lkml.org 
[lkml]   [2021]   [Jul]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] sched: Fix nr_uninterruptible race causing increasing load average
On Thu, Jul 08, 2021 at 09:54:06AM +0200 Peter Zijlstra wrote:
> On Thu, Jul 08, 2021 at 09:48:03AM +0200, Peter Zijlstra wrote:
> > On Thu, Jul 08, 2021 at 09:26:26AM +0200, Peter Zijlstra wrote:
> > > On Wed, Jul 07, 2021 at 03:04:57PM -0400, Phil Auld wrote:
> > > > On systems with weaker memory ordering (e.g. power) commit dbfb089d360b
> > > > ("sched: Fix loadavg accounting race") causes increasing values of load
> > > > average (via rq->calc_load_active and calc_load_tasks) due to the wakeup
> > > > CPU not always seeing the write to task->sched_contributes_to_load in
> > > > __schedule(). Missing that we fail to decrement nr_uninterruptible when
> > > > waking up a task which incremented nr_uninterruptible when it slept.
> > > >
> > > > The rq->lock serialization is insufficient across different rq->locks.
> > > >
> > > > Add smp_wmb() to schedule and smp_rmb() before the read in
> > > > ttwu_do_activate().
> > >
> > > > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > > > index 4ca80df205ce..ced7074716eb 100644
> > > > --- a/kernel/sched/core.c
> > > > +++ b/kernel/sched/core.c
> > > > @@ -2992,6 +2992,8 @@ ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
> > > >
> > > > lockdep_assert_held(&rq->lock);
> > > >
> > > > + /* Pairs with smp_wmb in __schedule() */
> > > > + smp_rmb();
> > > > if (p->sched_contributes_to_load)
> > > > rq->nr_uninterruptible--;
> > > >
> > >
> > > Is this really needed ?! (this question is a big fat clue the comment is
> > > insufficient). AFAICT try_to_wake_up() has a LOAD-ACQUIRE on p->on_rq
> > > and hence the p->sched_contributed_to_load must already happen after.
> > >
> > > > @@ -5084,6 +5086,11 @@ static void __sched notrace __schedule(bool preempt)
> > > > !(prev_state & TASK_NOLOAD) &&
> > > > !(prev->flags & PF_FROZEN);
> > > >
> > > > + /*
> > > > + * Make sure the previous write is ordered before p->on_rq etc so
> > > > + * that it is visible to other cpus in the wakeup path (ttwu_do_activate()).
> > > > + */
> > > > + smp_wmb();
> > > > if (prev->sched_contributes_to_load)
> > > > rq->nr_uninterruptible++;
> > >
> > > That comment is terrible, look at all the other barrier comments around
> > > there for clues; in effect you're worrying about:
> > >
> > > p->sched_contributes_to_load = X R1 = p->on_rq
> > > WMB RMB
> > > p->on_rq = Y R2 = p->sched_contributes_to_load
> > >
> > > Right?
> > >
> > >
> > > Bah bah bah.. I so detest having to add barriers here for silly
> > > accounting. Let me think about this a little.
> >
> > I got the below:
> >
> > __schedule() ttwu()
> >
> > rq_lock() raw_spin_lock(&p->pi_lock)
> > smp_mb__after_spinlock(); smp_mb__after_spinlock();
> >
> > p->sched_contributes_to_load = X; if (READ_ONCE(p->on_rq) && ...)
> > goto unlock;
> > smp_acquire__after_ctrl_dep();
> >
> > smp_cond_load_acquire(&p->on_cpu, !VAL)
> >
> > deactivate_task()
> > p->on_rq = 0;
> >
> > context_switch()
> > finish_task_switch()
> > finish_task()
> > smp_store_release(p->on_cpu, 0);
> >
> > ttwu_queue()
> > rq_lock()
> > ttwu_do_activate()
> > if (p->sched_contributes_to_load)
> > ...
> > rq_unlock()
> > raw_spin_unlock(&p->pi_lock);
> > finish_lock_switch()
> > rq_unlock();
> >
> >
> >
> > The only way for ttwu() to end up in an enqueue, is if it did a
> > LOAD-ACQUIRE on ->on_cpu,
>
> That's not completely true; there's the WF_ON_CPU case, but in that
> scenario we IPI the CPU doing __schedule and it becomes simple UP/PO and
> everything must trivially work.
>
> > but that orders with the STORE-RELEASE on the
> > same, which ensures the p->sched_contributes_to_load LOAD must happen
> > after the STORE.
> >
> > What am I missing? Your Changelog/comments provide insufficient clues..
>

Sorry... I don't have a nice diagram. I'm still looking at what all those
macros actually mean on the various architectures.

"Works great in practice but how does it work in theory?" :)

Using what you have above I get the same thing. It looks like it should be
ordered but in practice it's not, and ordering it "more" as I did in the
patch, fixes it.

Is it possible that the bit field is causing some of the assumptions about
ordering in those various macros to be off?

I notice in all the comments about smp_mb__after_spinlock etc, it's always
WRITE_ONCE/READ_ONCE on the variables in question but we can't do that with
the bit field.


I appreciate your time on this.


Cheers,
Phil

--

\
 
 \ /
  Last update: 2021-07-08 16:56    [W:0.071 / U:2.084 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site