libinput - a common input stack for Wayland compositors and X.Org drivers

Last November, Jonas Ådahl sent an RFC to the wayland-devel list about a common library to handle input devices in Wayland compositors called libinput. Fast-forward and we are now at libinput 0.6, with a broad support of devices and features. In this post I'll give an overview on libinput and why it is necessary in the first place. Unsuprisingly I'll be focusing on Linux, for other systems please mentally add the required asterisks, footnotes and handwaving.

The input stack in X.org

The input stack as it currently works in X.org is a bit of a mess. I'm not even talking about the different protocol versions that we need to support and that are partially incompatible with each other (core, XI 1.x, XI2, XKB, etc.), I'm talking about the backend infrastructure. Let's have a look:

The graph above is a simplification of the input stack, focusing on the various high-level functionalities. The X server uses some device discovery mechanism (udev now, previously hal) and matches each device with an input driver (evdev, synaptics, wacom, ...) based on the configuration snippets (see your local /usr/share/X11/xorg.conf.d/ directory).

The idea of having multiple drivers for different hardware was great when hardware still mattered, but these days we only speak evdev and the drivers that are hardware-specific are racing the dodos to the finishing line.

The drivers can communicate with the server through the very limited xf86 DDX API, but there is no good mechanism to communicate between the drivers. It's possible, just not doable in a sane manner. Most drivers support multiple X server releases so any communication between drivers would need to take version number mixes into account. Only the server knows the drivers that are loaded, through a couple of structs. Knowledge of things like "which device is running synaptics" is obtainable, but not sensibly. Likewise, drivers can get to the drivers of other devices, but not reasonably so (the API is non-opaque, so you can get to anything if you find the matching header).

Some features are provided by the X server: pointer acceleration, disabling a device, mapping a device to monitor, etc. Most other features such as tapping, two-finger scrolling, etc. are provided by the driver. That leads to an interesting feature support matrix: synaptics and wacom both provide two-finger scrolling, but only synaptics has edge-scrolling. evdev and wacom have calibration support but they're incompatible configuration options, etc. The server in general has no idea what feature is being used, all it sees is button, motion and key events.

The general result of this separation is that of a big family gathering. It looks like a big happy family at first, but then you see that synaptics won't talk to evdev because of the tapping incident a couple of years back, mouse and keyboard are have no idea what forks and knives are for, wacom is the hippy GPL cousin that doesn't even live in the same state and no-one quite knows why elographics keeps getting invited. The X server tries to keep the peace by just generally getting in the way of everyone so no-one can argue for too long. You step back, shrug apologetically and say "well, that's just how these things are, right?"

To give you one example, and I really wish this was a joke: The server is responsible for button mappings, tapping is implemented in synaptics. In order to support left-handed touchpads, gnome-settings-daemon sets the button mappings on the device in the server. Then it has to swap the tapping actions from left/right/middle for 1/2/3-finger tap to right/left/middle. That way synaptics submits right button clicks for a one-finger tap, which is then swapped back by the server to a left click.

The X.org input drivers are almost impossible to test. synaptics has (quick guesstimate done with grep and wc) around 70 user-configurable options. Testing all combinations would be something around the order of 10¹⁰¹ combinations, not accounting for HW differences. Testing the driver on it's own is not doable, you need to fire up an X server and then run various tests against that (see XIT). But now you're not testing the driver, you're testing the whole stack. And you can only get to the driver through the X protocol, and that is 2 APIs away from the driver core. Plus, test results get hard to evaluate as different modules update separately.

So in summary, in the current stack features are distributed across modules that don't communicate with each other. The stack is impossible to test, partially thanks to the vast array of user-exposed options. These are largely technical issues, we control the xf86 DDX API and can break it when needed to, but at this point you're looking at something that resembles a rewrite anyway. And of course, don't you dare change my workflow!

The input stack in Wayland

From the input stack's POV, Wayland simply merges the X server and the input modules into one item. See the architecture diagram from the wayland docs:

evdev gets fed into the compositor and wayland comes out the other end. If life were so simple... Due to the X.org input modules being inseparable from the X server, Weston and other compositors started implementing their own input stack, separately. Let me introduce a game called feature bingo: guess which feature is currently not working in $COMPOSITOR. If you collect all five in a row, you get to shout "FFS!" and win the price of staying up all night fixing your touchpad. As much fun as that can be, maybe let's not do that.

libinput

libinput provides a full input stack to compositors. It does device discovery over udev and event processing and simply provides the compositor with the pre-processed events. If one of the devices is a touchpad, libinput will handle tapping, two-finger scrolling, etc. All the compositor needs to worry about is moving the visible cursor, selecting the client and converting the events into wayland protocol. The stack thus looks something like this:

Almost everything has moved into libinput, including device discovery and pointer acceleration. libinput has internal backends for pointers, touchpads, tablets, etc. but they are not exposed to the compositor. More importantly, libinput knows about all devices (within a seat), so cross-device communication is possible but invisible to the compositor. The compositor still does configuration parsing, but only for user-specific options such as whether to enable tapping or not. And it doesn't handle the actual feature, it simply tells libinput to enable or disable it.

The graph above also shows another important thing: libinput provides an API to the compositor. libinput is not "wayland-y", it doesn't care about the Wayland protocol, it's simply an input stack. Which means it can be used as base for an X.org input driver or even Canonical's MIR.

libinput is very much a black box, at least compared to X input drivers (remember those 70 options in synaptics?). The X mantra of "mechanism, not policy" allows for interesting use-cases, but it makes the default 90% use-case extremely painful from a maintainer's and integrator's point of view. libinput, much like wayland itself, is a lot more restrictive in what it allows, specifically in the requirement it places on the compositor. At the same time aims for a better out-of-the-box experience.

To give you an example, the X.org synaptics driver lets you arrange the software buttons more-or-less freely on the touchpad. The default placement is simply a config snippet. In libinput, the software buttons are always at the bottom of the touchpad and also at the top of the touchpad on some models (Lenovo *40 series, mainly). The buttons are of a fixed size (which we decided on after analysing usage data), and you only get a left and right button. The top software buttons have left/middle/right matching the markings on the touchpad. The whole configuration is decided based on the hardware. The compositor/user don't have to enable them, they are simply there when needed.

That may sound restrictive, but we have a number of features on top that we can enable where required. Pressing both left and right software buttons simultaneously gives you a middle button click; the middle button in the top software button row provides wheel emulation for the trackstick. When the touchpad is disabled, the top buttons continue to work and they even grow larger to make them easier to hit. The touchpad can be set to auto-disable whenever an external mouse is plugged in.

And the advantage of having libinput as a generic stack also results in us having tests. So we know what the interactions are between software buttons and tapping, we don't have to wait for a user to trip over a bug to tell us what's broken.

Summary

We need a new input stack for Wayland, simply because the design of compositors in a Wayland world is different. We can't use the current modules of X.org for a number of technical reasons, and the amount of work it would require to get those up to scratch and usable is equivalent to a rewrite.

libinput now provides that input stack. It's the default in Weston at the time of this writing and used in other compositors or in the process of being used. It abstracts most of input away and more importantly makes input consistent across all compositors.

Stylus behaviour on Microsoft Surface 3 tablets

Note: The purpose of this post is basically just so we have a link when this comes up in future bugreports.

Some stylus devices have two buttons on the stylus, plus the tip itself which acts as a button. In the kernel, these two are forwarded to userspace as BTN_STYLUS and BTN_STYLUS2. Userspace then usually maps those two into right and middle click, depending on your configuration. The pen itself used BTN_TOOL_PEN when it goes into proximity.

The default stylus that comes with the Wacom Intuos Pro [1] has an eraser on the other side of the pen. If you turn the pen around it goes out of proximity and comes back in as BTN_TOOL_RUBBER. [2] In the wacom X driver we handle this accordingly, through two different devices available via the X Input Extension. For example the GIMP assigns different tools assigned to each device. [3]

In the HID spec there are a couple of different fields (In Range, Tip Switch, Barrel Switch, Eraser and Invert) that matter here. Barrel Switch is the stylus button, In Range and Tip Switch are proximity and "touching the surface". Invert signals which side of the pen points down, and Eraser is triggered when the eraser touches the surface. In Wacom tablets, Invert is always on when the eraser touches because that's how the pens designed.

Microsoft, in its {in}finite wisdom has decided to make the lower button an "eraser" button on its Surface 3 Pen. So what happens now is that once you press the button, In Range goes to zero, then to one in the next event, together with Invert. Eraser comes on once you touch the surface but curiously that also causes Invert to go off. Anyway, that's a low-level detail that will get handled. What matters to users is that on the press of that button, the pen goes virtually out of proximity, comes back in as eraser and then hooray, you can now use it as an eraser tool without having actually moved it. Of course, since the button controls the mode it doesn't actually work as button, you're left with the second button on the stylus only.

Now, the important thing here is: that's the behaviour you get if you have one of these devices. We could work around this in software by detecting the mode button, flipping bits here and there and trying to emulate a stylus button based on the mode switches. But we won't. The overlords have decreed and it's too much effort to hack around the intended behaviour for little gain.

[1] If marketing decides to rename products so that need a statement "Bamboo pen tablets are now Intuos. Intuos5 is now Intuos Pro." then you've probably screwed up.
[2] Isn't it nice to see some proper queen's English for a change? For those of you on the other side of some ocean: eraser.
[3] GIMP, rubber tool, I'm not making this up, seriously

pointer acceleration in libinput - a userstudy

This post is a summary, the full writeup with data is here (PDF). The texts are quite similar, so if you plan to read the paper you can skip this post.

Following pointer acceleration in libinput - an analysis, we decided to run an actual userstudy to gather some data on how our acceleration behaves, and - more importantly - test if a modified acceleration method is better.

We developed two new pointer acceleration methods (plus the one already in libinput). As explained previously, the pointer acceleration method is a function mapping input speed of a device into cursor speed in pixels. The faster one moves the mouse, the further the cursor moves per "mickey" (a 1-device unit movement). In a simplest example, input deltas of 1 may result in a 1 pixel movement, input deltas of 10 may result in a 30 pixel movement.

The three pointer acceleration methods used in this study were nicknamed:

smooth:

A shortening of "smooth and simple", this method is used in libinput 0.5 as well as in the X.Org stack since ~2008.

stretched:

a modification of 'smooth' with roughly the same profile, but the maximum acceleration is applied at a higher speed. This method was developed by Hans de Goede and very promising in personal testing.

linear:

a linear acceleration method with a roughly similar speed-to-acceleration profile as the first two. This method was developed to test if a simple function could achieve similar results, as the more complex "smooth" and "stretched" methods.

The input data expected by all three methods is in units/ms. Touchpad devices are normalised to 400 dpi, other devices are left as-is. It is impossible to detect in software what resolution a generic mouse supports, so any acceleration method differs between devices. This is intended by the manufacturer, high-resolution devices are sold as "faster" for this reason.

The three pointer acceleration methods

As the graph shows, the base profile is roughly identical and the main difference is how quickly the maximum acceleration factor is reached.

Study description

Central component was a tool built on libinput that displays a full-screen white window, with a round green target. Participants were prompted by GTK dialog boxes on the steps to take next. Otherwise the study was unsupervised and self-guided.

The task required participants to click on a round target with a radius of 15, 30 and 45 pixels. Targets were grouped, each "set" consisted of 15 targets of the same size. On a successful click within the target, a new target appeared on one out of 12 possible locations, arranged in a grid of 4x3 with grid points 300 pixels apart. The location of the target was randomly selected but was never on the same location twice in a row.

Screenshot of the study tool with the first target (size 45) visible.

Each participant was tested for two acceleration methods, each acceleration method had 6 sets of 15 targets (2 sets per target size, order randomised). The two acceleration methods were randomly selected on startup, throughout the study they were simply referred to as "first" and "second" acceleration method with no further detail provided. Acceleration changed after 6 sets (participants were informed about it), and on completion of all 12 sets participants had to fill out a questionnaire and upload the data.

Statistical concepts

A short foray into statistics to help explain the numbers below. This isn't a full statistics course, I'm just aiming to explain the various definitions used below.

The mean of a dataset is what many people call the average: all values added up divided by the number of values. As a statistical tool, the mean is easy to calculate but is greatly affected by outliers. For skewed datasets the median is be more helpful: the middle value of the data array (array[len/2]). The closer the mean and the median are together, the more symmetrical the distribution is.

The standard deviation (SD) describes how far the data points spread from the median. The smaller the SD, the closer together are the data points. The SD is also used to estimate causality vs randomly induced sampling errors. Generally, if the difference between two items is more than 2 standard deviations, there's a 95% confidence that this is a true effect, not just randomness (95% certainty is a widely accepted standard in this domain). That 95% directly maps to the p-value you may have seen in other studies. A p-value of less than 0.05 equals a less than 5% chance of random factors causing the data differences. That translates into "statistically significant".

The ANOVA method is a standard statistical tool for studies like ours. (we're using one-way ANOVA only here, Wikipedia has an example here). If multiple sets of samples differ in only a single factor (e.g. pointer acceleration method), we start with the so-called Null-Hypothesis of "the factor has no influence, all results are the same on average". Our goal is to reject that hypothesis so we can say that the factor did actually change things. If we cannot reject the Null-Hypothesis, either our factor didn't change anything or the results are caused by random influences. The tools for ANOVA compare the mean value within each sets to the mean value differences across the sets and spit out a p-value. As above, a p-value less than 0.05 means greater than 95% confidence that the Null-Hypothesis can be rejected, i.e. we can say our factor did cause those differences.

One peculiarity of ANOVA is that the sample sets have to be the same size. This affects our samples, more see below.

Study participants

An email was sent to three Red Hat-internal lists with a link to the study description. One list was a specific developer list, the other two list were generic lists. As Red Hat employees, participants are expected to be familiar with Linux-based operating systems and the majority is more technical than the average user. The data collected does not make it possible to identify who took part in the study beyond the information provided in the questionnaire.

44 participants submitted results, 7 left-handed, 37 right-handed (no ambidextrous option was provded in the questionnaire). Gender distribution was 38 male, 6 female. Mean age was 33.3 years (SD 6.7) and participants had an mean 21.2 years of experience with mouse-like input devices (SD 4.9) and used those devices an average 58.1 hours per week (SD 20.0).

As all participants are familiar with Linux systems and thus exposed to the smooth acceleration method on their workstations, we expect a bias towards the smooth acceleration method.

Study data

Data was manually checked and verified, three result files were discarded for bugs or as extreme outliers, leaving us with 41 data files. The distribution of methods in these sets was: 27 for smooth, 25 for stretched and 30 for linear.

The base measurement was the so-called "Index of Difficulty" (ID), the number obtained by distance-to-target/width-of-target. This index gives an indication on how difficult it is to hit the target; a large target very close is easier to hit than a small target that is some distance away.

Illustration of the Index of Difficulty for a target.

In hindsight, the study was not ideally suited for evaluation based on ID. The targets were aligned on a grid and the ID based on the pointer position was very variable. As is visible in the graph below, there are few clear dividing lines to categorise the targets based on their ID. For the evaluation the targets were grouped into specific ID groups: ID < 4.2, ID < 8.4, ID < 12.9, ID < 16.9 < ID < 25 and ID > 25. The numbers were selected simply because there are clear gaps between the ID clusters. This division results in uneven group sizes, (I ran the same calculations with different group numbers, it does not have any real impact on the results.)

ID for each target with the divider lines shown

The top ID was 36.44, corresponding to a 15px radius target 1093 pixels away, the lowest ID was 1.45, corresponding to a 45px radius target 130 pixels away.

Number of targets per ID group

As said above, ANOVA requires equal-sized sample sets. ANOVA was performed separately between the methods (i.e. smooth vs stretched, then smooth vs linear, then stretched vs linear). Before each analysis, the two data arrays were cut to be of equal length. For example, comparing smooth and stretched in the ID max group shortened the smooth dataset to 150 elements. The order of targets was randomised.

Study Results

The following factors were analysed:

• Time to click on target
• Movement efficiency
• Overshoot

Time to click on target

Time to click on a target was measured as the time between displaying the target and clicking on it. This does not take reaction time into account, but there is no reliable way of measuring reaction time in this setup.

Mean time to click on target

As is visible, an increasing ID increases the time-to-click. On a quick glance, we can see that the smooth method is slower than the other two in most ID groups, with linear and stretched being fairly close together. However, the differences are only statistically significant in the following cases:

• ID 8.4: linear is faster than smooth and stretched
• ID 12.9: linear and stretched are faster than smooth
• ID 25: linear and stretched are faster than smooth

In all other combinations, there is no statistically significant difference between the three methods, but overall a slight advantage for the two methods stretched and linear.

Efficiency of movement

The most efficient path from the cursor position to the target is a straight line. However, most movements do not follow that straight line for a number of reasons. One of these reasons is basic anatomy - it is really hard to move a mouse in a straight line due to the rotary action of our wrists. Other reasons may be deficiencies in the pointer acceleration method. To measure the efficiency, we calculated the distance to the target (i.e. the straight line) and compared that to all the deltas added up to the total movement. Note that the distance is to the center of the target, whereas the actual movement may be to any point in the target. So for short distances and large targets, there is a chance that a movement may be less than the distance to the target.

Straight distance to target vs. movement path shows the efficiency of movement.

The efficiency was calculated as movement-path/distance, then normalised to a percent value. A value of 10 thus means the movement path was 10% longer than the straight line to the target centre).

Extra distance covered

Stretched seems to perform better than smooth and linear in all but one ID group and smooth performing worse than linear in all but ID group 4.2. Looking at the actual values however shows that the large standard deviation prevents statistical significance. The differences are only statistically significant in the following cases:

• ID 4.2: stretched is more efficient than smooth and linear

In all other combinations, there is no statistically significant difference between the three methods.

Overshoot

Somewhat similar to the efficiency of movement, the overshoot is the distance the pointer has moved past the target. It was calculated by drawing a line perpendicular to the direct path from the pointer position to the target's far side. If the pointer moves past this line, the user has overshot the target. The maximum distance between the line and the pointer shows how much the user has overshot the target.

Illustration of pointer overshooting the target.
The red line shows the amount the pointer has overshot the target.

Overshoot was calculated in pixels, as % of the distance and as % of the actual path taken. Unsurprisingly, the graphs look rather the same so I'll only put one up here.

Overshoot in pixels by ID group

As the ID increases, the amount of overshooting increases too. Again the three pointer acceleration methods are largely the same, though linear seems to be slightly less affected by overshoot than smooth and stretched. The differences are only statistically significant in the following cases:

• ID 4.2: if measured as percentage of distance, stretched has less overshoot than linear.
• ID 8.4: if measured as percentage of movement path, linear has less overshoot than smooth.
• ID 16.8: if measured as percentage of distance, stretched and linear have less overshoot than smooth.
• ID 16.8: if measured as percentage of distance, linear has less overshoot than smooth.
• ID 16.8: if measured in pixels, linear has less overshoot than smooth.

In all other combinations, there is no statistically significant difference between the three methods.

Summary

In summary, there is not a lot of difference between the three methods, though smooth has no significant advantage in any of the measurements. The race between stretched and linear is mostly undecided.

Questionnaire results

The above data was objectively measured. Equally important is the subjective feel of each acceleration method. At the end of the study, the following 14 questions were asked of each participant, with answer ranges in a 5-point Likert scale, ranging from "Strongly Disagree" to "Strongly Agree".

1. The first acceleration method felt natural
2. The first acceleration method allowed for precise pointer control
3. The first acceleration method allowed for fast pointer movement
4. The first acceleration method made it easy to hit the targets
5. I would prefer the first acceleration method to be faster
6. I would prefer the first acceleration method to be slower
7. The second acceleration method felt natural
8. The second acceleration method allowed for precise pointer control
9. The second acceleration method allowed for fast pointer movement
10. The second acceleration method made it easy to hit the targets
11. I would prefer the second acceleration method to be faster
12. I would prefer the second acceleration method to be slower
13. The two acceleration methods felt different
14. The first acceleration method was preferable over the second

The figure below shows that comparatively few "strongly agree" and "strongly disagree" answers were given, hinting that the differences between the methods were small.

Distribution of answers in the questionnaire

Looking at statistical significance, the questionnaire didn't really provide anything of value. Not even the question "The two acceleration methods felt different" provided any answers, and the question "The first acceleration method was preferable over the second" was likewise inconclusive. So the summary of the questionnaire is pretty much: on the whole none of the methods stood out as better or worse.

Likert frequencies for the question of which method is preferable

Summary

Subjective data was inconclusive, but the objective data goes slightly in favour of linear and stretched over the current smooth method. We didn't have enough sample sets to analyse separately for each device type, so from a maintainer's point of view the vote goes to linear. It allows replacing a rather complicated pointer acceleration method with 3 lines of code.