Skip to content
Auriya

System Tweaks

The tweak layer is where the daemon actually touches the device: guarded writes to /proc and /sys, plus a few actions that must be routed through Android via the companion. Everything here is best-effort and device-dependent — a node that does not exist is skipped, never fatal.

Module map

Section titled “Module map”

src/core/tweaks/ (tweaks/mod.rs) splits the surface by concern:

Module Responsibility
paths.rs Scans and caches sysfs paths once (CPU governors, online, Snapdragon KGSL).
cpu.rs Governor, boost, core onlining, per-core affinity.
gpu.rs GPU performance/balanced mode.
sched.rs Scheduler tunables.
memory.rs Memory / swappiness / cache drop.
storage.rs Block-I/O tunables.
touchpanel.rs Touch “game mode”.
ceiling.rs Frequency-ceiling controller (Low/Balance).
init.rs One-shot general tweaks applied with the Performance profile.
vendor/ (detect.rs, mtk.rs, snapdragon.rs) SoC detection + vendor-specific hooks.
vendor_lock.rs Locks vendor perfmgr nodes so vendor services cannot fight Auriya.

Which of these a given profile triggers is the single-source table in Architecture overview → What each static profile changes.

Path detection and caching

Section titled “Path detection and caching”

Sysfs layout varies per device, so paths are discovered once and memoized in a OnceLock. CpuPaths::scan probes cpu0..15 and policy0..7 governor nodes plus cpu1..15/online, keeping only those that exist (paths.rs, cpu_paths()). set_governor_cached then writes every cached governor node, ignoring individual write errors:

pub fn set_governor_cached(governor: &str) {
    let paths = cpu_paths();
    for path in &paths.governors_cpu    { let _ = std::fs::write(path, governor); }
    for path in &paths.governors_policy { let _ = std::fs::write(path, governor); }
}

The same pattern caches Snapdragon KGSL/memlat paths and their original values so they can be restored later (SnapdragonPaths::scan, paths.rs).

Guarded, best-effort writes

Section titled “Guarded, best-effort writes”

The prevailing idiom (e.g. init.rs) is: check existence, write, ignore failure. Missing nodes are silently skipped so a different kernel layout keeps running. A concrete slice of the general tweaks (apply_general_tweaks, init.rs):

// disable kernel panics (only nodes that exist)
for (path, value) in [
    ("/proc/sys/kernel/panic", "0"),
    ("/proc/sys/kernel/panic_on_oops", "0"),
    ("/proc/sys/kernel/panic_on_warn", "0"),
    ("/proc/sys/kernel/softlockup_panic", "0"),
] {
    if Path::new(path).exists() { fs::write(path, value)?; }
}

apply_general_tweaks also: sets block-I/O tunables per /sys/block/*/queue (iostats=0, add_random=0, read_ahead_kb=32, nr_requests=32), picks the best available TCP congestion control from bbr3 → bbr2 → bbrplus → bbr → westwood → cubic, applies VM/scheduler tunables, and disables several OEM “assist”/bloat modules and the OEM battery-saver kernel module (init.rs). These are OEM-node-dependent and mostly no-ops on a device that lacks them.

Actions routed through Android — CmdWriter

Section titled “Actions routed through Android — CmdWriter”

Two decisions cannot be done from the root daemon because they require Android framework APIs (NotificationManager for Do-Not-Disturb, the display manager for refresh rate): DnD and refresh-rate changes. The daemon serializes these to a small command file that the companion service watches and replays through the proper APIs (cmd_writer/mod.rs module comment).

  • File: /data/adb/.config/auriya/auriya_cmd (CMD_FILE).

  • Wire format (mirrors the companion’s CmdFormat.kt):

    seq 42
    dnd 1              # 0 = All/off, 1 = Priority
    refresh_rate 90    # Hz; 0 means "restore previous"

  • Stateful re-emit: the companion reads the whole file each time (it is one command, not a queue). Two quick single-field writes — dnd then refresh_rate on a game switch — would otherwise clobber each other, so the writer remembers the last value of every field and re-emits the full state on each write (CmdWriter::write).

  • Atomic delivery: written to .auriya_cmd.tmp then renamed, so the companion’s inotify watcher only sees a complete payload (CmdWriter::write). seq is a process-monotonic counter for dedup.

There is one process-wide writer (shared()); using more than one would reset the seq counter and break the companion’s dedup.

SoC detection

Section titled “SoC detection”

Vendor hooks pick a SoC family once, cached in a OnceLock (vendor/detect.rs, detect_soc). Detection tries, in order: ro.board.platform prefixes (mt/k6 → MediaTek; sm/sdm/msm/apq → Snapdragon; exynos; ud710/ums → Unisoc; gs → Tensor), then ro.hardware substrings, then filesystem probes (/proc/ppm → MediaTek, /sys/class/kgsl/kgsl-3d0 → Snapdragon). Falls back to Unknown. Only MediaTek and Snapdragon have implemented hook modules at this revision.

Vendor lock — stopping vendor services from fighting back

Section titled “Vendor lock — stopping vendor services from fighting back”

On many devices a vendor “perfmgr”/game service continuously rewrites the same CPU/GPU nodes Auriya sets, undoing Auriya’s changes within milliseconds. VendorLock neutralizes this by pinning a set of vendor toggles read-only via a bind mount (vendor_lock.rs, lock_all):

For each existing path in VENDOR_PATHS it: saves the current value, writes the desired value, chmods the node to 0444, then bind-mounts a Auriya-owned file over it (MS_BIND | MS_REC) so even privileged writes hit the overlay, not the real node. unlock_all unmounts, restores permissions, and writes the saved value back. The locked set (VENDOR_PATHS):

/sys/module/mtk_fpsgo/parameters/perfmgr_enable      → 0
/sys/module/perfmgr/parameters/perfmgr_enable        → 0
/sys/module/perfmgr_policy/parameters/perfmgr_enable → 0
/sys/module/perfmgr_mtk/parameters/perfmgr_enable    → 0
/sys/module/migt/parameters/glk_fbreak_enable        → 0
/sys/module/migt/parameters/glk_disable              → 1
/proc/game_opt/disable_cpufreq_limit                 → 1

Locking happens when entering a game session and unlocking when leaving it (see Profile scheduler). A failed mount-bind reverts the node’s permissions and logs a warning rather than aborting.

VendorLock calls mount/umount2 and chmods kernel nodes — a genuine trust boundary. It only touches the fixed VENDOR_PATHS list above, and only nodes that already exist.

Likely to drift first

Section titled “Likely to drift first”

VENDOR_PATHS, the general-tweak node lists and their values in init.rs, and the SoC-detection prefixes. These are hardware-specific and change most often. Re-verify against src/core/tweaks/init.rs, vendor_lock.rs, and vendor/detect.rs.