Re: [PATCH 2/2] kthread: don't use to_live_kthread() in kthread_park() and kthread_unpark()

[Thomas Gleixner]

On Mon, 31 Oct 2016, Oleg Nesterov wrote:
> I think we need to unexport kthread_park/unpark, and either make it return
> "void" or actually fix the race with kthred_stop/exit. This patch just adds
> WARN_ON(PF_EXITING) for now.

I'll have a look.
 
> The usage of kthread_park() in cpuhp code (cpu.c, smpboot.c, stop_machine.c)
> is fine. It can never see an exiting/exited kthread, smpboot_destroy_threads()
> clears *ht->store, smpboot_park_thread() checks it is not NULL under the same
> smpboot_threads_lock. cpuhp_threads and cpu_stop_threads never exit, so other
> callers are fine too.
> 
> But it has two more users:
> 
> - watchdog_park_threads() and it does not look nice. The code is actually
>   correct, get_online_cpus() ensures that kthread_park() can't race with
>   itself (note that kthread_park() can't handle this race correctly), but
>   imo it should not use kthread_park() directly.

Should we provide an interface through the smpboot thread infrastructure for
this?

> - drivers/gpu/drm/amd/scheduler/gpu_scheduler.c and I think it should not
>   use kthread_park() too.
> 
>   But this patch should not break this code. kthread_park() must not be
>   called after amd_sched_fini() which does kthread_stop(), otherwise even
>   to_live_kthread() is not safe because task_struct can be already freed
>   and sched->thread can point to nowhere.

Right. That's why the smpboot thread code holds a task ref which is only
released after the thread has been stopped.

I can see why that gpu driver wants to use the park mechanism and I guess
there are other legitimate use cases as well. I prefer to implement a
park/unpark variant which is safe to use on arbitrary kthreads over forcing
driver writers to come up with even more broken open coded implementations
of that.

> Signed-off-by: Oleg Nesterov <oleg@xxxxxxxxxx>

Reviewed-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>