[belikh]

Summary

This ports the IRIX 6.5 idle-loop detector from Jupiter Belic's irix7/iris fork (PR #114, commits ee61088 and 52e5690) onto current upstream. With --features idle-pause, a settled IRIX 6.5 IP22 getty drops from ~100 % host CPU to ~1-2 %, and cold boots are not slowed.

Why the old detector never fired on 6.5

idleprof at a settled 6.5 IP22 getty (upstream build, xdm/amstart killed) shows the idle path is not a tight branch: it is a ~470-byte scheduler loop (0x880daedc..0x880db0b4, ~89 instructions) with calls to spl and the ll/sc lock and stores to kernel globals. The capture:

  ffffffff880db080  1.0%  100%  addiu sp, sp, -16
  ffffffff880db058  1.0%  100%  lw at, -24164(zero)
  ffffffff880db098  1.0%  100%  jal 0x80daedc
  ffffffff880daedc  1.0%  100%  addiu sp, sp, -32
  ... (the rest of the run-queue scan at ~1.0% each)
  ffffffff88019fb0  8.4%  100%  bgtz a0, -1        <- the DELAY loop
  ffffffff88019fb4  8.4%  100%  subu a0, a0, a1

Its registers evolve every iteration (the run-queue scan rotates state), so the PC+GPR hash never repeats exactly. The same hash did match a 3-instruction device-poll loop seen during boot (lbu/andi/bne -3 waiting on a SCSI bit), which must keep executing — parking it stalled the boot. The idle loop has no contiguous window covering half the samples (the helpers are far away), which is why the detector clusters the samples instead.

Detector

The run loop calls sample after each ten-instruction group; sample records the PC every 32 calls, so a record is ~320 guest instructions and the ring of 32 covers ~10,000. update sorts the samples into PC clusters (consecutive PCs ≤256 bytes apart) and parks when one kernel cluster holds ≥12 of the 32 samples with ≥12 distinct PCs. User-mode samples are allowed up to a quarter of the ring (a syscall wake runs a few user instructions between kernel stretches), but a user-dominated ring is refused.

That accepts IRIX 5.3's 14-instruction loop and 6.5's 89-instruction scheduler path while rejecting 2-instruction DELAY loops, 3-instruction device polls, the 8-instruction ll/sc lock spin and mixed work. The existing park rules are unchanged: interrupts enabled, a mask bit set, nothing pending, Compare armed; wake on any pending line; the masked-wake back-off (no re-park for ~20 ms after waking on a line the guest cannot take). Parking waits 60 s after the CPU thread starts (IRIS_IDLE_WARMUP_SECS overrides) so boot runs at full speed.

New monitor commands: idle on|off|status (runtime switch, park/sleep/wake counters, last window) and idleprof stats (the same counters).

The ci_clock ip7 unit tests now advance Count by retired cycles instead of sleeping, so the required --features ci_clock,idle-pause suite passes; no clock semantics change.

Measurements

Host: 6-core i5-8500T @ 2.1 GHz. Guest: IRIX 6.5 IP22, R4400, fresh sparse copy of the rig disk per run, nvram-rig.bin, headless, telnet forwarded; killall amstart; killall xdm after boot, then settle. Release build features: lightning,idle-pause unless noted.

measurement	result
settled getty, lightning,idle-pause, 4 × 10 s	1 %, 1 %, 1 %, 2 %
wall-clock cold boot to login	35 s quiet host, 46 s under moderate load
lightning,ci_clock,idle-pause cold boot to login	52 s
ci_clock,idle-pause getty, 3 × 10 s	19 %, 16 %, 18 % (see note)
monitor cpu stop; cpu start	resumes, parks again after the guest settles
guest date across a park	correct

Note on ci_clock: guest time is retired instructions, so the timer handler is a much larger share of host time (~16-19 % here). The wall-clock build is the one the 1-2 % figure describes. A later wall-clock boot on the same host under load average ~14 (three other emulators running) took 101 s, so the 35 s figure is a quiet-host sample; parking itself was not the cause.

Not tested: IRIX 5.3, because this rig has no 5.3 image. Its 14-instruction loop fits the cluster rules and the park/wake mechanics are unchanged.

Tests

build	command	result
default	cargo test --workspace	1348 passed, 0 failed
jitv2	cargo test --workspace --features jitv2	1830 passed, 0 failed
idle-pause	cargo test --workspace --features idle-pause	1366 passed, 0 failed
ci_clock,idle-pause	cargo test --workspace --features ci_clock,idle-pause	1366 passed, 0 failed

New tests cover the 6.5 window, distant helper calls, short device polls, a 2-instruction DELAY loop, user-mode loops, a moving window, the masked-wake back-off, and wake on Compare under both clock models. The masked-interrupt park test now samples the PC as the run loop does.

Provenance

This port is derived from Jupiter Belic's work in irix7/iris: PR irix7#114, commit irix7@ee61088 and docs correction irix7@52e5690. Taken from that work: the PC-cluster detector, the 60 s warm-up, the masked-wake back-off, the park statistics and idle/idleprof stats commands, the run-loop sampling points, and the ci_clock test adaptation. The capture and measurements were repeated here on upstream.