[PATCHSET sched_ext/for-7.1] Fix sub-sched locking issues

[Tejun Heo <tj@kernel.org>]

Hello,

Cheng-Yang reported a lockdep circular dependency between scx_sched_lock and
rq->__lock. scx_bypass() and sysrq_handle_sched_ext_dump() take
scx_sched_lock -> rq lock, while scx_claim_exit() (reachable from many paths
with rq lock held) takes rq -> scx_sched_lock. In addition, scx_disable()
directly calling kthread_queue_work() under scx_sched_lock creates another
chain through worker->lock -> pi_lock -> rq->__lock.

This patchset fixes these issues:

1. Fix wrong sub_detach op check.
2. Add scx_dump_lock and dump_disabled to decouple dump from scx_sched_lock.
3. Always bounce scx_disable() through irq_work to avoid lock nesting.
4. Flip scx_bypass() lock order and drop scx_sched_lock from sysrq dump.
5. Reject sub-sched attachment to a disabled parent.

Tested on three machines (16-CPU QEMU, 192-CPU dual-socket EPYC, AMD Ryzen)
with lockdep trigger tests and an 11-test stress suite covering
attach/detach, nesting, reverse teardown, rapid cycling, error injection,
SysRq-D/S dump/exit, and combined stress. Lockdep triggered on baseline,
clean after patches.

Based on sched_ext/for-7.1 (b8840942644c).

 1. sched_ext: Fix sub_detach op check to test the parent's ops
 2. sched_ext: Add scx_dump_lock and dump_disabled
 3. sched_ext: Always bounce scx_disable() through irq_work
 4. sched_ext: Fix scx_sched_lock / rq lock ordering
 5. sched_ext: Reject sub-sched attachment to a disabled parent

 kernel/sched/ext.c          | 59 +++++++++++++++++++++++++++++++++++----------
 kernel/sched/ext_internal.h |  3 ++-
 2 files changed, 48 insertions(+), 14 deletions(-)

--
tejun