On Thu 06-09-18 20:50:53, Tetsuo Handa wrote: > On 2018/09/06 20:35, Michal Hocko wrote: > > On Sat 01-09-18 20:48:57, Tetsuo Handa wrote: > >> On 2018/08/07 5:51, Michal Hocko wrote: > >>>> At the risk of continually repeating the same statement, the oom reaper > >>>> cannot provide the direct feedback for all possible memory freeing. > >>>> Waking up periodically and finding mm->mmap_sem contended is one problem, > >>>> but the other problem that I've already shown is the unnecessary oom > >>>> killing of additional processes while a thread has already reached > >>>> exit_mmap(). The oom reaper cannot free page tables which is problematic > >>>> for malloc implementations such as tcmalloc that do not release virtual > >>>> memory. > >>> > >>> But once we know that the exit path is past the point of blocking we can > >>> have MMF_OOM_SKIP handover from the oom_reaper to the exit path. So the > >>> oom_reaper doesn't hide the current victim too early and we can safely > >>> wait for the exit path to reclaim the rest. So there is a feedback > >>> channel. I would even do not mind to poll for that state few times - > >>> similar to polling for the mmap_sem. But it would still be some feedback > >>> rather than a certain amount of time has passed since the last check. > >> > >> Michal, will you show us how we can handover as an actual patch? I'm not > >> happy with postponing current situation with just your wish to handover. > > > > I am sorry but I am bussy with other higher priority issues. I believe I > > have outlined the scheme that might work (see above). All it takes is to > > look into that closer a play with it. > > If you are too busy, please show "the point of no-blocking" using source code > instead. If such "the point of no-blocking" really exists, it can be executed > by allocating threads. I would have to study this much deeper but I _suspect_ that we are not taking any blocking locks right after we return from unmap_vmas. In other words the place we used to have synchronization with the oom_reaper in the past. > I think that such "the point of no-blocking" is so late stage of > freeing that it makes no difference with timeout based back off. It is! It is feedback driven rather than a random time passed approach. And more importantly this syncing with exit_mmap matters only when there is going to be way much more memory in page tables than in mappings which is a _rare_ case. -- Michal Hocko SUSE Labs
