Re: [PATCH] sched_ext: allow scx_bpf_task_cgroup() in ops.dispatch()

[Changwoo Min]

Hello,

Thanks for the clarification.

On 9/17/25 14:00, Tejun Heo wrote:
> Hello,
>
> On Wed, Sep 17, 2025 at 11:12:18AM +0900, Changwoo Min wrote:
> > > I'm not sure this is safe tho. ops.dispatch() can release the rq lock it's
> > > holding to migrate tasks across rq's, which means that other operations can
> > > nest inside - ie. there can be an irq which triggers ops.enqueue() while
> > > ops.dispatch() is in progress. That can in turn overwrite
> > > current->scx.kf_tasks[].
> >
> > I thought that ops.dispatch() always holds an rq lock since there
> > is a lockdep_assert_rq_held(rq) check at the beginning of
> > balance_one(), which invokes the BPF scheduler's dispatch.
> > I guess I am missing an edge case?
>
> ops.dispatch() is always called and returns with rq locked; however, it
> sometimes needs to migrate tasks across rq boundaries and thus can
> temporarily release the rq lock it was called with. Please take a look at
> dispatch_to_local_dsq() which is called from balance_one() ->
> flush_dispatch_buf() -> finish_dispatch().

You are right. I overlooked that path.

> > > I wonder whether a better approach would be tracking cgroup membership from
> > > BPF side. ops.init_task() tells you the initial cgroup it's joining and if
> >
> > Currently, it is also not allowed to call scx_bpf_task_cgroup()
> > at ops.init_task() because the rq lock is not held at
> > ops.init_task(). The earliest possible moment to get a task's
> > cgroup ID is ops.enable().
>
> ops.init_task() is called with scx_init_task_args, which contains .cgroup
> which points to the CPU controller cgroup the task belongs to.

Yeah, that's better than relying on ops.enable().

Regards,
Changwoo Min