On Fri 19-03-21 17:29:01, Aaron Tomlin wrote: > Hi Michal, > > On Thu 2021-03-18 17:16 +0100, Michal Hocko wrote: > > On Mon 15-03-21 16:58:37, Aaron Tomlin wrote: > > > In the situation where direct reclaim is required to make progress for > > > compaction but no_progress_loops is already over the limit of > > > MAX_RECLAIM_RETRIES consider invoking the oom killer. > > Firstly, thank you for your response. > > > What is the problem you are trying to fix? > > If I understand correctly, in the case of a "costly" order allocation > request that is permitted to repeatedly retry, it is possible to exceed the > maximum reclaim retry threshold as long as "some" progress is being made > even at the highest compaction priority. Costly orders already do have heuristics for the retry in place. Could you be more specific what kind of problem you see with those? > Furthermore, if the allocator has a fatal signal pending, this is not > considered. Fatal signals pending are usually not a strong reason to cut retries count or fail allocations. > In my opinion, it might be better to just give up straight away or try and > use the OOM killer only in the non-costly order allocation scenario to > assit reclaim. Looking at __alloc_pages_may_oom() the current logic is to > entirely skip the OOM killer for a costly order request, which makes sense. Well, opinions might differ of course. The main question is whether there are workloads which are unhappy about the existing behavior. -- Michal Hocko SUSE Labs
