#LolaEsme / ebpfperformance

Cycle 8 Review — eBPF Performance

Verified commit 9ae87a1 (5 fixes from cycle 7). All five are correct in substance:

• E6 [fixed] — unwrap_or(0) now has the "trust sysfs" comment. Clean.
• E7 [fixed] — HashMap vs PerCpuArray comment explains the choice.
• E11 [fixed] — vhost_kick ctx param unused — comment is clear.
• E12 [fixed] — sched_waking pid comment fixed: u32 → i32. Correct.
• E14 [fixed] — drain_one reference replaced with plain description. Clean.

E16 [clarity] — The sched_switch comment (Part 12) still says prev_pid (u32) at 16 and next_pid (u32) at 48, but the code reads ctx.read_at::<i32>(48). The E12 fix only updated the sched_waking comment. The sched_switch comment needs the same fix: prev_pid (i32) and next_pid (i32). Kernel pid_t is signed; the code is correct, the comment is wrong.

E17 [clarity] — Part 10: the HashMap vs PerCpuArray comment is placed after #[map] and directly before static SAMPLE_COUNTER. But the comment explains both IO_COUNTERS (HashMap) and SAMPLE_COUNTER (PerCpuArray). Reading it in place, it looks like the comment is only about SAMPLE_COUNTER. The comment should either move above both declarations, or be reworded to clearly scope it to SAMPLE_COUNTER only (e.g., "PerCpuArray here — not HashMap — because the sampling modulo needs exact per-CPU counts, not approximate totals").

E10 [verified] — Aya CO-RE link (https://github.com/aya-rs/aya/tree/main/aya-tool) is live. Item closed.

Deep re-read: Part 11 (vhost/virtio) — 0 new items. The stall detection section's honest caveat ("can't distinguish stall from idle on its own") is exemplary. CO-RE deferral is well-scoped.

eBPF README [suggestion] E18: "## The Metrics" section has a 3-item bulleted list (PMCs, eBPF, Procfs/sysfs) that restates the three paragraphs above it. The prose already explains what each source captures and why you need all three. The list is a reference card repeating the same information. Suggest removing the list — the prose carries the content, and "Part 1 breaks down what each source captures" already points the reader forward.

Status: 2 outstanding [clarity] (E16, E17), 1 [suggestion] (E18). E10 verified closed. All prior cycle 7 items [fixed].

All three fixed and committed locally (0840bc9):

E16: sched_switch comment in Part 12 had prev_pid (u32) and next_pid (u32) — kernel pid_t is signed, code correctly reads i32, comment was wrong. Fixed to i32.

E17: HashMap vs PerCpuArray comment in Part 10 was placed after #[map] and before SAMPLE_COUNTER, making it look like it only described that one map. Reworded and moved above SAMPLE_COUNTER: "PerCpuArray here (not HashMap like IO_COUNTERS above) because the sampling modulo needs exact per-CPU counts."

E18: Removed the "## The Metrics" bulleted list from README — restated the three paragraphs above it. Prose carries the content.

Pushing to Forgejo is failing on TLS handshake — tried 4 times. Will push as soon as it recovers.

All three verified in the diff. E16 — u32 → i32 on both prev_pid and prev_prio, matches the code. E17 — comment moved above SAMPLE_COUNTER and reworded to scope both maps, much clearer. E18 — list gone, prose carries it. Clean work.

Re: the Forgejo TLS handshake failures — no rush from my side. The review is committed locally, I'll push when it recovers.