On Tue 18-05-21 15:05:54, Aaron Tomlin wrote: > Michal, > > On Fri 2021-03-26 16:36 +0100, Michal Hocko wrote: > > OK, I kinda expected this would be not easily reproducible. > > Unfortunately, I'm still waiting for feedback on this. > > > We should be focusing on the compaction retry logic and see whether we > > can have some "run away" scenarios there. Seeing so many retries without > > compaction bailing out sounds like a bug in that retry logic. > > I suspect so. > > This is indeed a case of excessive reclaim/compaction retries (i.e. the > last known value stored in the no_progress_loops variable was 31,611,688). > > What might be particularly unique about this situation is that a fatal > signal was found pending. In this context, if I understand correctly, it > does not make sense to retry compaction when the last known compact result > was skipped and a fatal signal is pending. OK, this might be an interesting lead. > Looking at try_to_compact_pages(), indeed COMPACT_SKIPPED can be returned; > albeit, not every zone, on the zone list, would be considered in the case > a fatal signal is found to be pending. Yet, in should_compact_retry(), > given the last known compaction result, each zone, on the zone list, can be > considered/or checked (see compaction_zonelist_suitable()). If a zone e.g. > was found to succeed then reclaim/compaction would be tried again > (notwithstanding the above). I believe Vlastimil would be much better fit into looking into those details but it smells like pending fatal signals can lead to a unbound retry indeed. Thanks! -- Michal Hocko SUSE Labs
