Hi,
I still see high (upto 30%) ksoftirqd cpu use with 4.19.101+ after these 2 back ports went in for 4.19.101
(had all 4 backports applied earlier to our tree):
commit f6783319737f28e4436a69611853a5a098cbe974 sched/fair: Fix insertion in rq->leaf_cfs_rq_list
commit 5d299eabea5a251fbf66e8277704b874bbba92dc sched/fair: Add tmp_alone_branch assertion
perf shows for any given ksoftirqd, with 20k-30k processes on the system with high scheduler load:
58.88% ksoftirqd/0 [kernel.kallsyms] [k] update_blocked_averages
Can we backport these 2 also, confirmed that it fixes this behavior of ksoftirqd.
commit 039ae8bcf7a5f4476f4487e6bf816885fb3fb617 upstream
commit 31bc6aeaab1d1de8959b67edbed5c7a4b3cdbe7c upstream
The second one doesn’t apply cleanly, there’s a small change for the last diff, where pelt needs renaming to task (cfg_rq_clock_pelt was cfs_rq_clock_task in 4.19.y) in that unchanged part of the diff:
@@ -7700,10 +7720,6 @@ static void update_blocked_averages(int cpu)
for_each_leaf_cfs_rq(rq, cfs_rq) {
struct sched_entity *se;
- /* throttled entities do not contribute to load */
- if (throttled_hierarchy(cfs_rq))
- continue;
-
if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq))
update_tg_load_avg(cfs_rq, 0);
I can post that patch with that rename if required.
Thanks
Vishnu
