On Thu 12-10-17 16:57:22, Greg Thelen wrote: [...] > Overcharging kmem with deferred reconciliation sounds good to me. > > A few comments (not reasons to avoid this): > > 1) If a task is moved between memcg it seems possible to overcharge > multiple oom memcg for different kmem/user allocations. > mem_cgroup_oom_synchronize() would see at most one oom memcg in > current->memcg_in_oom. Thus it'd only reconcile a single memcg. But > that seems pretty rare and the next charge to any of the other memcg > would reconcile them. This is a general problem for the cgroup v2 memcg oom handling. > 2) if a kernel thread charges kmem on behalf of a client mm then there > is no good place to call mem_cgroup_oom_synchronize(), short of > launching a work item in mem_cgroup_oom(). I don't we have anything > like that yet. So nothing to worry about. If we do invoke the OOM killer from the charge path, it shouldn't be a problem. > 3) it's debatable if mem_cgroup_oom_synchronize() should first attempt > reclaim before killing. But that's a whole 'nother thread. Again, this shouldn't be an issue if we invoke the oom killer from the charge path. > 4) overcharging with deferred reconciliation could also be used for user > pages. But I haven't looked at the code long enough to know if this > would be a net win. It would solve g-u-p issues failing with ENOMEM unexpectedly just because of memcg charge failure. -- Michal Hocko SUSE Labs
