[Japabu]

Problem

SchedPass::begin drains the mailbox before the pass knows its disposition. A fair wake posted on the poster's own CPU was judged against the running task in the very pass that parks or ends that task, and was forwarded to an idle CPU: an IPI and a cold cache where a local switch a moment later was free. The same forward moves communicating pairs apart. In fs-rtt's r2 sweep on the AMD bench, a client that landed on the server's CPU had the server forwarded away (10→14, 14→8), and a pair split across the two CCXs pays about 1.9 µs per round trip (read_4k_at 16.1–16.4 µs in one LLC against 19.9–20.4 µs across).

Behaviour

• Step 1: the opening drain holds any wake that would queue behind the running task. Each disposition then places it against the CPU as it leaves it:

  • after a park, a stop or an exit, the wake runs here;
  • before a yield, or with the task kept, it is forwarded as before.

  Only the opening drain holds.

• Step 2: WakeCause::parks_next(cpu) names the CPU of a poster that parks next. The home CPU hands such a fair wake there inside one LLC, and the poster's park runs it. No caller sets it yet; it is the placement input for the IPC phase-2 combined call.

Checks

check	revision	result
cargo run -- --ci host	the tree before 161fc1e1c was dropped from this branch	105 steps, exit 0
cargo test -p kernel --lib, cargo test -p kernel-sim, kernel clippy	b24b37a56	exit 0 each

• Host tests: six (park, exit and yield for step 1; same LLC, other LLC and RT for step 2), plus the simulator's turn_taking_pair over 40 seeds: no turn forwarded off its CPU.
• Mutations: six, each turning only its own test red. The simulator test also reds without step 1.

The rebase changed nothing these checks cover, so the host CI result carries over.

Evidence and what is still owed

• Logs: (evidence link added with the A/B).
• Metal: an A/B on the T14 and the AMD bench is requested, bundled with the affinity PR stacked on this one. Base is 370d7d9f6; rows: osbench's pingpong same_cpu, ipc, wake.onto_busy, spawn_cost and fs-rtt's fsrtt.* from osbench: files reads read_4k_at's parts, fstat, and the file server's client alone #1103. This PR's effect is read from the ipc and fsrtt rows.

Risks

• A held wake waits for the disposition within the same pass. That costs nothing measurable on the host, and the metal price is not yet read.
• The one-LLC limit for step 2 is a one-line change if the fs-rtt sweep argues otherwise.

🤖 Generated with Claude Code