Re: [PATCH 1/1] kthread: break dependency between worker->lock and task_struct->pi_lock

[Suren Baghdasaryan]

On Tue, Apr 28, 2020 at 11:04 AM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
>
> On Tue, Apr 28, 2020 at 9:31 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >
> > On Mon, Apr 27, 2020 at 11:43:58AM -0700, Suren Baghdasaryan wrote:
> > > A number of kthread-related functions indirectly take task_struct->pi_lock
> > > while holding worker->lock in the call chain like this:
> > >     spin_lock(&worker->lock)
> > >     kthread_insert_work
> > >     wake_up_process
> > >     try_to_wake_up
> > >     raw_spin_lock_irqsave(&p->pi_lock, flags)
> > >
> > > This lock dependency exists whenever kthread_insert_work is called either
> > > directly or indirectly via __kthread_queue_delayed_work in the following
> > > functions:
> > >     kthread_queue_work
> > >     kthread_delayed_work_timer_fn
> > >     kthread_queue_delayed_work
> > >     kthread_flush_work
> > >     kthread_mod_delayed_work
> > >
> > > This creates possibilities for circular dependencies like the one reported
> > > at: https://lkml.org/lkml/2020/4/24/954
> >
> > Please, do not use lkml.org links.
>
> Thanks for the review! Would
> https://lore.kernel.org/lkml/CAJuCfpG4NkhpQvZjgXZ_3gm6Hf1QgN_eUOQ8iX9Cv1k9whLwSQ@xxxxxxxxxxxxxx
> be better or should I just add the body of that report here? Or do not
> mention it at all?

Sorry, this time in plain text mode...
If there are no more comments on this patch I'll post a v2 with
lore.kernel.org instead of lkml.org link. Please let me know if there
are more issues that you would like to be addressed.
IMHO, taking kthread_queue_delayed_work() out from under rq->lock (if
we can figure out how to do that cleanly) can be a separate patch and
this one is still useful regardless of that.
Thanks!

>
> >
> > Also, ideally, we'd pull that kthread_queue_delayed_work() out from
> > under rq->lock.
>
> I understand but I don't see an easy way to do that. We need to start
> PSI polling whenever a monitored PSI state changes:
> https://elixir.bootlin.com/linux/v5.6.7/source/kernel/sched/psi.c#L783.
> This is happening under rq->lock because PSI accounting is done from
> inside enqueue_task/dequeue_task - the call chain is:
>
> enqueue_task > psi_enqueue > psi_task_change > psi_group_change >
> psi_schedule_poll_work > psi_task_change
>
> IIUC enqueue_task/dequeue_task are called with rq->lock taken, so
> moving kthread_queue_delayed_work out is not trivial.
>
> >
> > In fact, looking at it, WTH is the delayed branch of
> > kthread_queue_delayed_work() under that lock? That whole
> > delayed_work_list thing smells like bong-hits.
>
> I have the poll_scheduled atomic specifically to ensure that
> kthread_queue_delayed_work does not block as commented here:
> https://elixir.bootlin.com/linux/v5.7-rc3/source/kernel/sched/psi.c#L551.
> I understand this is not ideal. If there is a better way to schedule
> that kworker while ensuring it does not block I would be happy to
> rework this. Any suggestions?
>
> >
> > --
> > To unsubscribe from this group and stop receiving emails from it, send an email to kernel-team+unsubscribe@xxxxxxxxxxxx
> >