Re: [RFC] linux-input alps - loosing precision

[Peter Hutterer <peter.hutterer@who-t.net>]

On Tue, Jan 21, 2020 at 01:34:41PM -0800, dmitry.torokhov@gmail.com wrote:
> Hi Pali,
> 
> On Fri, Jan 17, 2020 at 10:42:06AM +0100, Pali Rohár wrote:
> > On Thursday 09 January 2020 14:26:16 Leutwein Tobias (BEG/ENC1) wrote:
> > > Hello,
> > > 
> > > At the file
> > > https://github.com/torvalds/linux/blob/master/drivers/input/mouse/alps.c
> > > I've seen that values coming from the trackpoint/trackstick are divided at some devices, which results in a loss of precision.
> > > 
> > > As I was not lucky with the behavior of the trackpoint of my computer I've made a fork of libinput https://gitlab.freedesktop.org/leutwe/libinput/tree/master/
> > > where I changed  src/filter-trackpoint.c . With this change, the values from lib evdev are multiplied by a factor (dpToFac_au16 []). The array element used from dpToFac_au16 [] is also the value received by evdev.
> > > 
> > > At higher forces I use a factor much higher than 1, so the undivided value from the device would be the best for me.
> > > 
> > > In order not to change the behavior for other users, it might be possible to change the divider at runtime via the sys- file system - like it is at trackpoint.c and take the currently used divider as default value.
> > > 
> > > Positions at source code were I see the division:
> > > 
> > > alps_process_trackstick_packet_v3()
> > > /*
> > > * The x and y values tend to be quite large, and when used
> > > * alone the trackstick is difficult to use. Scale them down
> > > * to compensate.
> > > */
> > > x /= 8;
> > > y /= 8;
> > > 
> > > alps_process_packet_v6()
> > > /* Divide 4 since trackpoint's speed is too fast */ input_report_rel(dev2, REL_X, (char)x / 4); input_report_rel(dev2, REL_Y, -((char)y / 4));
> > 
> > Hello Dmitry!
> > 
> > This problem is about fact that alps.c for some ALPS devices already
> > truncate precision of trackpoint movement.
> > 
> > Tobias is unhappy with this fact and would like to get all bits of
> > movement information, not just some high.
> > 
> > But existing userspace application expects that for these devices kernel
> > already truncate precision and process truncated data from kernel.
> > 
> > Now question is: Are we able to send to userspace input data without
> > doing truncation and without breaking existing applications?
> > 
> > It looks like that for such thing we need some userspace <--> kernel API
> > which disable truncating of precision.
> > 
> > Have you been solving similar issue for other other drivers or in other
> > area of input code?
> 
> Unfortunately I do not think we can automatically "recover" the lost
> precision without help of libinput, which would need to tweak the
> trackpoint [sysfs] property letting kernel know that it should send
> original data.

Right, so the problem here is simply: we don't have write access to sysfs
and most of libinput's users don't either. For evdev devices it's fine
thanks to logind but that won't apply for anything else. Which means that
sysfs is no-go unless you want to shell-script your way around it.

It's been a while since I looked at the trackpoint code but from what I
remember it's a historical mess of inter-dependencies. The kernel driver did
something, then userspace adjusted based on that, then the kernel driver
couldn't fix anything because of userspace, and that looped 3-4 times.

libinput has the magic trackpoint multiplier quirk to work around this
deadlock but multiplying doesn't give you precision back. As an immediate
brain fart, maybe we need a "subpixel" REL_X_FRACTION axis?
Not ideal, since every relative device other than trackpoints just pass on
device units and rely on userspace to make sense of it, so the trackpoints
will likely stay the only users of it.

Alternatively - add the sysfs file, make libinput read it and adjust its
behaviour and then eventually toggle the default. Less breakage that way but
now you're talking about a multi-year inertia. And you rely on userspace
updating with the kernel.

Or introduce a resolution field for the EV_REL scope. Not sure how to do
that either and let's be honest, it wouldn't really be set by the kernel for
most devices anyway because it doesn't scale well (would've come in useful
for the various dpi settings on mice though).

That's all I can think of right now. For any solution - the libinput bits
are the easy ones, it's the inertia and possible other input stacks that are
the killer here.

Cheers,
   Peter