Timer::unregister() taking longer and longer as session time increases

[JustinWayland]

Making this issue to keep potential fixes in one place. Title changes are welcome, just describing what is being seen downstream.

This issue has popped up twice with Niri. Once here and again here.

Apparently there is a potential fix, but @YaLTeR could not find it.

Any help would be greatly appreciated.

Relevant info:
Tests by @abhbh
Niri Overview with just Alacritty just after starting Niri
Niri Overview with just Alacritty 1 hour after starting Niri

[YaLTeR]

Yeah, I remember some potential fix, possibly from @cmeissl, that required using some newer MSRV std function (retain_mut on HashMap maybe?). But honestly I don't entirely remember if that fix was for this performance problem.

[YaLTeR]

To bring some context from the linked issues: niri frequently (every input event from a potentially high Hz mouse) re-registers a timer under some conditions, e.g.:

• idle tracking (reset idle timer at every event)
• "hide cursor on timeout" tracking (also reset timer at every mouse event)

This for some reason causes calloop to gradually spend longer and longer processing Timer::unregister(), to the point where it starts dominating perf flamegraphs after an hour of use.