[Please do not top post] On Fri 20-12-19 17:56:20, 彭志刚 wrote: > certainly. > > Steps to reproduce: > (1)Create a mm cgroup and set memory.limit_in_bytes > (2)Move the bash process to the newly created cgroup, and set the > oom_score_adj of the bash process to -998. > (3)In bash, start multiple processes, each process consumes different > memory until cgroup oom is triggered. > > The triggered phenomenon is shown below. We can see that when cgroup oom > happened, process 23777 was killed, but in fact, 23772 consumes more memory; > > [ 591.000970] Tasks state (memory values in pages): > [ 591.000970] [ pid ] uid tgid total_vm rss pgtables_bytes swapents oom_score_adj name > [ 591.000973] [ 23344] 0 23344 2863 923 61440 0 -998 bash > [ 591.000975] [ 23714] 0 23714 27522 25935 258048 0 -998 test > [ 591.000976] [ 23772] 0 23772 104622 103032 876544 0 -998 test points = 103032 + 0 + 876544/4096 = 103246 > [ 591.000978] [ 23777] 0 23777 78922 77335 667648 0 -998 test points = 77335 + 0 + 667648/4096 = 77498 It is not clear what is the actual hard limit but let's assume that rss+page_tables is the only charged memory (or at least the majority of it). That would be 207680 so the normalized oom_score_adj would be -206586 which is way too big for both tasks so from the OOM killer perspective both tasks are equal. The question is whether this is a bug or a (mis)feature. The oom_score_adj je documented as follows: Documentation/filesystems/proc.txt : Consequently, it is very simple for userspace to define the amount of memory to : consider for each task. Setting a /proc/<pid>/oom_score_adj value of +500, for : example, is roughly equivalent to allowing the remainder of tasks sharing the : same system, cpuset, mempolicy, or memory controller resources to use at least : 50% more memory. A value of -500, on the other hand, would be roughly : equivalent to discounting 50% of the task's allowed memory from being considered : as scoring against the task. Which implies that we are talking about the budget based on a usable memory (aka hard limit in this case). I do agree that the semantic is awkward. I know there are usecases which try to use the existing scheme for oom_score_adj to fine tune oom decisions and I am worried your patch might break those. That being said, I am not sure this change is safe wrt. to backward compatibility. I would rather recommend to not using oom_score_adj for anything but OOM_SCORE_ADJ_MIN resp OOM_SCORE_ADJ_MAX. -- Michal Hocko SUSE Labs
