On Mon, 29 Oct 2018 at 14:12, Linus Torvalds [off-list ref]

So what the half-multiplier did, assuming a multiplier of 8 (which is
what my MX Anywhere 2S reports) would be:

 - remainder starts at 3
 - high-res is +1
 - now remainder is 3+1, and it triggers the >= half logic
 - 4/8 is 0, but then the code added 1 because high-res was positive,
so the code decides to add 1
 - the code does a wheel update of 1, and updates remainder with -8,
so now it's -4

Next time around, if the high-res update is 0 or -1, it will go the
other direction. And then it will oscillate.

Notice how tiny movements of +1/-1 in the *high-res* count can
translate into +1/-1 in the regular wheel movement.

Ah, I see what you mean. So, if we move the threshold to (multiplier -
1)/multiplier (7/8) in this case, I think the equivalent scenario
would be:

- remainder starts at 7
- high-res is +1
- remainder is now 7+1, triggering a low-res update
- 7/8 is 0, but we add one to the remainder in the check making it (7+1)/8 == 1
- we update remainder to -1

This way we're still at least 7/8ths of a notch from the threshold in
either direction, so we shouldn't get the oscillation problem. Does
that sound reasonable, or do you think I've missed something?

I do not believe that you actually ever *used* that code, or if you
did, you only did so with applications that were high-res aware and
ignored the regular wheel entirely because you were testing in an
environment with other changes than just the kernel.

I tested these changes with 5 different Logitech mice (see the
Logitech high-res support patch [0] for details), and did so mainly
with applications that were *not* high-res aware, using a mix of
clicky and smooth modes. Admittedly the MX Anywhere 2S was not one of
my test devices; I had assumed that its behaviour would be
sufficiently similar to that of the MX Anywhere 2 and the MX Master
2S.

Harry Cutts
Chrome OS Touch/Input team

[0]: https://patchwork.kernel.org/patch/10582935/