On Thu 19-04-18 20:51:45, Tetsuo Handa wrote: > Michal Hocko wrote: > > > We need to teach the OOM reaper stop reaping as soon as entering > > > exit_mmap(). > > > Maybe let the OOM reaper poll for progress (e.g. none of > > > get_mm_counter(mm, *) > > > decreased for last 1 second) ? > > > > Can we start simple and build a more elaborate heuristics on top _please_? > > In other words holding the mmap_sem for write for oom victims in > > exit_mmap should handle the problem. We can then enhance this to probe > > for progress or any other clever tricks if we find out that the race > > happens too often and we kill more than necessary. > > > > Let's not repeat the error of trying to be too clever from the beginning > > as we did previously. This are is just too subtle and obviously error > > prone. > > > Something like this?
Not really. This is still building on the tricky locking protocol we have and proven to be error prone. Can we simply take the mmap_sem for write for oom victims before munlocking and release it after munmapping? I am OK with building on the current protocol if taking the mmap_sem for the whole section has some serious down sides but I haven't heard any yet, to be honest. -- Michal Hocko SUSE Labs
