[Matsha]

What changed

On a laptop whose backlight is driven by nvidia (nvidia_0), ryoku-shell daemon sits at 100% CPU after waking from suspend until it is restarted.

Cause. nvidia unregisters and re-registers its backlight across suspend, so /sys/class/backlight/nvidia_0 is a new kernfs node after resume. The OSD watcher in ipc/osd.go keeps its fd on the old actual_brightness node. On that dead node poll(2) returns POLLERR at once, pread fails with ENODEV, and watchBacklight does continue straight back into poll. Nothing ever re-arms the notify, so the loop never blocks.

Seen on the live daemon (1.0.5, unchanged in osd.go on unstable-dev):

$ ls -l /proc/<pid>/fd | grep backlight
19 -> /sys/devices/.../backlight/nvidia_0/actual_brightness
$ stat -L -c %i /proc/<pid>/fd/19; stat -c %i /sys/class/backlight/nvidia_0/actual_brightness
34320
73544
$ cat /proc/<pid>/fd/19
cat: /proc/<pid>/fd/19: No such device

/proc/<pid>/io showed about 1.2M read syscalls per second with rchar not moving, all on one thread, and the node's timestamp matched the PM: suspend exit line.

Fix.

• The poll loop moves into watchBacklightNode, which returns once the node is dead: a read that fails or comes back empty.
• watchBacklight then re-resolves the device with backlightDevice() and reopens it, backing off from 1s to 30s while it is missing and resetting after a node that worked.
• readActual now returns an error, so a read that succeeds but does not parse keeps the fd: that read already re-armed the notify.
• The saved level is still restored only at daemon start, and a reopen publishes without bumping seq, so resume does not flash the OSD.

Area

ryoku

How it was tested

Arch, Hyprland, laptop with an nvidia_0 backlight.

• Diagnosed on the running daemon as above. Restarting ryoku-shell opens the new node and the daemon drops to 0% CPU, which confirms the stale fd is the cause.
• go build ./... && go vet ./... && go test ./... in ryoku/shell/ipc passes.
• New tests: TestReadActualClassifiesDeadNode (closed fd and empty read are dead-node errors, an unparsable value is errBacklightValue) and TestWatchBacklightNodeReturnsOnMissingNode. The sysfs POLLPRI path itself cannot be driven from a regular file in a test.
• Not yet run: the patched daemon through a real suspend and resume.

Checklist

• One logical change, with a clear [area] scope: summary commit subject.
• Matching CHANGELOG.md updated in the area I touched.
• The git hooks pass locally; I did not use --no-verify.
• Lua parses (luac -p), shell scripts pass bash -n, and QML passes
  qmllint where applicable.
• No duplicated config, no dead code, no commented-out code, no em-dash.
• Docs updated if behavior or layout changed.

[neur0map]

🦊 mika's review

Incremental review: new commits since the last checkpoint.

Summary

• Reopen vanished backlight nodes with bounded retry backoff.
• Retry devices unreadable at startup without restoring brightness on reopen.
• Preserve the watcher on unparsable readings and avoid resume OSD flashes.
• Add tests for read error classification and missing nodes.

Change diagram

flowchart LR
  Watcher["Backlight watcher"] --> Node["actual_brightness node"]
  Node -->|read failure| Retry["Retry with backoff"]
  Retry -->|re-resolve and reopen| Node
  Node -->|brightness change| OSD["OSD topic"]

Loading

Estimated review effort: 2/5 · risk: low

📁 Walkthrough details

ryoku/shell/ipc/osd.go

• Reopens missing or failed backlight nodes with exponential backoff, preventing a dead-node polling loop.
• Distinguishes unparsable values from failed or empty reads, retaining the descriptor when polling can continue.
• Restores saved brightness only at startup and publishes reopened levels without incrementing the OSD sequence.
  • Extracted watchBacklightNode and added device rediscovery.
  • Added retry delays from one to 30 seconds.
  • Changed readActual to return classified errors.

ryoku/shell/ipc/osd_test.go

• Tests valid, closed, empty, and unparsable backlight reads.
• Checks that missing and unspecified devices return without being marked as read.
  • Added TestReadActualClassifiesDeadNode.
  • Added TestWatchBacklightNodeReturnsOnMissingNode.

ryoku/shell/CHANGELOG.md

• Documents the resume-related CPU usage fix and backlight reopen behavior.
  • Added an unreleased brightness watcher fix entry.

No actionable findings. Mika looked for bugs, security issues, and risks in the diff and came up empty on clear defects.

🛠 Deterministic checks run

• linters: skipped (no checkout)

🚦 Merge recommendation: no_concerns_found

---

_{🔍 mika-agent · self-hosted review for the Ryoku org · @mika-agent help for commands}

[neur0map]

🔍 mika reviewed this change; details in the walkthrough comment.