Why is my device a touchpad and a mouse and a keyboard?

If you have spent any time around HID devices under Linux (for example if you are an avid mouse, touchpad or keyboard user) then you may have noticed that your single physical device actually shows up as multiple device nodes (for free! and nothing happens for free these days!). If you haven't noticed this, run libinput record and you may be part of the lucky roughly 50% who get free extra event nodes.

The pattern is always the same. Assuming you have a device named FooBar ExceptionalDog 2000 AI[1] what you will see are multiple devices

/dev/input/event0: FooBar ExceptionalDog 2000 AI Mouse
/dev/input/event1: FooBar ExceptionalDog 2000 AI Keybard 
/dev/input/event2: FooBar ExceptionalDog 2000 AI Consumer Control 

The Mouse/Keyboard/Consumer Control/... suffixes are a quirk of the kernel's HID implementation which splits out a device based on the Application Collection. [2]

A HID report descriptor may use collections to group things together. A "Physical Collection" indicates "these things are (on) the same physical thingy". A "Logical Collection" indicates "these things belong together". And you can of course nest these things near-indefinitely so e.g. a logical collection inside a physical collection is a common thing.

An "Application Collection" is a high-level abstractions to group something together so it can be detected by software. The "something" is defined by the HID usage for this collection. For example, you'll never guess what this device might be based on the hid-recorder output:

# 0x05, 0x01,                    // Usage Page (Generic Desktop)              0
# 0x09, 0x06,                    // Usage (Keyboard)                          2
# 0xa1, 0x01,                    // Collection (Application)                  4
...
# 0xc0,                          // End Collection                            74

Yep, it's a keyboard. Pop the champagne[3] and hooray, you deserve it.

The kernel, ever eager to help, takes top-level application collections (i.e. those not inside another collection) and applies a usage-specific suffix to the device. For the above Generic Desktop/Keyboard usage you get "Keyboard", the other ones currently supported are "Keypad" and "Mouse" as well as the slightly more niche "System Control", "Consumer Control" and "Wireless Radio Control" and "System Multi Axis". In the Digitizer usage page we have "Stylus", "Pen", "Touchscreen" and "Touchpad". Any other Application Collection is currently unsuffixed (though see [2] again, e.g. the hid-uclogic driver uses "Touch Strip" and other suffixes).

This suffix is necessary because the kernel also splits out the data sent within each collection as separate evdev event node. Since HID is (mostly) hidden from userspace this makes it much easier for userspace to identify different devices because you can look at a event node and say "well, it has buttons and x/y, so must be a mouse" (this is exactly what udev does when applying the various ID_INPUT properties, with varying levels of success).

The side effect of this however is that your device may show up as multiple devices and most of those extra devices will never send events. Sometimes that is due to the device supporting multiple modes (e.g. a touchpad may by default emulate a mouse for backwards compatibility but once the kernel toggles it to touchpad mode the mouse feature is mute). Sometimes it's just laziness when vendors re-use the same firmware and leave unused bits in place.

It's largely a cosmetic problem only, e.g. libinput treats every event node as individual device and if there is a device that never sends events it won't affect the other event nodes. It can cause user confusion though: "why does my laptop say there's a mouse?" and in some cases it can cause functional degradation - the two I can immediately recall are udev detecting the mouse node of a touchpad as pointing stick (because i2c mice aren't a thing), hence the pointing stick configuration may show up in unexpected places. And fake mouse devices prevent features like "disable touchpad if a mouse is plugged in" from working correctly. At the moment we don't have a good solution for detecting these fake devices - short of shipping giant databases with product-specific entries we cannot easily detect which device is fake. After all, a Keyboard node on a gaming mouse may only send events if the user configured the firmware to send keyboard events, and the same is true for a Mouse node on a gaming keyboard.

So for now, the only solution to those is a per-user udev rule to ignore a device. If we ever figure out a better fix, expect to find a gloating blog post in this very space.

[1] input device naming is typically bonkers, so I'm just sticking with precedence here
[2] if there's a custom kernel driver this may not apply and there are quirks to change this so this isn't true for all devices
[3] or sparkling wine, let's not be regionist here

High resolution wheel scrolling on Linux v4.21

Disclaimer: this is pending for v4.21 and thus not yet in any kernel release.

Most wheel mice have a physical feature to stop the wheel from spinning freely. That feature is called detents, notches, wheel clicks, stops, or something like that. On your average mouse that is 24 wheel clicks per full rotation, resulting in the wheel rotating by 15 degrees before its motion is arrested. On some other mice that angle is 18 degrees, so you get 20 clicks per full rotation.

Of course, the world wouldn't be complete without fancy hardware features. Over the last 10 or so years devices have added free-wheeling scroll wheels or scroll wheels without distinct stops. In many cases wheel behaviour can be configured on the device, e.g. with Logitech's HID++ protocol. A few weeks back, Harry Cutts from the chromium team sent patches to enable Logitech high-resolution wheel scrolling in the kernel. Succinctly, these patches added another axis next to the existing REL_WHEEL named REL_WHEEL_HI_RES. Where available, the latter axis would provide finer-grained scroll information than the click-by-click REL_WHEEL. At the same time I accidentally stumbled across the documentation for the HID Resolution Multiplier Feature. A few patch revisions later and we now have everything queued up for v4.21. Below is a summary of the new behaviour.

The kernel will continue to provide REL_WHEEL as axis for "wheel clicks", just as before. This axis provides the logical wheel clicks, (almost) nothing changes here. In addition, a REL_WHEEL_HI_RES axis is available which allows for finer-grained resolution. On this axis, the magic value 120 represents one logical traditional wheel click but a device may send a fraction of 120 for a smaller motion. Userspace can either accumulate the values until it hits a full 120 for one wheel click or it can scroll by a few pixels on each event for a smoother experience. The same principle is applied to REL_HWHEEL and REL_HWHEEL_HI_RES for horizontal scroll wheels (which these days is just tilting the wheel). The REL_WHEEL axis is now emulated by the kernel and simply sent out whenever we have accumulated 120.

Important to note: REL_WHEEL and REL_HWHEEL are now legacy axes and should be ignored by code handling the respective high-resolution version.

The magic value of 120 is taken directly from Windows. That value was chosen because it has a good number of integer factors, so dividing 120 by whatever multiplier the mouse uses gives you a integer fraction of 120. And because HW manufacturers want it to work on Windows, we can rely on them doing it right, provided we use the same approach.

There are two implementations that matter. Harry's patches enable the high-resolution scrolling on Logitech mice which seem to mostly have a multiplier of 8 (i.e. REL_WHEEL_HI_RES will send eight events with a value of 15 before REL_WHEEL sends 1 click). There are some interesting side-effects with e.g. the MX Anywhere 2S. In high-resolution mode with a multiplier of 8, a single wheel movement does not always give us 8 events, the firmware does its own magic here. So we have some emulation code in place with the goal of making the REL_WHEEL event happen on the mid-point of a wheel click motion. The exact point can shift a bit when the device sends 7 events instead of 8 so we have a few extra bits in place to reset after timeouts and direction changes to make sure the wheel behaviour is as consistent as possible.

The second implementation is for the generic HID protocol. This was all added for Windows Vista, so we're only about a decade behind here. Microsoft got the Resolution Multiplier feature into the official HID documentation (possibly in the hope that other HW manufacturers implement it which afaict didn't happen). This feature effectively provides a fixed value multiplier that the device applies in hardware when enabled. It's basically the same as the Logitech one except it's set through a HID feature instead of a vendor-specific protocol. On the devices tested so far (all Microsoft mice because no-one else seems to implement this) the multipliers vary a bit, ranging from 4 to 12. And the exact behaviour varies too. One mouse behaves correctly (Microsoft Comfort Optical Mouse 3000) and sends more events than before. Other mice just send the multiplied value instead of the normal value, so nothing really changes. And at least one mouse (Microsoft Sculpt Ergonomic) sends the tilt-wheel values more frequently and with a higher value. So instead of one event with value 1 every X ms, we now get an event with value 3 every X/4 ms. The mice tested do not drop events like the Logitech mice do, so we don't need fancy emulation code here. Either way, we map this into the 120 range correctly now, so userspace gets to benefit.

As mentioned above, the Resolution Multiplier HID feature was introduced for Windows Vista which is... not the most recent release. I have a strong suspicion that Microsoft dumped this feature as well, the most recent set of mice I have access to don't provide the feature anymore (they have vendor-private protocols that we don't know about instead). So the takeaway for all this is: if you have a Logitech mouse, you'll get higher-resolution scrolling on v4.21. If you have a Microsoft mouse a few years old, you may get high-resolution wheel scrolling if the device supports it. Any other vendor or a new Microsoft mouse, you don't get it.

Coincidentally, if you know anyone at Microsoft who can provide me with the specs for their custom protocol, I'd appreciate it. We'd love to have support for it both in libratbag and in the kernel. Or any other vendor, come to think of it.