scheds/experimental/scx_flow: v2.2.4 — priority-aware rt_sensitive and relaxed preempt threshold

[galpt]

Summary

This is an incremental update on top of the scx_flow v2.2.0 scheduler that
landed in PR #3525. It carries two changes since that submission, both aimed
at improving worst-case wakeup latency without adding code paths that bypass
the existing lane analysis or containment machinery:

• Priority-aware rt_sensitive — the rt_sensitive_ready condition now
  checks p->prio so that kernel RT-class tasks (SCHED_FIFO / SCHED_RR at
  any priority) qualify for WAKE_PROFILE_RT_SENSITIVE on sufficient refill,
  rather than needing the combination of being pinned and having the 80us
  refill floor. Fixed in v2.2.4: the refill threshold is properly ANDed with
  the genuine RT classification (pinned || prio < 100) so well-refilled
  SCHED_OTHER tasks are not falsely classified as RT-sensitive.
• Relaxed preempt_ready refill check — the original preempt_ready
  condition required both last_refill_ns >= preempt_refill_min_ns (200us)
  and budget_ns >= preempt_budget_min_ns (150us). For short-interval
  wakeups such as cyclictest at 200us the per-wakeup refill is only
  sleep_ns / 100 = 2us, far below the 200us threshold. Dropped the
  refill gate so that accumulated positive budget alone qualifies a task for
  the preempt path.

Changes

The BPF changes are confined to one function in main.bpf.c.

1. Priority-aware rt_sensitive (v2.2.3, fixed in v2.2.4)

// v2.2.0 baseline — pinned + refill >= 80us
p->nr_cpus_allowed == 1 &&
tctx->last_refill_ns > 0 &&
tctx->last_refill_ns >= (s64)FLOW_INTERACTIVE_FLOOR_MIN_NS

// v2.2.3 (buggy) — last_refill_ns >= 80us was a standalone OR
tctx->last_refill_ns > 0 &&
(p->nr_cpus_allowed == 1 ||
 p->prio < 100 ||
 tctx->last_refill_ns >= (s64)FLOW_INTERACTIVE_FLOOR_MIN_NS)

// v2.2.4 (fixed) — refill threshold ANDed with RT classification
tctx->last_refill_ns > 0 &&
(p->nr_cpus_allowed == 1 || p->prio < 100) &&
tctx->last_refill_ns >= (s64)FLOW_INTERACTIVE_FLOOR_MIN_NS

p->prio < 100 identifies kernel SCHED_FIFO and SCHED_RR tasks (which have
priority 0--99).

2. Relaxed preempt_ready refill check

// before
(tctx->last_refill_ns >= (s64)preempt_refill_min_ns &&
 tctx->budget_ns >= (s64)preempt_budget_min_ns)

// after
tctx->budget_ns >= (s64)preempt_budget_min_ns

The idle-CPU local-reserved path that was present in v2.2.2/v2.2.3 has been
removed — it bypassed the latency lane routing, causing periodic tasks to
lose their priority dispatch slot. The net submission only carries the two
changes above.

Benchmark Results

Full tagged artifacts (PNG, SVG, CSV, report):
https://github.com/galpt/testing-scx_flow/tree/benchmark-archives/20260409_scx_flow_v2.2.0_release/mini/v2.2.4

All runs on CachyOS 7.0.3-1-cachyos, 16-core AMD system, Balanced power profile.

Normal mode (cyclictest -D 30 -t 4 -a 0 -m -v)

Scheduler	Max Latency (us)	Spikes >100us
baseline (CFS/EEVDF)	1106	304
scx_cosmos	852	821
scx_bpfland	2724	222
scx_flow v2.2.4	142	2

Hard RT mode (cyclictest --priority=99 --smp --interval=200 --histogram=20)

Scheduler	Overflows >20us	Max Latency (us)
baseline (CFS/EEVDF)	344	407
scx_cosmos	351	227
scx_bpfland	376	375
scx_flow v2.2.4	402	339

Note: hard RT cyclictest runs at SCHED_FIFO 99 which is dispatched by the
kernel's rt_sched_class, not by ext_sched_class. The hard RT results reflect
system noise under SCX background load rather than the scheduler policy
itself, and are consistent across all tested schedulers.

Files Changed

 Cargo.lock                                      | 2 +-
 scheds/experimental/scx_flow/Cargo.toml         | 2 +-
 scheds/experimental/scx_flow/README.md          | 2 +-
 scheds/experimental/scx_flow/src/bpf/main.bpf.c | 5 ++---
 4 files changed, 4 insertions(+), 5 deletions(-)

Signed-off-by: Galih Tama galpt@v.recipes

[Copilot]

Pull request overview

Updates the experimental scx_flow sched_ext scheduler to v2.2.3 by refining wakeup classification and local dispatch behavior to improve responsiveness (especially for RT-priority tasks) without reintroducing the prior short-sleep lane-analysis bypass.

Changes:

• Make rt_sensitive wake profiling priority-aware by marking RT-priority tasks (p->prio < 100) as WAKE_PROFILE_RT_SENSITIVE on positive refill.
• Relax preempt_ready gating to depend on accumulated budget only (drops the refill-threshold requirement).
• Extend the local-reserved fast path to cover wakeups targeting an idle CPU.
• Bump version references to 2.2.3 in Cargo.toml, Cargo.lock, and README text.

Reviewed changes

Copilot reviewed 3 out of 4 changed files in this pull request and generated 3 comments.

File	Description
scheds/experimental/scx_flow/src/bpf/main.bpf.c	Adjusts wake-profile logic and enqueue fast-path routing (RT sensitivity, preempt gating, idle-CPU local dispatch).
scheds/experimental/scx_flow/README.md	Updates benchmark snapshot label to v2.2.3 (link currently appears inconsistent).
scheds/experimental/scx_flow/Cargo.toml	Bumps crate version to 2.2.3.
Cargo.lock	Updates locked scx_flow package version to 2.2.3.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[sirlucjan]

Tested, everything is OK.