On Mon 20-04-20 18:51:50, Yafang Shao wrote: > On Mon, Apr 20, 2020 at 6:31 PM Michal Hocko <mhocko@xxxxxxxxxx> wrote: [...] > > If this turns out to be infeasible then we should be addressing that > > problem for all possible cases and that means to allow breaking the > > oom_score_adj contract and kill also hidden tasks. > > Breaking the oom_score_adj contract seems another possible way, that > would be accepted by the user - misconfiguration or bugs in user code > should be punished. IOW this is the fault of the user and the kernel > should tell the user the result of this fault. This is not how the kernel behaves in the vast majority of cases. We allow users to shoot their feet. Especially for root only interfaces. We simply rely that admins know what they are doing. So let me repeat there has to be a very strong justification - e.g. it is impossible to remove oom_score_adj OOM killer disabling for some reason and this really happens out there in real deployments. We are surely not going to break the contract based on artificial misconfigured test cases. -- Michal Hocko SUSE Labs
