On Thu, 27 May 2021 at 11:46, Odin Ugedal [off-list ref] wrote:

Hi,

quoted

I finally got it this morning with your script and I confirm that the
problem of load_sum == 0 but load_avg != 0 comes from
update_tg_cfs_load(). Then, it seems that we don't call
update_tg_load_avg for this cfs_rq in __update_blocked_fair() because
of a recent update while propagating child's load changes. At the end
we remove the cfs_rq from the list without updating its contribution.

I'm going to prepare a patch to fix this

Yeah, that is another way to look at it. Have not verified, but
wouldn't update_tg_load_avg() in this case
just remove the diff (load_avg - tg_load_avg_contrib)? Wouldn't we
still see some tg_load_avg_contrib
after the cfs_rq is removed from the list then? Eg. in my example
above, the cfs_rq will be removed from
the list while tg_load_avg_contrib=2, or am I missing something? That
was my thought when I looked
at it last week at least..

1st : ensure that cfs_rq->load_sum is not null if cfs_rq-> load_isn't too
2nd : call update_tg_load_avg() during child update so we will be sure
to update tg_load_avg_contrib before removing the cfs from the list

Thanks
Odin