fix: three defects around offline CPUs and a blocked UI thread

[vietanhdev]

Three independent runtime defects found while auditing the backend. All three are invisible in normal operation on a machine with every core online.

1. The Monitor view blocks the UI thread ~10% of the time

get_system_monitor was a non-async #[tauri::command]. Tauri's default is ExecutionContext::Blocking, so the body runs inline on the IPC thread — not on the async runtime.

That body sleeps 200ms between the two /proc/stat samples it has to diff, and additionally shells out to df and ps aux synchronously. monitor.js polls every 2s.

So for as long as the Monitor view is open, the main thread is hard-blocked for ≥200ms out of every 2000ms — a visible duty cycle of dropped frames in scrolling, hover states, and window dragging.

spawn_blocking moves the whole collection off, subprocesses included, rather than just converting the sleep.

2. Setting the CPU governor fails wholesale if any CPU is offline

The old script counted cpuN directories, then emitted a bare redirect per index 0..count under set -e. Two bad assumptions: the count includes CPUs that are present but offline — whose cpufreq/ directory the kernel removes — and the numbering is assumed contiguous.

Disable SMT (echo 0 > .../cpu5/online) and the redirect for cpu5 fails with ENOENT, set -e aborts, pkexec returns non-zero — after cpu0–cpu4 were already changed. Mixed governors, and a UI saying the operation failed.

Now globs at execution time (as root, so it sees whatever is online at that moment) and treats only a total failure as an error.

3. Per-core stats are shifted past an offline CPU

parse_cpu_snapshot parsed each core id purely to check it was numeric, then discarded it and let the caller's enumerate() index stand in.

/proc/stat omits offline CPUs entirely. With cpu5 offline the file lists cpu0–4, cpu6, cpu7 — so vector index 5 is really cpu6. Every core past the gap was labelled one too low and read its frequency from the wrong CPU. Index 5 pointed at the offline cpu5, which has no cpufreq directory: 0 MHz.

The id is now carried through, and the two snapshots are matched by id rather than by position, so a CPU going offline between the two 200ms-apart reads can't diff one core's counters against another's.

Verification

104 tests, clippy clean.

The governor tests are checked against the genuine original implementation rather than a paraphrase — restoring it fails all four:

one_refused_write_does_not_discard_the_successful_ones ... FAILED
no_writable_cpu_at_all_is_still_an_error ... FAILED
non_contiguous_cpu_numbering_is_handled ... FAILED
an_offline_cpu_does_not_abort_the_others ... FAILED

Worth recording: an earlier mutation that only re-added set -e passed, because bash suppresses errexit inside an if condition. The original's bare redirects were what made it fire — so the weaker mutation proved nothing, and the test needed the real thing to be trustworthy.

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Updated (UTC)
❌ Deployment failed View logs	thinkutils	6a72077	Jul 20 2026, 03:14 AM