Kernel: NVIDIA, Wayland and Intel

-
NVIDIA Fixes Flaws in Linux4Tegra Driver for Jetson AI Supercomputers
NVIDIA released a security update for the Jetson TX1 and TX2 to patch vulnerabilities discovered in the Linux for Tegra (L4T or Linux4Tegra) driver package that could enable local attackers with basic user privileges to elevate privileges and to perform privilege escalation, denial-of-service (DoS) or information disclosure attacks.
As described by NVIDIA, Jetson TX1 and TX2 are high-performance and low-power embedded AI supercomputers on a module designed to be used for compute-intensive deep learning and computer vision projects.
-
RFC: Wayland high-resolution wheel scrolling
I've been trying to sort out the new hi-res wheel scrolling that was added to linux 5.0 but it's been a bit of a struggle to say the least. For the impatient, skip forward to the protocol diff but the background info is important. Everything below applies equally to horizontal and vertical scrolling, both directions and wheels and wheel tilts. The APIs that are involved for low-resolution wheel scrolling: kernel: - REL_WHEEL for every wheel click libinput uses pointer axis events for wheels: - pointer axis value: the movement of the wheel in degrees - pointer axis discrete value: the number of wheel clicks - pointer axis source: set to 'wheel' Usally this means you get discrete 1, value 15 for the standard 15-degree mouse wheel, discrete 2 value 30 is two wheel clicks within one hw frame. There is one implicit assumption here: the discrete value must be 1 or more for every event. Otherwise it's impossible to use the value and calculate the angle. More on that later. This API is effectively the same in the wayland protocol except for the unit of the 'value' which is is in screen coordinates, not degrees. wayland protocol: - pointer axis value: the movement in screen coordinates - pointer axis discrete: the number of wheel clicks - pointer axis source: set to 'wheel' The big difference here though: libinput's API is per-device, wayland's API is per wl-seat pointer. There's no way to tell in the protocol that you're scrolling simultaneously with wheels on two different mice. Let's see how the common compositors implement this: weston: - for source wheel, only the discrete value is handled and: - pointer axis discrete is sent as-is from libinput - pointer axis source is sent as-is from libinput - pointer axis value is always (discrete * 10) - otherwise the libinput pointer value (degrees) from a source wheel is ignored The magic multiplier 10 is afaik for historical reasons, weston used to do 10 so early libinput took that. When libinput switched to use physical dimensions instead of 10, that value was kept for backwards compatibility with existing wayland clients. mutter: - for source wheel, only the discrete value is handled and: - clutter-internal event for the discrete **direction** from libinput - i.e. discrete > 1 will still only generate one event,some events drop to the floor here for fast wheel movements - clutter-internal event with (value * 10) marked as 'emulated' but that event is ignored later when we're writing to the protocol - the discrete event is written as wl_pointer.axis_discrete and wl_pointer.axis with a value of 10 So, the same multiplier of 10 but fast scroll events with more than one click per frame appear to get dropped. kwin (through qtwayland): - doesn't handle discrete at all afaict - pointer axis events are passed through but I got a bit confused over the actual value written on the wire here. wlroots: - pointer axis discrete is sent as-is from libinput - pointer axis source is sent as-is from libinput - pointer axis value is sent as-is from libinput Because the physical angle is passed on from libinput as screen coordinates, this means wlroots has different scroll behaviour to GNOME/sway and it does scroll a larger distance for wheels with higher click angles. This is the opposite of the intended behaviour, manufacturers advertise these wheels as slow-scrolling wheels. Realistically though the physical movements is the same-ish and if a client uses discrete it doesn't matter anyway. xf86-input-libinput: - X devices need to set up a scroll distance for scroll motion, the driver uses 120 (since Feb, before it was 15). This means any multiple of 120.0 triggers legacy button emulation in the server, otherwise the number has no meaning. - for source wheel it uses the discrete value only, so effectively 120 * discrete. This driver had a division by 0 for discrete values of 0, fixed in the current release. More on this later. Let's look at how wheel events are processed in some of the clients: GTK: - for a discrete event, GDK_SCROLL_UP is passed on flagged as 'emulated' - the axis value is passed on as GDK_SCROLL_SMOOTH Qt: - doesn't handle discrete - multiplies the value by -12. This gives it a 120-based value which matches the Windows API provided the axis value is 10 of course. Qt on wlroots is probably off here. Xwayland: - X devices need to set up a scroll distance for scroll motion, Xwayland currently uses 1.0 for that. This means any multiple of 1.0 triggers legacy button emulation in the server. - wl_pointer.axis_discrete are used where available - wl_pointer.axis uses the value * 0.1. This effectively means Xwayland relies on the deltas to be 10 like in mutter/weston, X clients under wlroots will have off wheel click emulation. With me so far? Hooray, now let's introduce hi-res scrolling. All the examples below are for a quarter scroll wheel movement, i.e. you get 4 hi-res events for every 1 lo-res event. The kernel uses a fraction/multiple of 120 for logical clicks, same as the Windows API. So one wheel click is 120, half a click is 60, etc. I'm calling this 120-based value the v120 because I'm very creative. kernel: - REL_WHEEL for every wheel click - REL_WHEEL_HI_RES for every fraction of a wheel click (as portion of 120), so a quarter-wheel stop gives you a value of 30. - there is no guarantee that REL_WHEEL is sent every full 120 because e.g. Logitech mice may reset halfway through a wheel motion. So the two axes must be considered completely independent. libinput: Basically: the current API is fairly useless in handling fractional scroll because libinput uses physical distances. libinput could send four events: - value 3.75, discrete 0 - value 3.75, discrete 0 - value 3.75, discrete 0 - value 3.75, discrete 1 Which is technically correct. But callers have no reliable way how much of a fraction of a wheel the 3.75 represents until the first discrete event comes in. Guessing is unreliable, if you scroll faster you may get 2+ fractional units in one event. And mice may reset halfway through a scroll motion so you don't always get the same number of events per discrete. So libinput needs some extra API that gives us some baseline to compare values against, i.e. same as the kernel's 120-based API does. There are other problems, specifically related to sending discrete 0 events: weston: - Current weston: a discrete value of 0 ends up as value of 0 which weston treats as axis_stop event. So our event sequence is axis_source, axis_stop, pointer_frame. this breaks scrolling. mutter: - Current mutter: a discrete value of 0 ends up in noops. mutter still generates internal clutter events but they are never pushed onto the wire. The normal values are ignored, so we basically get low-res scrolling just as before. Note: this is based on reading the code, not test runs wlroots: - will forward value/discrete as it comes in from libinput kwin: - doesn't use discrete so won't be affected by changes xf86-input-libinput: - anything pre 0.28.2 will get a division by 0 - current state: discrete values of 0 end up as noops and we we get lo-res scrolling only So basically: if we change libinput to send discrete 0 events, we'll break weston and all but the most recent xf86-input-libinput. Code changes are needed just to cope with the different event sequence, let alone the actual hi-res bits. The solution to this is to add a new event, LIBINPUT_EVENT_POINTER_AXIS_WHEEL. That event provides libinput_event_pointer_axis_value_v120() which is basically a mirror of the kernel API. This event obsoletes POINTER_AXIS events for sources WHEEL and WHEEL_TILT, so callers should just ignore those if they support the new event. A branch for this is available here: https://gitlab.freedesktop.org/whot/libinput/commits/wip/hi-res-scrolling The event sequence would now be: - WHEEL: value 3.75, v120 30 - WHEEL: value 3.75, v120 30 - WHEEL: value 3.75, v120 30 - WHEEL: value 3.75, v120 30 - AXIS: value 3.75, discrete 1, v120 120 Implementation-wise: AXIS and WHEEL are independent, so there's no guarantee that they add up to the same 120 values. This new event is easy to add to compositors. Let's look at the Wayland protocol with the current API given hi-res scroll events: - value 3.75, no discrete event - value 3.75, no discrete event - value 3.75, no discrete event - value 3.75, discrete 1 This is assuming 15 degrees/4 == 3.75 but if the compositor force wheels to a scale of 10, the value would be 2.5. Qt: - doesn't use discrete so won't be affected by changes GTK: - generates 4 GDK_SCROLL_SMOOTH events, - generates 1 GDK_SCROLL_DOWN event ('emulated') - GTK doesn't really care as such about whether the distance-to-wheel is 10 or something else, it'll just have different scroll distances (e.g. under wlroots). Afaict this works correctly (tested with stock F29 nautilus) Xwayland: - current Xwayland: values are handled as fraction as before, so we get 3 smooth-scroll events at 25%. Event 4 has a discrete event which is handled as full event, so we scroll by a logical click. Total distance is 175%. This means: Xwayland cannot work at all without any extra wayland protocol unless it assumes factor 10 is the base unit. So we have at least one ubiquitous Wayland client that cannot handle it. And the issues are very similar to the ones libinput's API has, added to that is that you cannot guarantee that two events come from the same device. What we need is a new event, but I'm struggling to add this to the wayland protocol in a good way, mostly for backwards-compatibility reasons. Much of the below would be easier to solve if we can say "if you claim to support wl_pointer version 8, you get different behaviour for axis events". I don't think we can trust clients to handle this correctly, I got burned by those assumptions before (in XI2). So I'm ruling out changing any of the existing protocol behaviour. My current idea is to add a wl_pointer.axis_v120 event: wl_pointer.axis_source wheel wl_pointer.axis_v120 30 wl_pointer.frame wl_pointer.axis_source wheel wl_pointer.axis_v120 30 wl_pointer.frame wl_pointer.axis_source wheel wl_pointer.axis_v120 30 wl_pointer.frame wl_pointer.axis_source wheel wl_pointer.axis 10 wl_pointer.axis_discrete 1 wl_pointer.axis_v120 30 wl_pointer.frame Main problems here: this puts a fair bit of implementation requirements on the compositor - skip source/v120/frame for clients < supported version. wl_pointer.source cannot currently occur on its own, it requires a wl_pointer.axis event in the current protocol. Same with axis_discrete. Empty frames are allowed, I think, but pointless. It also puts a bit of extra logic in the client, you need to remember if you've seen a v120 in this frame and discard the axis/axis_discrete events. If not, then axis events must be processed as normal. A possible workaround would be to add v120 to all axis frames (including touchpad scrolling) but that would then put the compositor in charge of deciding what distance is a wheel click and encodes that information in the protocol. The v120 event must also carry the equivalent to the wl_pointer.axis value so that can be passed on into existing code. So the above would have a v120 event with v120 value of 30 and an axis value of 2.5 (if enforcing 10 units per wheel click). Because the hi-res and lo-res sources are independent, we need to allow for a 0-value v120 event, which basically just tags that frame as The compositor also need to *always* send v120 events for wheel/tilt sources because otherwise the client code becomes even more of a nightmare. Where not backed by libinput the compositor needs to handle that itself. Anyway, the below is the best I've come up with so far though I haven't implemented it client-side yet. I'd really appreciate any feedback and/or epiphanies. Cheers, Peter -
High Resolution Scroll Wheel Support Being Worked On For Wayland
The high resolution scrolling support in Linux 5.0 has been a headache to say the least. After being ejected from Linux 4.20 following early fall-out, the support was merged for Linux 5.0 but the user-space support has yet to stabilize.
Red Hat input expert Peter Hutterer who has been involved in this high-resolution scroll wheel support for the likes of Logitech and Microsoft mouse referred to the situation as "a bit of a struggle to say the least."
-
Intel Wi-Fi 6 AX200 Launches With Linux Support In Tow
The product page does list Linux support that comes as little surprise these days. In fact, back in January we wrote about Linux support for these new Intel adapters within the "IWLWIFI" driver and that enablement is now present in the Linux 5.1 kernel.
-

- Login or register to post comments
Printer-friendly version- 2179 reads
PDF version
More in Tux Machines
- Highlights
- Front Page
- Latest Headlines
- Archive
- Recent comments
- All-Time Popular Stories
- Hot Topics
- New Members
digiKam 7.7.0 is released
After three months of active maintenance and another bug triage, the digiKam team is proud to present version 7.7.0 of its open source digital photo manager. See below the list of most important features coming with this release.
|
Dilution and Misuse of the "Linux" Brand
|
Samsung, Red Hat to Work on Linux Drivers for Future Tech
The metaverse is expected to uproot system design as we know it, and Samsung is one of many hardware vendors re-imagining data center infrastructure in preparation for a parallel 3D world.
Samsung is working on new memory technologies that provide faster bandwidth inside hardware for data to travel between CPUs, storage and other computing resources. The company also announced it was partnering with Red Hat to ensure these technologies have Linux compatibility.
|
today's howtos
|








.svg_.png)
Content (where original) is available under CC-BY-SA, copyrighted by original author/s.

Recent comments
1 year 11 weeks ago
1 year 11 weeks ago
1 year 11 weeks ago
1 year 11 weeks ago
1 year 11 weeks ago
1 year 11 weeks ago
1 year 11 weeks ago
1 year 11 weeks ago
1 year 11 weeks ago
1 year 11 weeks ago