diff options
| author | Nick Moore <nicholasvmoore@gmail.com> | 2026-11-07 09:20:22 -0600 |
|---|---|---|
| committer | Nick Moore <nicholasvmoore@gmail.com> | 2026-11-07 09:20:22 -0600 |
| commit | 6c4039aebbe7ec25f47cb4623e44b96373e9e6e5 (patch) | |
| tree | 6b29ac15408e586c41d6bc91561294709c87a962 /rigel/.config/sway/ppd-switch.sh | |
| parent | 34abcb44caf7266ce40063daca05a2a06f1b26c7 (diff) | |
Track sway (shared + rigel machine overlay), replace fibonacci with dwm-dwindle listener
- sway package: shared config (styling, all muscle-memory binds, no exec
lines) including ~/.config/sway/machines/*.conf overlays included LAST;
floating_windows + machines/README contract; sway/AGENTS.md doctrine
- rigel package: machine overlay (outputs, touchpad, session stack, power
bindings) + session scripts now stow-managed; fibonacci.py replaced by
a POSIX sh+jq dwindle listener (dwm dwindle geometry, --monitor
subscription, mark+move re-home, flock singleton); machines/rigel.conf
exec line updated to the new path
- validated in a headless nested sway against a literal port of
fibonacci.c: five- and seven-window trees plus refocus-before-spawn
all match dwm's rects (modulo the bar offset)
Diffstat (limited to 'rigel/.config/sway/ppd-switch.sh')
| -rwxr-xr-x | rigel/.config/sway/ppd-switch.sh | 162 |
1 files changed, 162 insertions, 0 deletions
diff --git a/rigel/.config/sway/ppd-switch.sh b/rigel/.config/sway/ppd-switch.sh new file mode 100755 index 0000000..4d2f496 --- /dev/null +++ b/rigel/.config/sway/ppd-switch.sh @@ -0,0 +1,162 @@ +#!/bin/bash +# ppd-switch.sh — auto-switch the power-profiles-daemon profile on AC↔battery. +# +# power-profiles-daemon deliberately has no "switch profile on power source" +# public policy (that's a session thing; DEs like GNOME/KDE do it). 0.30 adds +# internal battery/AC awareness (BatteryAware) — feed events into drivers and +# actions, e.g. intel_pstate switches *internally* — but nothing emits a +# user-decidable balanced↔power-saver-on-ac rule. Sway has no DE policy +# engine, so this script is the policy. Everything talks to the *system* bus +# (ppd + UPower), which means no session env is required — the supervisor can +# spawn it from a bare context. +# +# The equality gate means we never fight ppd's own native choices: if the +# daemon already sits on the state-correct profile, apply_profile no-ops. +# +# Policy (agreed 2026-10-31): +# auto zone {balanced, power-saver} — on every power transition it flips +# to the state-correct profile: mains → balanced, battery → +# power-saver (first message's mapping; the "balanced (AC)" win). +# manual zone {performance} — never touched, on ANY transition. Sticks +# until the user changes it themselves (sway keybinds: +# Ctrl+Alt+Shift+3). Setting balanced/power-saver by hand +# re-enters auto, effective from the next transition: the +# check below re-reads ActiveProfile at every event, so the +# zone it lands in governs what happens next — no daemon state +# to track, the profile *is* the zone. +# +# Events: the battery's UPower PropertiesChanged signals. They also fire on +# plain charge-level updates every ~30s; those are cheap no-ops thanks to the +# equality check in apply(), so no debounce bookkeeping beyond one short +# settle-sleep per event (the kernel isn't atomic about adapter/battery +# reveals). +# +# Supervisor: part of the daemons=() array in session-daemons.sh — singleton +# flock, crash-restart, session teardown, and stderr rolling into +# session-daemons.log. Name must be ≤15 chars (it is), because the +# supervisor resolves pids via /proc/*/comm which truncates at 15. + +set -u + +BUS="--system" +PPI="org.freedesktop.UPower.PowerProfiles" +PPIPATH="/org/freedesktop/UPower/PowerProfiles" +# The only battery on this machine (checked: BAT0 absent, BAT1 present). +BATTERY="/org/freedesktop/UPower/devices/battery_BAT1" + +msg() { echo "$(date '+%F %T') [ppd-switch.sh] $*"; } + +# --- D-Bus helpers — raw gdbus calls, no powerprofilesctl dependency ------- +# (ppd 0.30: no SetProfile method on the interface; ActiveProfile is a +# readwrite property, Get/Set alongside every other property — verified by +# introspection and by flipping profiles live before writing this.) + +profile_get() { # -> ActiveProfile on stdout + gdbus call $BUS --dest "$PPI" --object-path "$PPIPATH" \ + --method org.freedesktop.DBus.Properties.Get "$PPI" ActiveProfile \ + | sed -n "s/^.*<'\([a-z][a-z-]*\)'>.*$/\1/p" # <'balanced'> -> balanced +} + +profile_set() { # $1 = profile; write is a variant "<'x'>" + gdbus call $BUS --dest "$PPI" --object-path "$PPIPATH" \ + --method org.freedesktop.DBus.Properties.Set "$PPI" ActiveProfile \ + "<'$1'>" >/dev/null || msg "set ActiveProfile='$1' FAILED" +} + +# --- power state ------------------------------------------------------------ + +# Any mains adapter online? Authored as a scan for type=Mains on purpose: +# the AC is ACAD but there are also four USB-C PD ports (ucsi-source-psy-*), +# and which one has the charger can vary by port. +mains_online() { + local t online + for d in /sys/class/power_supply/*/; do + read -r t <"$d/type" 2>/dev/null || continue + [ "$t" = Mains ] || continue + read -r online <"$d/online" 2>/dev/null || continue + [ "$online" = 1 ] && return 0 + done + return 1 +} + +desired_profile() { # state-correct auto profile + if mains_online; then echo balanced; else echo power-saver; fi +} + +# --- the one decision ------------------------------------------------------- + +apply_profile() { + local cur want + # Manual zone: performance is sticky across every transition. If it's set, + # decline entirely — including the startup one-shot. + cur=$(profile_get) + [ "$cur" = performance ] && return 0 + want=$(desired_profile) + [ "$cur" = "$want" ] && return 0 # no-op: also absorbs the + # charge-update event noise + msg "transition -> '$want' (was '$cur')" + profile_set "$want" +} + +# --- event loop --------------------------------------------------------------- + +# Teardown: the supervisors TERM-kills the whole leftover stack, and our +# instance was setsid'd into its own process group (we are the group leader), +# so every child — gdbus monitor, the settle sleeps — shares the group. kill +# 0 takes them all out at once instead of leaving the monitor orphaned until +# its next battery write hits a closed pipe (SIGPIPE, ~30s later). Guard: if +# we are NOT the process-group leader (e.g. someone ran this interactively), +# kill 0 would take the caller's terminal pgrp with it — skip the group kill +# there and just exit. +cleanup() { + # Clear the traps FIRST: leaving them active while kill 0 TERMs our own + # group means this very shell re-enters the trap from its own group kill, + # and the handler's exit 0 never wins the race (verified: unbounded trap + # replay). Cleared, the self-TERM lands on default disposition and drops + # the shell instantly — exactly probe2.sh's correct behaviour. + trap - EXIT HUP INT TERM + if [ "$(ps -o pgid= --no-headers -p $$ 2>/dev/null | tr -d '[:space:]')" = "$$" ]; then + kill 0 2>/dev/null + fi + exit 0 +} +trap cleanup EXIT HUP INT TERM + +# Startup one-shot, zone-respecting: a performance pick from a previous +# session therefore also survives a full session restart. Failures here (e.g. +# ppd briefly not ready at login) are not fatal — the loop below converges on +# the first battery signal, and profiles only need one read/write. +apply_profile + +while :; do + # monitor only watches the battery device: its PropertiesChanged is the + # trigger for "power config may have changed". If the monitor dies (bus + # hiccup), the outer while restarts it; the supervisor's crash-restart + # covers the script as a whole. + # Coupling subtlety #1 — pipeline: `gdbus monitor | while read` seats the + # main shell in do_wait on the pipeline and never services pending traps. + # Subtlety #2 — read builtin: even with the reader as the main shell + # (process substitution), bash defers pending traps until the read + # *returns* (verified live: TERM was ignored for tens of seconds). The + # pattern that works is reader-in-subshell + main shell waiting via + # `wait $!`: wait-on-async runs deferred traps IMMEDIATELY, so the + # supervisor's TERM -> our trap -> kill 0 teardown actually executes. + # The reader owns no state; it is pure "event -> apply" code, so it needn't + # handle a TERM itself — the group kill in cleanup covers it. + reader() { + while read -r; do + apply_profile + sleep 0.3 # settle: hot-trailing signals are gated in apply() + done + } + reader < <(gdbus monitor $BUS --dest org.freedesktop.UPower --object-path "$BATTERY") & + wait $! + msg "monitor exited; restarting in 30s" + # sleep is a builtin run to completion, and bash defers pending traps + # until a builtin RETURNS (this file's history is one big proof of that) — + # so the 30s restart backoff sleeps in 1s slices: a pending TERM never + # waits longer than ~1s from here. + for _ in $(seq 30); do + sleep 1 + done +done |
