Re: [PATCH 1/1] psi: eliminate kthread_worker from psi trigger scheduling mechanism

[Suren Baghdasaryan]

On Mon, Jun 8, 2020 at 7:56 PM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
>
> On Thu, Jun 4, 2020 at 12:20 PM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
> >
> > On Thu, Jun 4, 2020 at 6:12 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > >
> > > On Thu, May 28, 2020 at 12:54:42PM -0700, Suren Baghdasaryan wrote:
> > > > Each psi group requires a dedicated kthread_delayed_work and
> > > > kthread_worker. Since no other work can be performed using psi_group's
> > > > kthread_worker, the same result can be obtained using a task_struct and
> > > > a timer directly. This makes psi triggering simpler by removing lists
> > > > and locks involved with kthread_worker usage and eliminates the need for
> > > > poll_scheduled atomic use in the hot path.
> > > >
> > > > Signed-off-by: Suren Baghdasaryan <surenb@xxxxxxxxxx>
> > > > ---
> > > > This patch is meant to address Peter's request in [1] to pull
> > > > kthread_queue_delayed_work() out from under rq->lock. This should also address
> > > > the lockdep warning about possibility of a circular dependency described in [2]
> > >
> > > I think you could've just fixed kthread_queue_delayed_work(), that code
> > > is sub-optimal.
>
> After some more staring into kthread code I think I understand what
> Peter's comment meant about delayed_work_list.
> worker->delayed_work_list seems to be unnecessary because each
> kthread_delayed_work has its own timer which will add the work into
> worker->work_list when the time comes. So there is no need to store
> the delayed work in an intermediate worker->delayed_work_list.
> However I think kthread_destroy_worker() has an issue if it's called
> while worker->delayed_work_list is non-empty. The issue is that
> kthread_destroy_worker() does not stop all the
> kthread_delayed_work->timers scheduled on the
> worker->delayed_work_list. So if such a timer fires after a call to
> kthread_destroy_worker(), timer's handler will dereference the already
> destroyed worker.
>
> If I'm right and this is indeed an issue then I think we do need
> worker->delayed_work_list to cancel all the scheduled timers. The
> issue can be avoided if we assume that the caller will alway call
> kthread_cancel_delayed_work_sync() for each delayed_work scheduled on
> worker->delayed_work_list before calling kthread_destroy_worker(). If
> that's what we expect I think this expectation should be reflected in
> the comments and a WARN_ON(!list_empty(&worker->delayed_work_list)) be
> added in kthread_destroy_worker(). WDYT?
>
> >
> > Ok, let me look into it some more. My understanding was that the
> > worker->lock in kthread_queue_delayed_work() was needed to synchronize
> > worker->delayed_work_list access. But maybe I'm missing something... I
> > assume you are talking about optimizing this beyond what
> > https://lkml.org/lkml/2020/5/4/1148 was doing?
> >
> > BTW, any objections against taking https://lkml.org/lkml/2020/5/4/1148
> > ? It's not the ultimate fix but it is an improvement since it gets
> > some of the operations that were unnecessarily under worker->lock out
> > of it.
> >
> > >
> > > But I suppose this works too.
> >
> > In PSI's case there is always one work for each worker, so the
> > delayed_work_list and work_list are not needed and therefore I can
> > replace kthread_worker machinery with a task and a timer.
> > I think I can simplify this a bit further. For example
> > group->poll_wakeup doesn't have to be an atomic. Originally I wanted
> > to avoid a possibility of a race when poll_timer_fn sets it and
> > psi_poll_worker resets it and as a result misses a wakeup, however if
> > psi_poll_worker resets it before calling psi_poll_work then there is
> > no harm in missing a wakeup because we called psi_poll_work and did
> > the required work anyway.
> >
> > One question about this patch I'm not sure about and wanted to ask you
> > Peter is whether it's ok to call mod_timer from within a hotpath
> > (while holding rq->lock). As I described in the additional comment,
> > there is a possibility of a race between when I check timer_pending
> > and the call to mod_timer, so it's possible that mod_timer might be
> > called both from psi_poll_work (psi poll work handler) and from
> > psi_task_change (hotpath under rq->lock). I see that mod_timer takes
> > base->lock spinlock, and IIUC such a race might block the hotpath and
> > therefore is unacceptable. If this is true I'll need to revive the
> > poll_scheduled atomic to close this race and then I can change
> > mod_timer into add_timer.
> > WDYT? And sorry for my ignorance if this is a trivial question. I'm
> > not sure about the rules when it comes to rq->locks.

Thanks for taking this patch, Peter. I just wanted to double-check if
you considered the race I mentioned in the above paragraph and decided
it's a non-issue. If it is an issue I can send a small follow-up patch
to close the race or I can send a new version of the whole patch with
the fix if that's preferable. Please LMK.

> >
> > Thanks,
> > Suren.
> >
> > >
> > > --
> > > To unsubscribe from this group and stop receiving emails from it, send an email to kernel-team+unsubscribe@xxxxxxxxxxxx
> > >