Re: Add rq->nr_uninterruptible count to dest cpu's rq while CPU goes down.

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Mon, Aug 20, 2012 at 11:26:57AM +0200, Peter Zijlstra wrote:
> On Fri, 2012-08-17 at 19:39 +0600, Rakib Mullick wrote:
> > On 8/16/12, Peter Zijlstra <peterz@infradead.org> wrote:
> > > On Thu, 2012-08-16 at 21:32 +0600, Rakib Mullick wrote:
> > >> And also I think migrate_nr_uninterruptible() is meaning less too.
> > >
> > > Hmm, I think I see a problem.. we forget to migrate the effective delta
> > > created by rq->calc_load_active.
> > >
> > And rq->calc_load_active needs to be migrated to the proper dest_rq
> > not like currently picking any random rq.
> 
> 
> OK, so how about something like the below, it would also solve Paul's
> issue with that code.
> 
> 
> Please do double check the logic, I've had all of 4 hours sleep and its
> far too warm for a brain to operate in any case.
> 
> ---
> Subject: sched: Fix load avg vs cpu-hotplug
> 
> Rabik and Paul reported two different issues related to the same few
> lines of code.
> 
> Rabik's issue is that the nr_uninterruptible migration code is wrong in
> that he sees artifacts due to this (Rabik please do expand in more
> detail).
> 
> Paul's issue is that this code as it stands relies on us using
> stop_machine() for unplug, we all would like to remove this assumption
> so that eventually we can remove this stop_machine() usage altogether.
> 
> The only reason we'd have to migrate nr_uninterruptible is so that we
> could use for_each_online_cpu() loops in favour of
> for_each_possible_cpu() loops, however since nr_uninterruptible() is the
> only such loop and its using possible lets not bother at all.
> 
> The problem Rabik sees is (probably) caused by the fact that by
> migrating nr_uninterruptible we screw rq->calc_load_active for both rqs
> involved.
> 
> So don't bother with fancy migration schemes (meaning we now have to
> keep using for_each_possible_cpu()) and instead fold any nr_active delta
> after we migrate all tasks away to make sure we don't have any skewed
> nr_active accounting.
> 
> 
> Reported-by: Rakib Mullick <rakib.mullick@gmail.com>
> Reported-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
> ---
>  kernel/sched/core.c | 31 ++++++++++---------------------
>  1 file changed, 10 insertions(+), 21 deletions(-)
> 
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 4376c9f..06d23c6 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -5338,27 +5338,17 @@ void idle_task_exit(void)
>  }
>  
>  /*
> - * While a dead CPU has no uninterruptible tasks queued at this point,
> - * it might still have a nonzero ->nr_uninterruptible counter, because
> - * for performance reasons the counter is not stricly tracking tasks to
> - * their home CPUs. So we just add the counter to another CPU's counter,
> - * to keep the global sum constant after CPU-down:
> - */
> -static void migrate_nr_uninterruptible(struct rq *rq_src)
> -{
> -	struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
> -
> -	rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
> -	rq_src->nr_uninterruptible = 0;
> -}
> -
> -/*
> - * remove the tasks which were accounted by rq from calc_load_tasks.
> + * Since this CPU is going 'away' for a while, fold any nr_active delta
> + * we might have. Assumes we're called after migrate_tasks() so that the
> + * nr_active count is stable.
> + *
> + * Also see the comment "Global load-average calculations".
>   */
> -static void calc_global_load_remove(struct rq *rq)
> +static void calc_load_migrate(struct rq *rq)
>  {
> -	atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
> -	rq->calc_load_active = 0;
> +	long delta = calc_load_fold_active(rq);
> +	if (delta)
> +		atomic_long_add(delta, &calc_load_tasks);
>  }
>  
>  /*
> @@ -5652,8 +5642,7 @@ migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
>  		BUG_ON(rq->nr_running != 1); /* the migration thread */
>  		raw_spin_unlock_irqrestore(&rq->lock, flags);
>  
> -		migrate_nr_uninterruptible(rq);
> -		calc_global_load_remove(rq);
> +		calc_load_migrate(rq);

Not sure that it matters, but...

This is called from the CPU_DYING notifier, which runs with irqs
disabled, but in process context.  As I understand it, this means that
->nr_running==1.  If my understanding is correct (ha!), this means that
this change sets ->calc_load_active to one (rather than zero as in the
original) and that it subtracts one fewer from calc_load_tasks than did
the original.  Of course, I have no idea whether this matters.

If I am correct and if it does matter, one straightforward fix
is to add a "CPU_DEAD" branch to the switch statement and move the
"calc_load_migrate(rq)" to that new branch.  Given that "rq" references
the outgoing CPU, my guess is that locking is not needed, but you would
know better than I.

							Thanx, Paul

>  		break;
>  #endif
>  	}
>