windows. Prior to that, we didn’t have any reason that could possibly somehow justify the changes that we would have to make in code that was battle-tested for many years.
With Wayland, we know what changes have to be done in order to add support for client-side decorated clients. Surprisingly, the geometry abstractions that we chose specifically for client-side decorated Wayland clients turned out to be also pretty good for client-side decorated X11 clients, so we decided to add support for _GTK_FRAME_EXTENTS since it didn’t require any huge changes.
It still means that we will be screwed if GTK switches to something completely different on X11, though. But let’s hope that it won’t happen.
CSD and KDE Plasma
“But Vlad,” you may say. “Does this mean that KDE is going to switch CSD?” No, as far as I know, nothing has changed, we still use server-side decorations. Support for client-side decorations was added because it’s something that we need on Wayland and to make GTK applications usable on X11.
Frame, buffer, and client geometry
Warning: This is a brain dump. Please skip to the next section if you’re not interested in technical stuff.
For an old school window manager such as KWin, client-side decorated windows are troublesome mainly because due to the long history all rendering related code had been written with the assumption that the window frame wraps the window contents. If an application draws the window frame on its own, that’s not the case.
In 5.18, we tackled that problem by introducing two new geometries to separate window management from rendering – the frame geometry and the buffer geometry.
The frame geometry describes a rectangle that bounds the window frame. It doesn’t matter whether the window is client-side decorated or server-side decorated. This kind of geometry is used practically for everything, for example window snapping, resizing, etc. KWin scripts see and operate on this geometry.
The buffer geometry is used primarily during rendering, for example to build window quads, etc.
In 5.20, we introduced yet another new geometry, which existed prior to that in an implicit form – the client geometry. The client geometry indicates where the window contents inside the window frame  is on the screen. We use this geometry primarily for configuring windows.
It can be a bit challenging to deal with three different geometries at the same time, but there is not that much we can do about it, unfortunately. Each geometry has its own specific domain where the other geometries are inappropriate to use.
CSD is a rather controversial subject in the KDE community, but it’s here and it’s not going anywhere, anytime soon.
 On X11, the client geometry actually corresponds to the geometry of the client window.