[PATCH 0/2] sched: Fix SCX_ENQ_CPU_SELECTED for queued wakeups

[Kuba Piecuch]

SCX_ENQ_CPU_SELECTED is meant to tell ops.enqueue() whether
ops.select_cpu() ran for a wakeup. However, wakeups delivered through
the ttwu wakelist never have the flag set, because sched_ttwu_pending()
drops WF_RQ_SELECTED. Schedulers that set SCX_OPS_ALLOW_QUEUED_WAKEUP
therefore cannot tell whether ops.select_cpu() was skipped, and
SCX_EV_SELECT_CPU_FALLBACK is undercounted.

Patch 1 carries WF_RQ_SELECTED across the wakelist the same way
WF_MIGRATED already is. Patch 2 adds a selftest. In a 4-vCPU VM, without
patch 1 the test reports that all of the roughly 110k queued wakeups in
a 2s run had ops.select_cpu() called but SCX_ENQ_CPU_SELECTED clear.
With patch 1 there are none.

Patch 1 changes kernel/sched/core.c and include/linux/sched.h, but only
sched_ext sees the effect. Peter, Tejun: I'll leave it to you to decide
whether the series goes through tip or the sched_ext tree.

The bug was found, and the series was drafted, with the help of an AI
coding agent.

Kuba Piecuch (2):
  sched/core: Preserve WF_RQ_SELECTED across queued wakeups
  selftests/sched_ext: Test SCX_ENQ_CPU_SELECTED with queued wakeups

 include/linux/sched.h                         |   6 +-
 kernel/sched/core.c                           |  10 +-
 tools/testing/selftests/sched_ext/Makefile    |   1 +
 .../sched_ext/enq_cpu_selected.bpf.c          | 107 +++++++++++++
 .../selftests/sched_ext/enq_cpu_selected.c    | 149 ++++++++++++++++++
 5 files changed, 270 insertions(+), 3 deletions(-)
 create mode 100644 tools/testing/selftests/sched_ext/enq_cpu_selected.bpf.c
 create mode 100644 tools/testing/selftests/sched_ext/enq_cpu_selected.c


base-commit: 734bef548a96e8774d6eef5dfa2026c6edc4e80b
-- 
2.56.0.257.gf7497c7d27-goog