Remove SCHED_EEVDF, restore O(1) priority

[jserv]

EEVDF's two-pass picker walked every runnable peer under pcpu_runq_lock on every dispatch, which weakens the bounded dispatch-cost invariant the hard-RT contract rests on; its cross-CPU pcpu_min_vruntime read at the idle-steal and load-balance migration sites was unsynchronized; and its fair-share semantics are not property PSE51 callers express importance through (they use priority). Removing the entire path reduces the picker to a bitmap-and-FIFO dequeue and eliminates the race.

Replace the eevdf_fairness selftest with quantum_rotation_fairness: pin N CPU-bound equal-priority workers to one CPU, busy-loop on time_rdtime in each worker, and assert each worker's worst-case run-to-run latency stays within (N - 1) * quantum + jitter slack and that the worker was descheduled at least once. This exercises the timer-driven quantum expiry path that intra-band rotation actually relies on, not the voluntary-yield path the old test was inadvertently hitting via sleep_ms(0).

---

Summary by cubic

Remove EEVDF and return to O(1) priority scheduling with quantum-based FIFO rotation among equal-priority threads. This restores bounded dispatch cost, removes a cross-CPU race, and better aligns with PSE51 priority semantics.

• Refactors

  • Drop CONFIG_SCHED_EEVDF and all vruntime/deadline code; picker is now bitmap scan + FIFO within a priority level.
  • Equal-priority rotation is driven by timer quanta; no RR/OTHER behavior is exposed beyond SCHED_FIFO.
  • Update docs and configs to remove EEVDF and clarify the single normal-thread policy; sched_{set,get}scheduler accept SCHED_OTHER/SCHED_RR but always map/report SCHED_FIFO.
  • Replace eevdf_fairness with quantum_rotation_fairness: N CPU-bound peers on one CPU must see worst-case service gaps ≤ (N−1) quanta + small jitter, and each must be descheduled at least once.

• Bug Fixes

  • Eliminate the unsynchronized cross-CPU min-vruntime read in idle-steal and load-balance.
  • Restore predictable, bounded pick-next cost required for the hard-RT contract.

^{Written for commit 2f260a0. Summary will update on new commits.}

[cubic-dev-ai]

1 issue found across 13 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="tests/tests-sched.c">

<violation number="1" location="tests/tests-sched.c:558">
P1: The failure path returns before canceling `dom.refill_callout`, which can leave a timer referencing stack memory after `test_sched_domain_budget()` exits.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review, or fix all with cubic.}

[cubic-dev-ai]

P1: The failure path returns before canceling dom.refill_callout, which can leave a timer referencing stack memory after test_sched_domain_budget() exits.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At tests/tests-sched.c, line 558:

<comment>The failure path returns before canceling `dom.refill_callout`, which can leave a timer referencing stack memory after `test_sched_domain_budget()` exits.</comment>

<file context>
@@ -482,79 +482,113 @@ static i32 test_watchdog_activity(void)
+                              SCHED_PRIO_NORMAL, 0);
+        if (r.is_error) {
+            enable_interrupts();
+            return 1;
+        }
     }
</file context>