On Wed, 23 Jun 2021 at 17:13, Odin Ugedal [off-list ref] wrote:

ons. 23. jun. 2021 kl. 15:56 skrev Vincent Guittot [off-list ref]:

quoted

The pelt value of sched_entity is synced with  cfs and its contrib
before being removed.

Hmm. Not sure what you mean by sched_entity here, since this is only
taking the "removed" load_avg
and removing it from cfs_rq, together with (removed.load_avg *
divider) from load_sum. (Although. ".removed" comes from
a sched entity)

The sched_entity's load_avg that is put in removed.load, is sync with
the cfs_rq PELT signal, which includes contrib, before being added to
removed.load.

quoted

Then, we start to remove this load in update_cfs_rq_load_avg() before
calling __update_load_avg_cfs_rq so contrib should not have change and
we should be safe

For what it is worth, I am now able to reproduce it (maybe
CONFIG_HZ=300/250 is the thing) as reported by Sachin,
and my patch makes it disappear. Without my patch I see situations
where _sum is zero while _avg is eg. 1 or 2 or 14 (in that range).

hmm, so there is something wrong in the propagation

This happens for both load, runnable and util.

Lets see what Sachin reports back.

Thanks
Odin