[PATCHSET v2 sched_ext/for-7.3-fixes] sched_ext: Fix missing ops.dequeue() on remote local DSQ moves

[Kuba Piecuch <jpiecuch@google.com>]

Hi,

When a task in the BPF scheduler's custody is moved to another CPU's
local DSQ, ops.dequeue() is deferred until the task is picked and then
reported with SCX_DEQ_CORE_SCHED_EXEC, instead of being called without
flags when the task is inserted into the destination DSQ. Patch 1 fixes
this by reverting the enqueue side of b75aaea24c9f ("sched_ext: Properly
mark SCX-internal migrations via sticky_cpu"). Patch 2 adds a selftest.

For 7.1.y, patch 1 also needs 18d62044cda7 ("sched_ext: Preserve rq
tracking across local DSQ dispatch"), as noted in its stable tags.

This is based on sched_ext/for-7.3-fixes (d35a535d3e3e). Merging it into
sched_ext/for-next gives a trivial conflict in the selftests Makefile,
where enq_blocked was added next to dequeue_remote; keep both.

Testing was done on x86_64 in virtme-ng with 4 vCPUs in an SMT topology
(2 cores x 2 threads) and CONFIG_SCHED_CORE=y. Without patch 1,
dequeue_remote fails in every run (30/30), e.g.:

  sched_ext: dequeue_remote: dequeue_remote.bpf.c:141: 15 (rcu_preempt): late ops.dequeue() with SCX_DEQ_CORE_SCHED_EXEC (enq_cpu=3 cpu=2 seq=1)
     ...
     ops_dequeue+0x114/0x170
     set_next_task_scx+0x104/0x1e0
     __pick_next_task+0xc7/0x180
     __schedule+0x154/0x1870

and the full sched_ext selftest suite reports 31 passed, 1 skipped
(nohz_tick), 1 failed (dequeue_remote).

With patch 1, dequeue_remote passes in every run (30/30), with ~140k
custody enqueues per run, ~90k of them followed by the task running on
another CPU. The full suite reports 32 passed, 1 skipped (nohz_tick),
0 failed. dequeue_remote also passes 10/10 with the runner and its
workers sharing a core cookie, where legitimate core-sched picks out of
custody do happen.

v2:
 - Reordered to put the fix first and folded the Makefile entry into the
   test patch (Tejun, Andrea).
 - Fix: clear p->scx.sticky_cpu right after reading it, making the fix a
   revert of the enqueue side of b75aaea24c9f (Andrea). Reworded the
   comment and shortened the description (Tejun). Noted 18d62044cda7 as
   a 7.1.y prerequisite (Tejun, Andrea). Added Andrea's Reviewed-by.
 - Test: check p->core_cookie on SCX_DEQ_CORE_SCHED_EXEC instead of
   skipping when core scheduling is in use (Tejun).
 - Test: count CPUs with sched_getaffinity() (Tejun), pop past stale queue
   entries in ops.dispatch(), reset and print all counters per scenario
   (Tejun), destroy the struct_ops link and reap workers on error paths
   (Sashiko).
 - Test: deduplicated the descriptions and lowercased single-line
   comments (Tejun).

v1: https://lore.kernel.org/r/20260929161730.185271-1-jpiecuch@google.com

Thanks,
Kuba

Assisted-by: Claude:claude-opus-5.5

Kuba Piecuch (2):
  sched_ext: Call ops.dequeue() when a task arrives on a remote local
    DSQ
  selftests/sched_ext: Add a test for ops.dequeue() on remote local DSQ
    moves

 kernel/sched/ext/ext.c                        |  10 +-
 tools/testing/selftests/sched_ext/Makefile    |   1 +
 .../selftests/sched_ext/dequeue_remote.bpf.c  | 265 ++++++++++++++++++
 .../selftests/sched_ext/dequeue_remote.c      | 202 +++++++++++++
 4 files changed, 475 insertions(+), 3 deletions(-)
 create mode 100644 tools/testing/selftests/sched_ext/dequeue_remote.bpf.c
 create mode 100644 tools/testing/selftests/sched_ext/dequeue_remote.c


base-commit: d35a535d3e3e71f92a051d94e415f43612c4edfc
-- 
2.56.0.rc1.315.gc6ed9934b7-goog