Re: [PATCH] sched_ext: Fix sched_ext_dead() race with scx_root_enable_workfn()

[Tejun Heo]

Hello,

The race seems real, thanks for catching it, but I'm not sure the
reader-side fix is the right shape. The new branch in sched_ext_dead()
resets state to NONE without a matching ops.exit_task(cancelled=true),
leaking what ops.init_task() set up; and the list_empty() gate sits
before scx_set_task_sched(), so a sched_ext_dead() that races after
sch is installed but before state goes READY would still flip state
to NONE under us.

Worth exploring on the writer side instead: reorder so p->scx.sched
is installed before state transitions off NONE. That restores the
"state != NONE -> p->scx.sched != NULL" invariant and the existing
sched_ext_dead() handles the rest. I haven't fully traced this
through - there may still be a residual window between INIT and the
workfn's READY write - but it seems like a more promising direction.

Thanks.

--
tejun