On Tue 24-07-18 06:55:28, Tejun Heo wrote: > Hello, > > On Tue, Jul 24, 2018 at 03:50:22PM +0200, Michal Hocko wrote: > > > So, one good way of thinking about this, I think, could be considering > > > it as a scoped panic_on_oom. However panic_on_oom interacts with > > > memcg ooms, scoping that to cgroup level should likely be how we > > > define group oom. > > > > So what are we going to do if we have a reasonable usecase when somebody > > really wants to have group kill behavior depending on the oom domain? > > I have hard time to imagine such a usecase but my experience tells me > > that users will find a way I have never thought about. > > So, I don't know when that happend but panic_on_oom actually has 0, 1, > 2 settings - 0 no group oom, 1 system kill on oom of any origin, 2 > system kill only if it was a system oom. Maybe we should just follow > that but just start with 1? I am sorry but I do not follow. Besides that modeling the behavior on panic_on_oom doesn't really sound very appealing to me. The knob is a crude hack mostly motivated by debugging (at least its non-global variants). So can we get back to workloads and shape the semantic on top of that please? -- Michal Hocko SUSE Labs
