On Fri, 2016-04-01 at 15:42 -0400, Chris Metcalf wrote: > On arm64, when calling enqueue_task_fair() from migration_cpu_stop(), > we find the nr_running value updated by add_nr_running(), but the > cfs.nr_running value has not always yet been updated. Accordingly, > the sched_can_stop_tick() false returns true when we are migrating a > second task onto a core.
I don't get it.
Looking at the enqueue_task_fair(), I see this:
for_each_sched_entity(se) {
cfs_rq = cfs_rq_of(se);
cfs_rq->h_nr_running++;
...
}
if (!se)
add_nr_running(rq, 1);
What is the difference between cfs_rq->h_nr_running,
and rq->cfs.nr_running?
Why do we have two?
Are we simply testing against the wrong one in
sched_can_stop_tick?
> Correct this by using rq->nr_running instead of rq->cfs.nr_running.
> This should always be more conservative, and reverts the test to the
> form it had before commit 76d92ac305f2 ("sched: Migrate sched to use
> new tick dependency mask model").
That would cause us to run the timer tick while running
a single SCHED_RR real time task, with a single
SCHED_OTHER task sitting in the background (which will
not get run until the SCHED_RR task is done).
I don't think that is the quite behaviour we want.
> Signed-off-by: Chris Metcalf <cmetc...@mellanox.com>
> ---
> I found this bug because I had a program running in nohz_full
> on a core, and from a different core I called sched_setaffinity()
> to force that task onto the nohz_full core, but I did not end up with
> a kick to the nohz_full core, so tick-based scheduling did not start.
> This is probably bad enough that we should fix it for 4.6.
>
> Strangely, for some reason, the existing code worked correctly for me
> for tilegx, but not for arm64. I see that the enqueue_task_fair()
> code calls enqueue_entity(), which calls account_entity_enqueue() to
> adjust cfs.nr_running. That seemed to happen on tilegx, but not
> arm64.
> Perhaps there is some difference in how the sched_entity stuff is
> done,
> but frankly that took me a little deeper into the CFS stuff than I
> was
> willing to dive in this moment.
>
> I could also argue that sched/core.c shouldn't have a lot of CFS
> stuff in it anyway, and if we view the FIFO/RR stuff as handling the
> real special cases in sched_can_stop_tick() anyway, then just
> checking
> the core nr_running feels like the right thing to do regardless.
>
> kernel/sched/core.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 00649f7ad567..1737d63c65fa 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -599,7 +599,7 @@ bool sched_can_stop_tick(struct rq *rq)
> }
>
> /* Normal multitasking need periodic preemption checks */
> - if (rq->cfs.nr_running > 1)
> + if (rq->nr_running > 1)
> return false;
>
> return true;
--
All Rights Reversed.
signature.asc
Description: This is a digitally signed message part
